Colored-noise-like waveforms and codes in ultrasound imaging

Abstract

Techniques, structures, and systems are disclosed for ultrasound imaging using noise-like waveforms. In some aspects, a system for synthesizing noise-like waveforms for ultrasound imaging includes an array of transducer elements to transmit an ultrasound signal at a target volume; one or more digital-to-analog converters (DACs) in communication with the array of transducer elements; a computing device, comprising a processor and a memory, in communication with the one or more DACs; and one or more analog-to-digital converters (ADCs) in communication with the array of transducer elements and the computing device, wherein the computing device is configured to generate one or more noise-like coded waveforms that possess a number of signed amplitude levels of at least three levels.

Description

PCT Patent Application Attorney Docket No. 116291.8025. WO00COLORED-NOISE-LIKE WAVEFORMS AND CODES IN ULTRASOUND IMAGING CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This patent document claims priority to and benefits of U. S. Provisional Patent Application No. 63 / 729,242 titled “NOISE-LIKE WAVEFORMS IN ULTRASOUND IMAGING” filed on December 6, 2024. The entire contents of the aforementioned patent application are incorporated by reference as part of the disclosure of this patent document.TECHNICAL FIELD

[0002] This patent document relates to systems, devices, and processes that use ultrasound imaging.BACKGROUND

[0003] Ultrasound imaging is an imaging modality that employs the properties of sound waves traveling through a medium to render a visual image. Ultrasound imaging has been used as an imaging modality for decades in various biomedical fields to view the internal structures and functions of animals and humans. Ultrasound waves used in biomedical imaging may operate in different frequencies, e.g., between 1 and 20 MHz or even higher frequencies. Some factors, including inadequate contrast, dynamic range, depth-of-penetration, signal-to-noise ratio, and spatial and Doppler resolution, can lead to less-than-desirable image quality using conventional techniques of ultrasound imaging, which can limit its use for many clinical applications.SUMMARY

[0004] Techniques, systems, and apparatuses are disclosed for ultrasound imaging that utilize waveforms that are produced in the time domain to produce waveforms that can be characterized as being composite, spread-spectrum, coherent, and noise-like. In some aspects, in accordance with the present technology, the disclosed techniques, systems, and apparatuses can produce waveforms that are noise-like, which can be characterized as coded waveforms. Unlike commonly- encountered coded waveforms, which are based on well-known, deterministic, mathematically derived and perhaps ideal properties, such as for example, Barker, Golay, Gold codes, etc., noiselike waveforms can improve medical ultrasound images over what can be achieved with commonly1184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00used coded waveforms. For example, developed noise-like waveforms can improve medical ultrasound images by optimizing waveform characteristics, some of which can also be characterized as being colored-noise-like or white-noise-like. The disclosed techniques, systems, and apparatuses are created so as to produce noise-like, coded waveforms that have similar advantages as the amplitude and phase-coded composite waveforms produced in the time domain to achieve improvements in ultrasound imaging performance. The disclosed techniques, systems, and apparatuses can involve hardware configurations that optimize use of low-power Digital-to-Analog Converters (DACs), linear power amplifiers (LPAs), and / or low complexity transmitters to reduce space, weight, and power for an ultrasound imaging system while achieving improvements in ultrasound imaging performance.

[0005] In some aspects of the disclosed technology, a method of creating an image from an acoustic colored-noise-like coded waveform includes setting a transmit / receive switch into transmit mode, employing a stored composite, time-domain, digital waveform in one or more waveform synthesizers, transmitting an acoustic composite waveform based on the stored composite digital waveform toward a target, setting the transmit / receive switch into receive mode, receiving a returned echo (e.g., the acoustic waveform returned from at least part of the target area that is the tissue volume being studied), converting the received echo from analog format to digital format as a received composite waveform comprising information of the target area, and processing the received composite waveform to produce an image of at least part of the target area.

[0006] In some aspects, a method for synthesizing noise-like waveforms for ultrasound imaging includes producing a set of one or more candidate seed waveforms based on at least one waveform parameter; producing an optimized noise-like waveform set by iterating a set of trial waveforms using the set of one or more candidate seed waveforms and based on one or more optimization parameters; and selecting one or more output waveforms from the optimized noiselike waveform set based on one or more selection criteria to produce a set of noise-like waveforms.

[0007] In some aspects, a method for synthesizing noise-like waveforms for ultrasound imaging includes producing a set of one or more candidate seed codes based on at least one code parameter; producing an optimized noise-like code set by iterating a set of trial codes using the set of one or more candidate seed codes and based on one or more code optimization parameters; selecting one or more codes from the optimized noise-like code set based on one or more code selection criteria to form a selected set of waveform-formation codes; and generating a set of noise-2184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00like waveforms by processing the selected set of waveform-formation codes with respect to one or more waveform parameters.

[0008] In some aspects, a system for synthesizing noise-like waveforms for ultrasound imaging includes an array of transducer elements to transmit an ultrasound signal at a target volume; one or more digital-to-analog converters (DACs) in communication with the array of transducer elements: a computing device, comprising a processor and a memory, in communication with the one or more DACs; and one or more analog-to-digital converters (ADCs) in communication with the array of transducer elements and the computing device, wherein the computing device is configured to generate one or more noise-like coded waveforms that possess a number of signed amplitude levels in a range of 3 to 4,096 levels.

[0009] In some aspects, a machine learning-based method for generating optimized coded waveforms for ultrasound imaging includes training a machine learning model on a dataset comprising waveform parameters and associated performance metrics including at least one of a sidelobe level (SLL) or a noise gain (NG): receiving optimization parameters specifying target performance characteristics for a coded waveform; generating, using the trained machine learning model, a candidate coded waveform predicted to meet the target performance characteristics; and outputting the candidate coded waveform for use in an ultrasound transmission.

[0010] The subject matter described in this patent document can provide one or more of the following features and be used in many applications. For example, the disclosed technology can be used during routine primary care screenings to identify and locate early-stage malignancies, as well as later-stage cancers, which can potentially raise survival rates of hard-to-diagnose asymptomatic patients. The disclosed technology can be used by board-certified radiologists to diagnose neoplasms as benign or malignant prior to any surgical biopsy or resection intervention, which may also improve patient survival rate while reducing unnecessary biopsies. The disclosed technology can, when integrated with surgical instruments, be used in medical procedures to confirm noninvasive diagnoses by image-guided fine needle aspiration or to ablate lesions by fine needle radiofrequency probes, which can improve the outcome of such procedures. The disclosed technology can w'hen integrated with minimally invasive surgical high-definition video instrumentation, fuse optical and ultrasound images, which can further give surgeons added abilities to locate and surgically excise diseased tissue without excising excessive healthy tissue. The disclosed technology can reduce the amount of time for the brachytherapy treatment of3184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00malignant neoplasms by, for example, precisely guiding the insenion of catheters and sealed radioactive sources into the proper location. Similarly, the disclosed technology can aid the insertion of high-dose, time-released localized pharmaceuticals for treatments of diseases.BRIEF DESCRIPTION OF THE DRAWINGS

[0011] FIG. 1A-1 shows a block diagram of an example embodiment of a system for synthesizing colored-noise-like waveforms and codes for ultrasound imaging, in accordance with the present technology.

[0012] FIG. 1 A-2 shows a diagram of an example embodiment of a system for implementing composite, coherent, spread- spectrum, colored-noise-like coded waveforms in ultrasound imaging, in accordance with the present technology.

[0013] FIG. IB shows a diagram of a method for synthesizing a composite, coherent, spread¬ spectrum, colored-noise-like coded waveforms for ultrasound imaging, in accordance with the present technology.

[0014] FIG. IC shows an example embodiment of an ultrasound system in accordance with the disclosed noise-like waveform ultrasound imaging technology.

[0015] FIG. ID shows an example embodiment of a colored-noise-like waveform exciter¬ transmitter modules depicted in the system of FIG. IC.

[0016] FIG. IE shows an example embodiment of the analog Rx modules depicted in the system of FIG. IC.

[0017] FIG. 2A shows a diagram illustrating an example embodiment of a method for producing colored-noise-like coded waveforms, in accordance with the present technology.

[0018] FIG. 2B shows a diagram illustrating an example embodiment of a method for producing waveform parameters that can be utilized in the method shown in FIG. 2A.

[0019] FIG. 2C shows a diagram illustrating an example embodiment of a method for producing colored-noise-like coded waveforms from selected waveform codes, in accordance with the present technology.

[0020] FIG. 3 shows a data plot depicting an electrical sinusoidal excitation conversion to an acoustic output response for a medical ultrasound transducer array.

[0021] FIG. 4 show's an example embodiment of a digital signal processing technique that can be employed by an example Digital Signal Processor of the system shown in FIG. IB for receiving4184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00digital Colored-Noise-Like Waveform(s) that is / are the output signal produced by A / D Converter(s) of the system.

[0022] FIGS. 5A-5D show data plots depicting examples features associated with colored¬ noise-like waveforms produceable by implementations of the method shown in FIG. 2A.

[0023] FIGS. 6A-6C show data plots depicting three scenarios of example implementations for producing colored-noise-like coded waveforms from selected waveform codes, in accordance with aspects of the method shown in FIG. 2C.

[0024] FIGS. 7 A and 7B show data plots depicting three scenarios of example implementations of a pre-conditioning process and a filtering process, respectively, in accordance with the processes depicted in FIG. 4.

[0025] FIGS. 8A-8F show diagrams depicting optimized sets of codes for sidelobe level (SLL) and noise gain (NG) of particular lengths, for example 32, 64, 48, 64, 128, and 256, for use in generating coded colored-noise-like waveforms produceable by implementations of the example method shown in FIG. 2 A.

[0026] FIG. 9 shows a diagram of an example embodiment of a machine learning-based method for generating optimized coded waveforms for ultrasound imaging, in accordance with the present technology.DETAILED DESCRIPTION

[0027] Some ultrasound imaging techniques can use noise-like, spread-spectrum, coded waveforms that offer several significant advantages as compared to conventional pulse waveforms used in most ultrasound imaging systems. These advantages include higher axial resolution, lower range sidelobes, higher contrast ratio, better simultaneous axial and Doppler resolution, and greater processing gain can improve image quality over conventional ultrasound imaging techniques used by most current systems. Example systems and methods that utilize noise-like, spread-spectrum, coded waveforms in ultrasound imaging and other applications are described in U. S. Patent Publication No. 2013 / 0123635 Al, which is incorporated by reference in its entirety for all purposes.

[0028] Further, noise-like, spread-spectrum, coded waveforms can be used in synthetic aperture image formation offer additional significant improvements conventional pulse waveforms used in most ultrasound imaging products. Using noise-like, spread-spectrum, coded waveforms5184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00in synthetic aperture ultrasound imaging can provide additional advantages that include improved 2D and 3D point-spread functions and optimal image focusing that is independent of depth. Example systems and methods that utilize noise-like, spread-spectrum, coded waveforms in synthetic aperture ultrasound imaging and other synthetic aperture applications are described in U. S. Patent Publication No. 2015 / 0080725 Al, which is incorporated by reference in its entirety for all purposes.

[0029] Yet, to implement noise-like, spread- spectrum, coded waveforms ultrasound imaging techniques, whether in real aperture or synthetic aperture applications, the hardware requirements typically involve medium resolution. 8 to 16-bit, high-speed. low-power Digital-to-Analog Converters (DACs), whose analog output signal is increased by linear power amplifiers (LPAs), and frequently, high-voltage Multiplexers (MUXs) amplifiers to drive acoustic transducers to produce composite acoustic waveforms. Yet, this combination of high-speed, low-power DACs plus linear output amplifiers, plus high voltage MUSs can be costly in terms of acquisition, space, weight, and power. Thus, a new approach that can reap the advantages of noise-like, spread¬ spectrum. coded waveforms ultrasound imaging techniques without their current size, power, and cost expensive requirements would be beneficial.

[0030] Thus, this disclosed new approach, which can reap the advantages of noise-like, spread¬ spectrum, coded waveforms in either real aperture or synthetic aperture ultrasound imaging applications, without the current size, power, and cost expensive burdens, would be beneficial.

[0031] Techniques, systems, and apparatuses are described for synthesizing, generating, transmitting, receiving, and processing, composite, coherent, spread- spectrum, colored-noise-like coded waveforms used in ultrasound imaging.

[0032] In acoustics, electronics, and physics, the term “noise” refers to the power spectrum of a signal that is produced by a sequential process that can be a stochastic process, but the sequential process is typically not well-defined in common usage. For the purposes of this disclosure, “noise” is characterized by its state space, which includes the cardinality, the index set, the time increments, and the interdependence of the variables in the sequential process that generates the noise.

[0033] The term “noise-like,” as used in this disclosure, is a type of noise that is generated by a process that can be either random, pseudorandom, deterministic, or some combination of these types of processes. For example, a noise-like signal, which is also known as a waveform, can be6184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00generated by a pseudorandom number generator algorithm that generates a sequence of signal amplitudes that approximate the properties of a finite sequence of sequential, random signal amplitudes.

[0034] Also in acoustics, the “color of noise,” which is also known as noise spectrum, refers to the characteristics of the power spectrum of the noise. Depending on the characteristics of its power spectrum, colored noise can exhibit significantly different properties. For example, white noise is an infinite bandwidth, infinite duration signal whose spectrum is statistically flat and has equal average power for any selected sub-band as compared to any other sub-band of equal width in the white noise spectrum. The term “colored-noise-like,” as used in this disclosure, is a type of noise-like signal that does not necessarily have equal average power for any selected sub-band as compared to any other sub-band of equal width in the band-limited and time-limited spectrum of the colored-noise-like waveform. Waveforms that are generated by the processes described herein can have colored noise like attributes.

[0035] Ultrasound imaging can be performed by emitting a time-gated, single frequency or a pulsed, earner frequency (pulse) acoustic waveform, which is partly reflected from a boundary between two mediums (e.g., biological tissue structures) and partially transmitted. The reflection can depend on the acoustic impedance difference between the two mediums. Ultrasound imaging by some techniques may only use amplitude information from the reflected signal. For example, when one pulse is emitted, the reflected signal can be sampled continuously. In biological tissue, sound velocity can be considered fairly constant, in which the time between the emission of a waveform and the reception of a reflected signal is dependent on the distance the waveform travels in that tissue structure (e.g., the depth of the reflecting structure). Therefore, reflected signals may be sampled at multiple time intervals to receive the reflected signals being reflected from multiple depths. Also, different tissues at the different depths can (partially) reflect the waveform with different amounts of energy, and thus the reflected signal from different mediums can have different amplitudes. A corresponding ultrasound image can be constructed based on depth. The time before a new waveform is emitted can therefore be dependent of the maximum depth that is desired to image. Ultrasound imaging techniques employing pulsed monochromatic and / or narrow instantaneous bandwidth waveforms can suffer from poor resolution of image processing and production. Yet, waveforms with spread-spectrum, wide instantaneous bandwidth, noise-like characteristics can yield higher quality images.7184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00

[0036] The disclosed colored-noise-like, coded waveforms have similar advantages as the amplitude and phase-coded composite waveforms produced in the time domain to achieve improvements in ultrasound imaging performance. The disclosed technology can involve hardware other than using a combination of low-power DACs and LPAs to reduce space, weight, and power while achieving improvements in ultrasound imaging performance.

[0037] In one aspect of the disclosed technology, a method of creating an image from an acoustic colored-noise-like coded waveform includes setting a transmit / receive switch into transmit mode, employing a stored composite, time-domain, digital waveform in one or more waveform synthesizers, transmitting an acoustic composite waveform based on the stored composite digital waveform toward a target, setting the transmit / receive switch into receive mode, receiving a returned echo (e.g., the acoustic waveform returned from at least part of the target area that is the tissue volume being studied), converting the received echo from analog format to digital format as a received composite waveform comprising information of the target area, and processing the received composite waveform to produce an image of at least part of the target area.

[0038] FIG. 1A-1 shows a block diagram of an example embodiment of a system 1000 for synthesizing noise-like waveforms and codes for ultrasound imaging, in accordance with the present technology. The system 1000 includes an array of transducer elements 1010 configured to transmit an ultrasound signal at a target volume; one or more digital-to-analog converters (DACs) 1020 in communication with the array of transducer elements 1010; a data processing device 1030, e.g., a computing device, comprising one or more processor(s) in communication with one or more memory(ies) and optionally with one or more (optional) input / output units, where the data processing device 1010 is in communication with the one or more DACs 1020; and one or more analog-to-digital converters (ADCs) 1040 in communication with the array of transducer elements 1010 and the data processing device 1030. The data processing device 1030 is configured to generate one or more noise-like coded waveforms that possess a number of signed amplitude levels, e.g., which can be in a range of 3 to 4,096 levels.

[0039] In some implementations of the system 1000, the data processing device 1030 is configured to produce a set of one or more candidate seed waveforms based on at least one waveform parameter; produce an optimized noise-like waveform set by iterating a set of trial waveforms using the set of one or more candidate seed waveforms and based on one or more optimization parameters; and select one or more output waveforms from the optimized noise-like8184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00waveform set based on one or more selection criteria to produce a set of the one or more noise¬ like coded waveforms. In some implementations of the system 1000, the data processing device 1030 is configured to produce a set of one or more candidate seed codes based on at least one code parameter; produce an optimized noise-like code set by iterating a set of trial codes using the set of one or more candidate seed codes and based on one or more code optimization parameters; select one or more codes from the optimized noise-like code set based on one or more code selection criteria to form a selected set of waveform-formation codes; and generate a set of the one or more noise-like coded waveforms by processing the selected set of waveform-formation codes with respect to one or more waveform parameters. In some embodiments of the system 1000. for example, the one or more DACs 1020 includes one or more low-power DACs or a one or more high-power DACs. In some embodiments of the system 1000, for example, the system 1000 can optionally include one or more output amplifiers (not shown) in communication with the array of transducer elements 1010 and the one or more DACs 1020, wherein the one more output amplifiers is configured to amplify an analog signal output of the one or more DACs 1020. For example, the (optional) one or more output amplifiers can include one or more linear power amplifiers (LPAs). In some embodiments of the system 1000, for example, the system 1000 can optionally include one or more pre-amplifiers (not shown) in communication with the array of transducer elements 1010 and the one or more ADCs 1040, wherein the one more pre-amplifiers is configured to amplify a received ultrasonic signal that is returned from the volume of interest based on a transmitted ultrasound signal.

[0040] Example embodiments of the system 1000 are described below, including but not limited to system 100 and 100C described in FIG. 1A-2 and FIG. 1C.

[0041] FIG. 1A-2 shows a block diagram of an example embodiment of an ultrasound system 100 that can produce acoustic waveforms with enhanced waveform properties that include a composite, coherent, spread-spectrum, colored-noise-like coded waveforms. System 100 can be configured in one of many system designs. In the example shown in FIG. 1A-2, the system 100 includes a System Controller 102 that can include a processing unit, e.g., including but not limited to a central processing unit (CPU) of RISC-based or other types of CPU architectures. System Controller 102 can also include at least one input / output (I / O) unit(s) and / or memory unit(s), which are in communication with the processing unit, to support various functions of the System Controller 102. In some embodiments of the system 100, the System Controller 102 that may be9184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00interfaced with a (optional) Master Clock 101 configured for time synchronization of components and modules of the system 100. For example, the processing unit can be associated with a system control bus, e.g.. illustrated in the diagram of FIG. 1A-2 as Data & Control Bus 103. In various embodiments of the system 100, the System Controller 102 can be implemented as one of various data processing architectures, such as a personal computer (PC), laptop, tablet, and / or mobile communication device architectures.

[0042] Die memory unit(s) in the System Controller 102 can store other information and data, such as instructions, software, values, images, and other data processed or referenced by the processing unit. Various types of Random Access Memory (RAM) devices, Read-Only Memory (ROM) devices, Flash Memory devices, and other suitable storage media can be used to implement storage functions of the memory unit(s). The memory unit(s) can store pre-store digital composite time domain waveforms, ancillary, and other information. The memory unit(s) can store data and information obtained from received and processed waveforms, which can be used to generate and transmit new waveforms. The memory unit(s) can be associated with the system control bus, e.g., Data & Control Bus 103.

[0043] The I / O unit(s) can be connected to an external interface, source of data storage, and / or display device. The I / O unit(s) can be associated with the control bus of the system 100, e.g., Data & Control Bus 103. The I / O unit(s) can interface with other components of the system 100 using various types of wired or wireless interfaces compatible with typical data communication standards, such as but not limited to Universal Serial Bus (USB). IEEE 1394 (FireWire). Bluetooth, Bluetooth Low Energy (BLE), Zigbee, IEEE 802.11, Wireless Local Area Network (WLAN), Wireless Personal Area Network (WPAN), Wireless Wide Area Network (WWAN), WiMAX, IEEE 802.16 (Worldwide Interoperability for Microwave Access (WiMAX)), 3G / 4G / LTE / 5G / 6G cellular communication methods, Near Field Communication (NFC), and / or parallel interfaces. The I / O unit(s) can interface with an external interface, source of data storage, or display device to retrieve and transfer data and information that can be processed by the processor unit, stored in the memory unit, or exhibited on an output unit.

[0044] In some embodiments of the system 100, the System Controller 102 can be configured to control at least one, some, or all of the modules of system 100, e.g., through wired or wireless connection to such module(s), such as via the Data & Control Bus 103 as an example. For example, in some embodiments of the system 100, the Data & Control Bus 103 can link the System10184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00Controller 102 to one or more attached, connected, or interfaced digital signal processors, e.g., Digital Signal Processor 104, for processing waveforms for their functional control. The Digital Signal Processor 104 can include one or many processors, such as but not limited to ASIC (application-specific integrated circuit), FPGA (field-programmable gate array), DSP (digital signal processor), AsAP (asynchronous array of simple processors), and other types of data processing architectures. In some embodiments, for example, the Digital Signal Processor 104 can be embodied in the same computing device as the System Controller 102; whereas in some embodiments, for example, the Digital Signal Process 104 can be embodied in a different (but interfaced) computing device as the System Controller 102. In some embodiments of the system 100, for example, the Digital Signal Processor 104 can be embodied in software (e.g., instructions stored in memory unit(s) of the system 100 as code), which can be executed by processing unit(s) (e.g., CPU) of the System Controller 102 and / or other processing unit(s) interfaced to the System Controller 102.

[0045] In some embodiments, for example, the system 100 can include one or more display units that constitute a Display 105. For example, the Data & Control Bus 103 can link System Controller 102, as well as Digital Signal Processor 104, to the one or more display units with modules for user interfaces, e.g., Display 105 with a module User Interface 106. Display 105 can include many suitable display units, such as but not limited to cathode ray tube (CRT), light emitting diode (LED), and liquid crystal display (LCD) monitor and / or screen as a visual display. Display 105 can also include various types of display, speaker, haptic, or printing interfaces. In other examples, Display 105 can include other output apparatuses, such as toner, liquid inkjet, solid ink, dye sublimation, inkless (such as thermal or UV) printing apparatuses and various types of audio signal transducer apparatuses. User Interface 106 can include many suitable interfaces including various types of keyboard, mouse, voice command, touchpad, and brain-machine interface apparatuses.

[0046] The system 100 includes a waveform generator that is configured as a Colored-Noise-Like Waveform Transmit (Tx) Beamformer 107. In some embodiments, Colored-Noise-Like Waveform Tx Beamformer 107 may be embodied as a hardware and / or software module of a computing device, which comprises instructions executable by a computer processor and memory (e.g., of System Controller 102) to shape and steer one or more transmitted acoustic beams from an array of transducer elements (e.g., Transducer Array 112). The Colored-Noise-Like Waveform11184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00Tx Beamformer 107 can compute waveform delays and amplitude weights for each individual or a selected group of the Colored-Noise-Like Waveform Analog Exciter-Transmitters 108. Using information provided by Colored-Noise-Like Waveform Synthesizer 122, the Colored-Noise-Like Waveform Transmit (Tx) Beamformer 107 can use some combination of linear and nonlinear operations applied to delay and weight the noise-like digital waveforms that can be communicated to the Colored-Noise-Like Waveform Analog Exciter-Transmitters 108 to produce one or more transmitted, beamformed acoustic beams as outputs from the array of transducer elements (e.g., Transducer Array 112). Some aspects of beamforming and other techniques implementable by the Colored-Noise-Like Waveform Tx Beamformer 107 are described in U. S. Patent Publication No. 2013 / 0123635 Al and U. S. Patent Publication No. 2015 / 0080725 Al, which are incorporated by reference in their entirety for all purposes.

[0047] In example implementations of the system 100, the Colored-Noise-Like Waveform Tx Beamformer 107 can be controlled by the System Controller 102 for producing one or more digital composite waveforms to be transduced as one or more composite acoustic waveforms that are to be transmitted by system 100, e.g., at a target volume of interest. The System Controller 102 is in communication with a Colored-Noise-Like Waveform Synthesizer 122, which can provide individually synthesized, digital colored-noise-like waveforms for producing one or more composite, colored-noise-like coded waveforms that can be transmitted by the system 100 to the target volume. The Composite Waveform Synthesizer 122 is configured to provide the System Controller 102 with a set of digital waveform data (representative of synthesized individual waveforms having colored-noise-like properties) from which the System Controller 102 can process and / or select as one or more composite digital waveforms to be communicated to the Colored-Noise-Like Waveform Tx Beamformer 107, which processes the digital waveform data as digital inputs for generation of one or more composite analog waveforms by one or more (or an array of) Analog Exciter-Transmitters 108 of the system 100 (also referred to as Colored-Noise- Like Waveform Analog Exciter-Transmitters). In this disclosure, the one or more composite analog waveform(s) is / are the output(s) of the Colored-Noise-Like Waveform Analog Exciter-Transmitters 108 and can be in the form of composite analog RF waveforms, composite analog waveform beams, or other forms, and may be referred to herein as “analog colored-noise-like electronic waveforms.” For example, in some embodiments of the system 100, the Analog Exciter- Transmitters 108 can be embodied with a single component of 1 to 8-bit low-resolution high-power12184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00DACs, which are also known as multi-level pulser transmitters, instead of the more expensive high resolution, low power DACs that require high power, high voltage amplifiers to boost the output of lower DACs to excite the transducer elements for noise-like waveforms.

[0048] The Analog Exciter-Transmitters 108 are in communication with a Transducer Array 112 of the system 100. In some embodiments, which utilize most of currently available types of transducers, the Analog Exciter- Transmitters 108 can be in direct communication with the Transducer Array 112. Although typically not needed, in some embodiments of the system 100, the Analog Exciter-Transmitters 108 can be in communication through (optional) Power Amplifiers 109 with the Transducer Array 112, e.g., which can be due to some unique requirements of the selected Transducer Array 112 or some other requirement for system 100. For example, the (optional) Power Amplifiers 109 (also referred to as (optional) Output Amplifiers 109) can be configured to match the excitation voltage, current or impedance requirements of the transducer elements, or they can be configured to also provide the bias voltages or current required by the transducers of Transducer Array 112. Also for example, in some embodiments where the Transducer Array 112 is configured as a transmitter and receiver (Tx / Rx), the Analog Exciter-Transmitters 108 are in communication with the Transducer Array 112 via transmit / receive TR Switches 110, and (optional) MUX 111 (e.g., high voltage, two-way multiplexer).

[0049] The Transducer Array 112, which is also referred to an ultrasound probe or as a synthetic aperture array or as a tomographic array, is configured to transduce the analog colored¬ noise-like electronic waveforms from the Analog Exciter-Transmitters 108 into Colored-Noise-Like Acoustic Waveforms 113 in an Acoustic Medium 114, e.g., such as a body part or region of a living subject, including but not limited to a breast, arm, leg, neck, throat, head or face, chest, abdominal area, pelvis / waist, shoulder, knee joint, hip joint, ankle, or other anatomical structure of a human subject or animal subject (e.g., canine, equine, feline, etc.). The Transducer Array 112 in system 100 can be implemented using lead zirconate titanate (PZT), lead magnesium niobite (PMN), lead magnesium niobite - lead titanate (PMN-PT), Capacitive Micromachined Ultrasonic Transducer (CMUT), or any other transducer technology as the system design requirements require. The Colored-Noise-Like Acoustic Waveforms 113 propagate from the Transducer Array 112 through the intervening Acoustic Medium 114 to a Target Volume 115. The Target Volume 115 scatters the incident Colored-Noise-Like Acoustic Waveforms 113 back through the intervening Acoustic Medium 114 as receive acoustic waveforms (having the colored-noise-like13184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00waveform properties) to the Transducer Array 112, where it is received and transducted by the Transducer Array 112 into an electronic composite receive waveform.

[0050] Referring back to the Analog Exciter-Transmitters 108 shown in FIG. 1A-2, the individual analog waveforms can be generated by an array of I’ number Colored-Noise-Like Waveform Analog Exciter-Transmitters 108 (where I’ includes, 1, 2,... n number of individual analog exciter-transmitters), whose output can be amplified by an array of I’ number of (optional) Output Amplifiers 109. For example, if the output power of the Exciter-Transmitters 108 is insufficient or more amplitude or phase control of individual F number of analog colored -noise-like electronic waveforms is desired, then the Analog Exciter- Transmitters 108 can be included to directly drive the Transducer Array 112 without the need for output amplifiers to save cost and space; or, alternatively, the (optional) Output Amplifiers 109 can be optionally added to the system 100 to drive the Transducer Array 112, if, for example, more power is needed for a specific use medical procedure or additional waveform control is desired, or if the Transducer Array 112 requires bias voltages.

[0051] The analog colored-noise-like electronic waveforms can be passed to transmit / receive T / R Switches 110, which can be implemented by several types of low-cost hardware, e.g., diode bridges. N-pole double-throw transmit / receive switches, etc. T / R Switches 110 that can be interfaced with a Transducer Array 112.

[0052] For example, in some implementations, if the I number of elements in the Transducer Array 11 is greater than the I’ number of transmitter channels of the Analog Exciter Transmitters 108 (and / or the (optional) Output Amplifiers 109), the individual analog colored-noise-like electronic waveforms can be multiplexed by the (optional) MUX 111 to be communicated sequentially to the I’ number transducer channels available.

[0053] In some implementations, for example, an analog colored-noise-like electronic waveform that is to be transmitted into a target medium can be transduced into an analog colored- noise-like acoustic waveform by the Transducer Array 112, which can include an array of transducer elements comprising I number of elements. Each array element of the Transducer Array 112 may generate one or more analog colored-noise-like acoustic waveform that correspond to the individual waveforms determined by the Colored-noise-Like Waveform Tx Beamformer 107.

[0054] For example, the hardware design of the system 100 can be configured to reduce overall costs, electrical energy (power), and data processing burdens. For example, this can be14184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00accomplished by the system 100 employing low-resolution, 1 to 8-bit, high-speed, high-power, high-voltage DACs, which are also known as multi-level pulsers, to directly transmit composite, colored-noise-like waveforms without the need for a Linear Power Amplifier (LPA), which substantially at a higher cost financially and power-consumption. Some commercially available high-power DACs can also have selectable clock cycle periods between output selectable amplitude pulses. Example hardware embodiments of the system 100, which employ high-speed, high-power DACs, can produce a composite, spread-spectrum, coherent, colored-noise-like coded waveforms to achieve improvement in ultrasound image quality.

[0055] In some embodiments of the system 100, for example, the Colored-Noise-Like Waveform Tx Beamformer 107 can also include at least one or more memory unit(s) that can store digital composite waveform copies produced by the Colored-Noise-Like Waveform Synthesizer 122, which are also termed as “composite waveform replicas.” In such embodiments, for example, the Colored -Noise-Like Waveform Tx Beamformer 107 can modify the composite waveform replicas to incorporate time delays and amplitudes for the purposes of beamforming, focusing and steering, as commanded by the System Controller 102. In some implementations, the output of the Colored-Noise-Like Waveform Synthesizer 122 can be processed locally or remotely and stored for use by the system 100. Whereas, in some implementations, the output of the Colored- Noise-Like Waveform Synthesizer 122 can be processed locally or remotely and provided to the Colored-Noise-Like Waveform Tx Beamformer 107 in real-time.

[0056] In various example implementations, while the Colored-Noise-Like Waveform Synthesizer 122 can use methods for synthesizing optimized noise-like waveforms, the system 100 does not require the noise-like waveforms to be optimized. In some implementations, the system 100 can utilize any analog or digital method that produces noise-like waveforms that are compatible with the hardware specifications of the Colored-Noise-Like Waveform Analog Exciter-Transmitters 108 to produce medical ultrasound images. For example, any digital random number generator or random number look-up table that describes the sequence of amplitudes and durations of the waveform to be produced as analog waveforms by the Colored-Noise-Like Waveform Analog Exciter-Transmitters 108 will be noise-like, and thereby can be utilized by the system 100 to produce medical ultrasound images. Alternatively, for example, any analog noise source that produces the sequence of amplitudes and durations of the waveform to be produced as analog waveforms by the Colored-Noise-Like Waveform Analog Exciter-Transmitters 108 (and15184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00that can be stored in memory by the Digital Signal Processor 104) will be noise-like, and thereby can be utilized by the system 100 to produce medical ultrasound images.

[0057] FIG. IB shows an example embodiment of a method 150 for synthesizing colored¬ noise-like waveforms for ultrasound imaging, in accordance with the present technology. The method 150 includes a process 152 to produce a set of one or more candidate seed waveforms based on at least one waveform parameter. The method 150 includes a process 154 to produce an optimized noise-like waveform set by iterating a set of trial waveforms using the set of one or more candidate seed waveforms and based on one or more optimization parameters. The method 150 includes a process 156 to select one or more output waveforms from the optimized noise-like waveform set based on one or more selection criteria to produce a set of noise-like waveforms, which based on their synthesis can possess properties akin to spread-spectrum ‘coded waveforms’. For example, the set of noise-like waveforms can be characterized as colored-noise-like or white¬ noise-like.

[0058] System 100 can be configured in one of many system designs. Some example embodiments of the system 100 are shown in FIGS. 1C, ID, and IE.

[0059] FIG. 1 C shows an example embodiment of an ultrasound system 100C that can provide some performance and / or cost advantages for some desired applications of the disclosed noise-like waveform ultrasound imaging technology. For example, the modular nature of System 100C allows configuring the system for an initial minimum number of transducer elements, I, in Transducer Array 118, while permitting the system to be expanded at some future date to accommodate larger Transducer Arrays 118 which have an increased number of transducer elements. As depicted in the diagram of FIG. 1C, the system 100C includes an embodiment of the Colored-Noise-Like Waveform Tx Beamformer 107 of system 100, which can be implemented as multiple, separate Tx Beamformer & Tx Controller (TxB-TxC) modules 107C to control Colored-Noise-Like Waveform Exciter-Transmitter Modules 108C to produce colored-noise-like, analog waveforms for transduction by Transducer Array 112. The Tx Beamformer & Tx Controller modules 107C can be configured as one or more (e.g., 1 to m), which are in communication with the one or more (e.g., 1 to m) Colored-Noise-Like Waveform Exciter-Transmitters 108C. In some embodiments, for example, the Tx Beamformer & Tx Controller modules 107C can be implemented in hardware using the type of processing unit required by a given application, such as for example either as one or more CPUs (central processing unit), GPUs (graphics processing16184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00unit), APUs (accelerated processing unit), TPUs (tensor processing unit), VPUs (vector processing unit), PLDs (programmable logic device), FPGAs (field-programmable gate array), QPUs (quantum processing unit), or other processing unit, which can be any combination thereof, or in combination as separate processing units or as a shared resource of the System Controller 102. In some embodiments, for example, the Colored-Noise- Like Waveform Exciter-Transmitter Modules 108C can include digital-to-analog converters (DAC). power amplifiers if required, and transducer bias circuits,

[0060] The system 100C also includes Analog Receiving (Rx) Modules 118, which can include one or more preamplifiers or signal conditioning circuitry, which is in signal communication with the exemplary T / R Switches 110 for processing the (transduced) received returned acoustic signals by the Transducer Array 112. In some embodiments, the Analog Rx Modules 118 may further include one or more analog-to-digital converters (ADC) in signal communication with the one or more preamplifiers or signal conditioning circuitry. The system 100C can include a Receiver (Rx) Beamformer & ADC Controller module 119, which is in communication with the Analog Rx Modules 118. e.g., that operates in conjunction with the DSP 104 of the system 100C to off-load a portion of the Digital Signal Processor 104 computational burden to increase the frame rate of the system.

[0061] FIG. ID shows an example embodiment of the Colored-Noise-Like Waveform Exciter-Transmitter Modules 108C depicted in the system 100C of FIG. 1C. For example, the Colored- Noise-Like Waveform Exciter-Transmitter Modules 108C can be implemented, as required by specific system design input requirements, as an assembly comprised of one or more Colored- Noise-Like Waveform Analog Exciter-Transmitters 108, previously discussed above. As in the system 100, the (optional) Output Amplifiers 109 can be used to match the requirements of the Transducer Array 112 technology, such as for example CMUT technology, selected to satisfy specific system design requirements for a given application.

[0062] FIG. IE shows an example embodiment of the Analog Rx Modules 118 depicted in the system 100C of FIG. 1C. For example, the Analog Rx Modules 118 can include example embodiments of the Pre-Amplifiers 116 and the A / D Converters 117, which can be controlled by one or more Rx beamformer & ADC controllers of the Rx Beamformer & ADC Controller module 119. For example, in a similar manner as the Tx Beamformer & Tx Controllers 107C, the Rx Beamformer & ADC Controllers 119 can be implemented in hardware by one or more of various17184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00types of processing units or as a shared resource of the System Controller 107, as the system design requires for a given application.

[0063] FIG. 2A shows a diagram illustrating an example embodiment of a method 200 for synthesizing colored-noise-like spread-spectrum coded waveforms, in accordance with the present technology.

[0064] For some embodiments, for example, the method 200 can be implemented by the Colored-Noise-Like Waveform Synthesizer 122 of the system 100, for example, which can produce one or more (digital) composite waveforms that are constituted by a Selected Set of Colored-Noise-Like Coded Waveforms 209 as shown in FIG. 2A. For example, the one or more composite waveforms (e.g., the Selected Set of Colored-Noise-Like Coded Waveforms 209) can be processed by the Colored-Noise-Like Waveform Tx Beamformer 107 of the system 100 to produce the digital waveform data that are transferred to the Analog Exciter-Transmitters 108 of the system 100 (e.g., which can be embodied with Multi-Level Pulser Transmitters, i.e., low 1 to 8-bit resolution, high-power DACs) for generating the analog colored-noise-like electronic waveforms, and, ultimately for the system 100 to initiate transmission of the Colored-Noise-Like Acoustic Waveform(s) 113 by the Transducer Array 112 of the system 100 toward a target, based on the analog colored-noise-like electronic waveforms generated by the system 100.

[0065] While colored-noise-like coded-waveforms can be formed by any number of synthesis techniques, the method 200 of FIG. 2A illustrates some example techniques. The method 200, also referred to as an example Colored-Noise-Like Waveform Synthesizing Process 200, can include three main processes: [1] a process to randomly produce and / or select candidate seed waveform(s) based on one or more waveform parameter(s) 201 (e.g., at process module 202, labeled Set of Random Seed Waveform Generator in FIG. 2A); [2] a process to optimize a multivariate waveform, e.g., iteratively, to produce a candidate optimized colored-noise-like waveform set 205 from the randomly-produced candidate seed waveform(s) based on one or more optimization parameter(s) 203 (e.g., at process module 204, labeled Waveform Multivariate Optimizer in FIG. 2A); and [3] a process to select one or more final waveforms from the optimized colored-noise-like waveform set based on selection criteria 207 to produce a set of colored-noise- like coded waveforms 209 (e.g., at process module 208, labeled Waveform Selector in FIG. 2A).

[0066] In implementations of the method 200, for example, the Random Seed Waveform Generator process module 202 randomly generates (e.g., creates and / or selects) a set (or sets) of18184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00candidate seed waveform(s) that are to be processed into waveforms in subsequent processes of the method 200. In implementations of the method 200, for example, the Waveform Multivariate Optimizer process module 204 produces a set (or sets) of candidate optimized colored-noise-like waveforms 205. In some embodiments, for example, the Waveform Multivariate Optimizer 204 can be a gradient-based heuristic or a gradient-free heuristic or another type of optimizer. In implementations of the method 200, for example, the Waveform Selector process module 208 produces a Set of Colored-Noise-Like Coded Waveforms 209, which is the output of method 200.

[0067] For some embodiments of the method 200. for example, to initiate the method 200, waveform parameter(s) 201 are specified and / or designed to provide the desired properties of the colored-noise-like waveforms that can be produced in accordance with the hardware requirements and limitations of the system 100. As an example, the waveform parameter(s) 201 can be specified as inputs by a user of the system 100 prior to implementation of the method 200. In other examples, the waveform parameter(s) 201 can be designed based on a protocol implanted by the system 100 where the system 100 (i) analyzes the hardware of the system 100 (e.g., DACs and / or LPAs and / or transducer(s) and / or other components) and / or (ii) interacts with a user of the system 100, e.g., via the User Interface 106, to probe for the desired imaging application and / or output specifications of the ultrasound images to be obtained by the system 100. and using the analyzed and / or probed information, produce the waveform parameter(s) 201. For example, the specifications of the waveform parameter(s) 201 can include, but are not limited to, analog output amplitude level, whether the output levels are unipolar or bipolar; sample update rate; waveform duration, sampling frequency, etc., e.g., which can be supported by Analog Exciter-Transmitters 108.

[0068] In some implementations of the process module 202, a set of randomly produced and / or selected seed waveform(s) is obtained to initiate the Waveform Multivariate Optimizer 204. For example, each of the seed waveform(s) in the set can be defined by a sequence of waveform levels (e.g., either unipolar or bipolar), by the waveform level durations (e.g., the number of intra-sample clock cycles), which is also termed pulse width, and / or by inter-amplitude sample intervals (e.g., the number of inter-sample clock cycles between amplitude samples) which is also termed pulse spacing,. For example, the seed waveform(s) for the set can be produced by a random number generator, selected from a data store of used or unused seed waveforms, or obtained by other means to ultimately produce a set of seed waveform(s) that is in accordance with waveform parameter(s) 201. The set of seed waveform(s) can include information that can be processed to digitally19184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00produce an individual waveform that meets criteria specified in the waveform parameter(s) 201.

[0069] In some implementations of the process module 204, the Waveform Multivariate Optimizer 204 inputs the one or more optimization parameter(s) 203 and iteratively processes the set of randomly produced and / or selected seed waveform(s) in a trial objective function based on the one or more optimization parameter(s) 203 to produce a set of candidate optimized colored- noise-like waveforms 205.

[0070] The optimization parameter(s) 203 can include, but are not limited to, for example, a maximum magnitude of first range sidelobe (hereinafter referred to as SLL, also known as peak side-lobe level or PSL), a peak-to-average-power ratio (PAPR), noise gain (NG, where noise gain represents the amplification of noise relative to signal), an average side-lobe level (related to the integrated side-lobe level by a scale factor, commonly referred to as ISLL), and / or other similar parameters). Optimization parameters can also include features of a discrete code sequence used to produce colored-noise-like coded waveforms, such as, but not limited to, non-zero first and last samples, a non-zero DC component (non-zero sum or non-zero mean value across all code samples), the probability of each quantized level (with a total probability of 1), a fixed occurrence (or count) of each quantized level, the probability of each position in the code for variability (with a total probability of 1), and a maximum and / or minimum number of consecutive levels.

[0071] During the development and optimization of colored-noise-like coded waveforms, it was discovered that optimized code sequences perform better when constrained to have a non-zero DC component (that is, a non-zero mean value across all code samples). This discovery emerged from empirical optimization studies wherein code sequences with zero DC component consistently exhibited suboptimal performance characteristics compared to code sequences with non-zero DC components. This non-zero DC component requirement finds support in the properties of known optimal codes from coding theory, specifically the Barker codes, which are well-established sequences with ideal autocorrelation properties for pulse compression. Of the nine known Barker codes, eight possess a non-zero DC component, with only the simplest length-two code comprised of [+1 -1] having zero DC component. This observation suggests that the non-zero DC component property may be a fundamental characteristic of codes with superior autocorrelation performance. While traditional white noise and colored noise signals in signal processing literature are typically defined as having zero DC component (zero mean), the present code sequence optimization context revealed that relaxing this constraint and instead enforcing a non-zero DC component on20184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00the discrete code sequence leads to improved optimization outcomes. Consequently, the optimization parameter(s) 203 can include an enforceable condition that requires candidate code sequences to maintain a non-zero DC component, wherein this condition can be implemented by constraining the sum of all code samples to be non-zero, or equivalently by requiring that the average value of the code sequence deviate from zero by a specified minimum threshold. It should be noted that this non-zero DC component requirement applies to the discrete code sequence itself, not to the final transmitted waveform, as the waveform generated from the code sequence may have zero DC component when modulated with a bipolar pulse (such as a single-cycle square wave pulse at the center frequency of transmission). This non-zero DC component requirement in the code sequence distinguishes the optimized colored-noise-like codes described herein from conventional noise-based code designs and contributes to achieving the superior sidelobe suppression and noise gain performance demonstrated in the optimization results.

[0072] The optimization parameter(s) 203 are used to generate a goal objective function that defines the characteristics of optimal colored-noise-like coded waveforms that are in accordance with the hardware limitations and requirements of the system 100. Put another way, for example, the goal objective function is used in a parameter generation process to produce the optimization parameters that can be used at the process module 204 (Waveform Multivariate Optimizer), which will ultimately produce digital waveforms that meet the ultrasound image features desired (e.g., a peak-to-first-sidelobe ratio (PFSR) to improve the contrast of the image, a signal-to-noise ratio (SNR) to improve the signal from the noise in the image, and / or a peak-to-average-sidelobe ratio (PASR) to reduce artifacts in the image and reduces blurriness of the features in the image). These inputs (i.e., optimization parameter(s) 203) can reflect the desired image quality produceable by the system 100. In addition, the optimization parameter(s) 203 can include additional, miscellaneous processing criteria, such as, e.g., the maximum number of iterations, the maximum allowable distance the trial objective function is from the goal objective function, etc. used by the Waveform Multivariate Optimizer 204.

[0073] For example, the process module 204 iterates trial waveforms, utilizing the seed waveform(s) with respect to inputted optimization parameter(s) 203 until a candidate waveform (or set of candidate waveforms) emerges from the Waveform Multivariate Optimizer 204 whose trial function is within the allowable distance from the objective function defined by the optimization parameter(s) 203. In some implementations, for example, for each candidate21184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00waveform produced by the Waveform Multivariate Optimizer, that candidate waveform can be stored in memory and becomes a member of the set of candidate optimized colored-noise-like waveforms 205.

[0074] In some implementations, for example, the processes 202 and 204 can repeat, where a new waveform seed or seeds is / are selected at the Random Seed Waveform Generator 202 and the Waveform Multivariate Optimizer 204 again optimization-processes the new waveform seed(s) based on the optimization parameter(s) 203 until a desired number of candidate waveforms is produced to complete the set of candidate optimized colored-noise-like waveforms 205.

[0075] The set of candidate optimized colored-noise-like waveforms 205 is provided to the process module 208 (Waveform Selector), which uses additional waveform selection criteria 207, such as, for example, Peak-to-Average-Power Ratio (PAPR), colored-noise-like spectral characteristics, ambiguity function characteristics, etc., to evaluate the candidate waveforms of the set of candidate optimized colored-noise-like waveforms 205. The Waveform Selector 208 determines which if any candidate waveforms of the set of candidate optimized colored-noise-like waveforms 205 meet the selection criteria 207 and outputs the waveforms that meet the criteria into a set of colored-noise-like coded waveforms 209. In some implementations of the method 200. the set of colored-noise-like coded waveforms 209 are outputted by the Colored-Noise-Like Waveform Synthesizer 122 of the system 100.

[0076] FIG. 2B shows a diagram illustrating an example embodiment of a method 250 to produce the waveform parameter(s) 201 used in the method 200, shown in FIG. 2A. For some implementations of the method 200, for example, the colored-noise-like coded waveforms can be synthesized by the method 200 in accordance with the method 250 shown in FIG. 2B. In some embodiments of the method 250, the waveform parameter(s) 201 that are outputted to the Random Seed Waveform Generator 202 (of the method 200) can be produced using an Analog Noise Source 210, an Analog Filter and Amplifier 211, an Analog Quantizer 212, and a Window 213. The Analog Noise Source 210 can include one of many different types of analog noise source devices, such as. for example, hot-cathode diode vacuum tubes, hot-cathode gas-discharge tubes, biased semiconductor diodes, biased avalanche diodes, biased Zener diodes, etc. The output analog noise signal from the Analog Noise Source 210 can be buffered, amplified, and / or filtered by the Analog Filter and Amplifier 211, whose characteristics are specified by an initial set of the Waveform Parameter(s) 201. The Analog Quantizer 212 converts a filtered / amplified output analog signal22184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00into a digital signal by discretizing the signal into a finite set of levels, which is then signal processed by Window 213. The output from Window 213 is outputted to both the Random Seed Waveform Generator 202, of method 200, and it can also be outputted to Digital Signal Processor (DSP) 104 (e.g., memory of or associated with DSP 104) for use by a filtering process (e.g., filtering process 402 discussed in FIG. 4) if required. In some implementations, for example, if the center frequency of the selected Analog Noise Source 210 is not high enough for the intended application of the system 100, it can be upconverted to a higher frequency. For such implementations, the method 250 can include an optional Frequency Upconverter 222 between The Analog Noise Source 210 and Analog Filter and Amplifier 211 to convert the outputted analog noise to the desired frequency range using any one of a number of well-known techniques.

[0077] FIG. 2C shows a diagram illustrating an example embodiment of a method 300 for producing colored-noise-like coded waveforms from selected waveform codes, in accordance with the present technology. In implementations of the method 300, for example, the colored -noise-like coded waveforms produced from the selected waveform codes (i.e., Selected Set of Waveform Codes 309) can be produced as a Set of Colored-Noise-Like Coded -Waveforms 319 that have similar properties to the waveforms produced by the method 200 shown in FIG. 2A. The method 300 can include a two-phase process that (i) develops optimized codes first and then (ii) produces colored-noise-like coded waveforms, which utilize the optimized codes, instead of the one-phase process of method 200 that directly produces optimized colored noise-like coded waveforms.

[0078] For some embodiments of the method 300, for example, to initiate the method 300, code parameter(s) 301 are specified and / or designed to provide the desired properties of the colored-noise-like waveforms that can be produced in accordance with the hardware requirements and limitations of the system 100. Similar to the method 200, in the method 300, for example, the waveform parameter(s) 301 can be specified as inputs by a user of the system 100 prior to implementation of the method 300. In some examples, the waveform parameter(s) 301 can be designed based on a protocol implanted by the system 100 where the system 100 (i) analyzes the hardware of the system 100 (e.g., DACs and / or LPAs and / or transducer(s) and / or other components) and / or (ii) interacts with a user of the system 100. e.g., via the User Interface 106. to probe for the desired imaging application and / or output specifications of the ultrasound images to be obtained by the system 100, and using the analyzed and / or probed information, produce the waveform parameter(s) 301. For example, the specifications of the code parameter(s) 301 can23184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00include, but are not limited to, the number of code levels, code length, whether the code is unipolar or bipolar: code length, the number of samples, etc,, that can be supported by Analog Exciter- Transmitters 108.

[0079] In implementations of the method 300, for example, a Random Seed Code Generator 302 randomly generates (e.g., creates and / or selects) a set (or sets) of candidate seed codes, which can be characterized by a series of code amplitudes and sample delays, which are to be processed into waveforms in a subsequent process of the method 300. In implementations of method 300, for example, a multivariate optimizer process module (referred to as Code Multivariate Optimizer 304) produces a set (or sets) of candidate optimized codes 305 from the set(s) of candidate seed codes that are based on code optimization parameter(s) 303. The Code Multivariate Optimizer 304 can iteratively optimize the inputted seed codes based on the code optimization parameter(s) 303 to produce the set(s) of candidate optimized codes 305 in a similar manner to the Waveform Multivariate Optimizer 204 in producing the set(s) of candidate optimized colored-noise-like waveforms 205, as discussed above. In some embodiments, for example, the Code Multivariate Optimizer 304 can be configured as a gradient-based heuristic, a gradient-free heuristic, or another type of optimizer.

[0080] Further, in implementations of the method 300, for example, a code selection process module (referred to as Code Selector 308) produces a set of waveform codes 309, based on code selection criteria 307. In some implementations, for example, the Code Selector 308 analyzes the optimized codes of the Candidate Optimized Code Set 305 with respect to the code selection criteria 307 to select a subset of codes that have more desirable properties than the other code(s) in accordance with waveform properties associated with the code selection criteria 307. For example, the code selection criteria 307 can specify code properties that produce desirable images, including but not limited to peak side-lobe and average side-lobe levels.

[0081] The set of waveform codes 309 produced by the Code Selector 308 can be used by a waveform generation process module (referred to as Waveform Generator 310) to produce the Set of Colored-Noise-Like Coded Waveforms 319 based on waveform parameters. In some implementations of the method 300, the waveform parameters used by the Waveform Generator 310 can be waveform parameter(s) 201, previously discussed in connection with the method 200. For example, in some implementations of the method 300, the Waveform Generator 310 can produce colored-noise-like coded waveforms by convolving a code impulse train with a defined24184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00waveform chip function of the desired period length. An example of a code impulse train is shown later in FIG. 6B. One example of a waveform chip function can be defined as having the value of 1 for t< T / 2 and 0 for T / 2<t< T.

[0082] FIG. 3 shows a data plot depicting an electrical sinusoidal excitation conversion to an acoustic output response for a typical medical ultrasound transducer array (e.g.. probe) whose nominal center frequency is 5 MHz. As shown by the data plot, there is a reduction in acoustic transductance conversion efficiency for frequencies away from approximately the relatively flat (± 0.5 db) 2 MHz portion of the transducer’s response.

[0083] For example, in some implementations when the Acoustic Medium 114 (shown in FIG. 1A-2) is soft mammalian tissue, the Acoustic Medium 114 may have an attenuation of acoustic waves between 0.5 - 0.75 db / MHz-cm attenuation in such media. Both transductance and acoustic medium attenuation phenomena can result in a reduction of resultant range resolution capability of colored-noise-like composite acoustic waveforms, e.g., as compared to mathematical predictions of unattenuated composite waveforms. To overcome the phenomena of the attenuation of higher frequency components of the set of colored-noise-like coded waveforms 209 like that in FIG. 2A, for example, the output synthesized waveforms from Colored-Noise-Like Waveform Synthesizer 122 can have the high and / or low-frequency components their spectrum emphasized by a filter whose transfer function is the inverse of the specific frequency transductance response of the Transducer Array 112 and Acoustic Medium 114.

[0084] Another example method that takes advantage of 1 to 8-bit. high-speed, high-power, high-voltage DACs is by using amplitude and phase-coded waveforms, e.g. like that described U. S. Patent Publication No. 2013 / 0123635 Al (previously incorporated by reference), as a substitute for the set of candidate seed waveforms produced by the Random Seed Waveform Generator 202 in FIG 2A. For example, since the specifications for amplitude and phase-coded waveforms are likely to have higher amplitude and time resolutions than specified by the waveform parameter(s) 201, their amplitude and time resolution can be decimated by truncating, rounding, or any other decimation process being used in Random Seed Waveform Generator 202. In such implementations, the amplitude and / or phase-coded waveforms are inputs into the Waveform Multivariate Optimizer 204, which implements the method 200 accordingly (e.g.. other process steps shown in FIG. 2A may remain unchanged).

[0085] The exemplary transducted acoustic waveform can be transmitted toward a Target25184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00Volume 115, e.g., biological tissue, and form a spatially combined acoustic waveform. The transmitted and transducted waveform can propagate through the Acoustic Medium 114 into the Target Volume 115. which, for example, can have one or more inhomogeneous mediums that partially transmit and partially reflect the transmitted acoustic waveform. The acoustic waveforms that are partially reflected also referred to as returned acoustic waveforms, can be received by Transducer Array 112. For example, each array element of I array elements of Transducer Array 112 can be configured to receive a returned acoustic waveform and convert it to an analog RF waveform signal.

[0086] If the I number of elements in the Transducer Array is greater than the F number of receiver channels comprised of Pre-Amplifiers 116 and A / D Converters 117, the individual received (analog) RF waveform signal returns can be multiplexed by an Optional MUX 111 to be communicated sequentially to the available receiver channels.

[0087] The individual received (analog) RF waveform signal returns can be modified by PreAmplifiers 116, which includes an array of F number of amplifiers, e.g., by amplifying the gain of a received waveform. The individual received waveforms can be converted from analog format to digital format by analog to digital (A / D) Converter module 117, which includes an array of I” number of A / D converters. A / D Converter module 117 can include A / D converters that have low least significant bit (LSB) jitter, spurious-free dynamic range (SFDR) and waveform dependency, such that the exemplary waveforms can be adequately decoded. The converted digital representations of the individual received composite waveforms can be processed by a processor, e.g., Digital Signal Processor 104, in a manner that creates and forms a representative image of the target medium.

[0088] FIG. 4 shows a diagram of an example embodiment of a digital signal processing (DSP) technique 400 that employs a signal conditioning process 401, a filtering process 402, a beamforming process 403, and / or an image processing process 404. In this example, the exemplary DSP technique 400 is also referred to as Colored-Noise-Like Waveform Filter Processing Technique 400.

[0089] In some embodiments of the DSP technique 400. the signal conditioning process 401 (also referred to as Pre-Conditioning process 401) can apply any number of desired DSP techniques to the converted digital waveforms produced by the A / D Converters 117 of the system 100. such as for example, level shifting, normalization, low-pass, bandpass, high-pass, or other26184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00filtering, windowing, and / or any other DSP common technique required for filtering a digital signal.

[0090] The Filtering process 402 can utilize any number of filters, such as a matched filter or other compressive filter. In some embodiments, the Filtering process 402 uses a DSP filter that is conjugate-matched to the specific Colored-Noise-Like Coded-Waveform 209, whose transducted Colored-Noise-Like Acoustic Waveform 113 that was transmitted through the Acoustic Medium 114 and that ensonified the Target Volume 115 and was reflected from the target volume and received. Most conventional filter DSP algorithms, including matched filtering algorithms, can be implemented in the Filtering process 402. In some embodiments, the Filtering process 402 implements a Range-Doppler Response algorithm to produce a Range-Doppler Response output. Examples of a Range-Doppler Response algorithm implemented at the Filtering process 402 are described in U. S. Patent Publication No. 2013 / 0123635 Al (previously incorporated by reference), e.g., at FIG. 5 and corresponding description.

[0091] The Beamforming process 403 can use any number of beamforming techniques depending on the Transducer Array 112 configuration, size, field-of-view (FOV), and System 100 mode of operation. For example, common beamforming techniques such as Delay and Sum (DAS), Bartlett, Capon, multi-line acquisition (MLA), multi-line transmission (MLT), plane wave, diverging wave, synthetic aperture, or other techniques can employ.

[0092] The Imaging Processing process 404 can be optionally employed after the Beamform 403 step to enhance the appearance of the resultant image. Imaging Processing process 404 is an optional step since it is not a required step for processing colored-noise-like waveforms, and it can include a variety of well-known image processing techniques such as logarithmic compression, spatial and frequency compounding, image noise reduction, and many other known techniques.

[0093] FIGS. 5A-5D show data plots depicting examples features that illustrate colored-noise-like waveforms produceable by implementations of the method 200 shown in FIG. 2A. In these examples, the Waveform Parameters 201 allowed the Waveform Multivariate Optimizer 204 to randomly vary not only the sequential amplitude chip levels of the waveforms but also randomly vary the duration of the sequential amplitude levels (pulse widths), commensurate with the Waveform Parameters 201 specifying a DAC clock rate. This random variation of the duration of the sequential amplitude levels includes the zero amplitude level samples, which can also be understood to be random varying intervals between amplitude chip levels.27184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00

[0094] FIG. 5A shows three examples of seed waveforms that are outputted from the Random Seed Waveform Generator 202 of the method 200. Each example seed waveform corresponds to a different Waveform Multivariate Optimizer 204 operation, with each operation using different Waveform Parameters 201. The three examples specified by the Waveform Parameters 201 are 31-chip long waveforms produced by 2-bit (3-level), 3-bit (5-level), and 4-bit (7-level) DAC s, that represent the output capabilities of different Colored-Noise-Like Waveform Analog Exciter-Transmitter 108 embodiment examples.

[0095] FIG. 5B shows three separate candidate optimized colored-noise-like waveforms that are outputted from the Waveform Multivariate Optimizer 204 (i.e., Candidate Optimized Colored-Noise-Like Waveform Set 205 examples) that can result from three different specifications for the Colored-Noise-Like Waveform Analog Exciter-Transmitters 108 in the Waveform Parameters 201.

[0096] FIG. 5C shows the resultant colored-noise-like coded waveform performance parameters for the three scenarios of the example candidate optimized colored-noise-like waveforms in three Optimized Colored-Noise-Like Waveform Sets 205 produced by the Waveform Multivariate Optimizer 204. Note that each data plot in FIG. 5C, for each scenario, shows a mean value of the performance parameters that is depicted by a circle and a standard deviation depicted by an ellipse surrounding the circle.

[0097] FIG. 5D shows examples of the result of the Waveform Selector 208 process that down¬ selects a specified small number of colored-noise-like coded waveforms out of the numerous members of the Candidate Optimized Colored-Noise-Like Waveform Set 205 based on the Selection Criteria 207 to form the Set of Colored-Noise-Like Waveforms 209. Note that each data plot in FIG. 5D, for each scenario, the dots represent the performance parameters of each selected colored-noise-like waveform: and the data plot shows a mean of these performance parameters depicted by a circle and a standard deviation depicted by an ellipse surrounding the circle.

[0098] It is worth noting that the 3-bit (5-level), and 3-bit (7-level) colored-noise-like coded- waveforms have significantly better performance characteristics than the 2-bit (3-level) waveforms. Some of the members of the 3-bit (5-level) waveform set exhibit side lobe levels below -60dB, and some of the members of the 3-bit (7-level) waveform set exhibit side lobe levels below -70dB, and none of the members of the 2-bit (3-level) waveform set exhibit side lobe levels below -50dB. The noise gain of the colored-noise-like waveforms produced by all three DACs in28184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00the examples are approximately -12 dB, which is not surprising since all of the waveforms are 31 chips long.

[0099] Also, notably, as shown in FIG. 5C, almost any arbitrary colored-noise-like waveform, which can also be characterized as a coded waveform, can provide adequate sidelobe levels and noise gains to enable ultrasound images to be produced. Hence, while the images produced by any arbitrary colored-noise-like transmitted wav eform may not be as desirable as those produced by employing optimizing methods, such as for example as methods 200 or 300, any colored-noise-like waveform can produce ultrasound images.

[0100] For example, for the number of codes analyzed (#-codes), the following table summarizes the performance that can be achieved by employing colored-noise-like coded waveforms that can be achieved with conventional, 1 - Bit pulsed mono-polar or bi-polar transmitted waveforms in embodiments of medical ultrasound systems.Table 1.1-Bit Binary 1-Bit Bipolar 2-Bit. 3-Levei 3-Bit, 5-Level 3-Bit. 7-Level Metric[0,1] [-1 1] [-1 0 1] [-2 -1 0 1 2] [-3 -2 -1 0 1 2 3] # codes 268,451,840 1,073,741,824 3,844 4,084 4,095 SLL -35.56 -47.56 -50.28 -67.10 -73.24 (dB)NG (dB) -9.39 -14.33 -13.87 -13.81 -13.47

[0101] As shown by Table 1, for example, the improved performance of the 3-bit, multi-level waveforms is markedly superior to the others, particularly with respect to SLL. A SLL value below -60 dB typically results in an artifact free or nearly artifact free ultrasound image.

[0102] Table 1 shows examples of the result of the Waveform Selector 208 process that down-selects a specified small number of colored-noise-like coded waveforms out of the numerous members of the Candidate Optimized Colored-Noise-Like Waveform Set 205 based on the Selection Criteria 207 to form the set of Colored-Noise-Like Waveforms 209.

[0103] FIGS. 6A-6C show data plots depicting three scenarios of example implementations for producing colored-noise-like coded waveforms from selected waveform codes, in accordance with aspects of the method 300 shown in FIG. 2C. Since a code can be defined by both a sequence of amplitudes plus a string of delays between amplitude changes, for example, FIGS. 6 A and 6B together specify example codes produced by the Code Multivariate Optimizer 304.29184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00

[0104] FIG. 6 A shows examples of the optimized code amplitudes produced by the Multivariate Optimizer 304 for three scenarios, which depict aspects of the candidate optimized code set 305 that is the output of Code Multivariate Optimizer 304. Scenario i is a 2-Bit 3-Level code. Scenario ii is a 3-bit 5-level code. And Scenario iii is a 3-Bit 7-Level code.

[0105] FIG. 6B shows examples of the optimized code delays (impulse trains) produced by the Multivariate Optimizer 304 for three scenarios, which depict aspects of the candidate optimized code set 305 that is the output of Code Multivariate Optimizer 304. Note the impulse train code delays can be uniform or random. The impulse train code delays represent the delay between amplitude level changes, which would be in the resultant Colored-Noise-Like Coded-Waveforms 319 produced by the Waveform Generator 310 as part of the method 300.

[0106] FIG. 6C illustrates the performance of the optimized codes after compression for three scenarios, which depict aspects of the candidate optimized code set 305 that is the output of Code Multivariate Optimizer 304. It should be noted that both the 3-Bit, 5-Level and the 3-Bit, 7-Level codes have lower sidelobe levels than the 2-Bit, 3-Level codes.

[0107] FIGS. 7A and 7B show data plots depicting three scenarios of example implementations of the Pre-conditioning process 401 and Filtering process 402 shown in FIG. 4, respectively, for signal conditioning and filtering of three different colored-noise-like coded- waveforms in their digital format.

[0108] FIG. 7A shows the received waveform outputs for three scenarios produced after Pre¬ conditioning 401 the received converted digital waveforms. In the FIG. 7 A example, scenario i is a 2-bit, 3-level waveform; scenario ii, is a 3-bit, 5-level waveform; and scenario iii is a 3-bit, 7- level waveform.

[0109] FIG. 7B shows three examples of the output from the Filtering 402 process for the same three example scenarios shown in FIG. 7A.

[0110] Referring back to FIG. 1A-2, in some implementations, the exemplary system 100 can be operated in one of many operation modes. In one example, Master Clock 101 can provide the time base for synchronizing the system 100, e.g., as a time base for the Colored-Noise-Like Waveform Tx Beamformer 107 and Analog Exciter Transmitters 108. Master Clock 101 can be configured as a low-phase noise clock such that the exemplary waveforms can be phase-encoded. An operator can select the mode of operation at User Interface 106. Exemplary modes of operation provided for the user to select at the User Interface 106 include Conventional A-Mode (e.g., ID30184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00Depth only image), Conventional B-Mode (e.g., 2D Plane image - transverse vs. depth), Conventional C-Mode (e.g., 2D Plane image at selected depth), and Conventional D-Modes (e.g., Doppler Modes). Exemplary Doppler modes include Color Doppler (e.g., superposition of color- coded Doppler and B-mode images), Continuous Doppler (e.g., 1D Doppler profile vs. depth), Pulsed Wave Doppler (e.g., Doppler vs. time for selected volume), and Duplex / Triplex Doppler (e.g.. superposition of Conventional B-Mode, Conventional C-Mode or Color Doppler, and Pulsed Wave Doppler). Some other exemplary modes of operations can include Conventional 3D and 4D (“real time” 3D) volume renderings of the previously described modes of operations. The exemplary system 100 can implement new modes of operation that can generate spread-spectrum, wide instantaneous bandwidth, coherent, colored-noise-like coded waveforms, in addition to frequency-and / or phase-coded waveforms. For example, a user can select exemplary ATS-Modes (Artificial Tissue Staining Modes) that can comprise a B-Mode, a C-Mode, a D-Mode, or other mode combined with image color coding to aid tissue differentiation - analogous to tissue staining for microscopic histological studies; and exemplary CAD-Modes (Computer Aided Diagnostic Modes) that differentiate and identify tissue type. ATS-Modes can employ the use of features for image color coding in image processing based on one or more of a number of measured properties that are obtained from the returned echo waveform from the target area, e.g., the returned echo from an exemplary transmitted spread-spectrum, wide instantaneous bandwidth, coded acoustic waveform. CAD-Modes can use classifiers (algorithms) to classify, for example, tissue types based on features of the measured properties of the returned echo from the target area, e.g., the returned echo from an exemplary spread-spectrum, wide instantaneous bandwidth, coded acoustic waveforms. The features properties can include differing impedances, amplitude reflections (as a function of wavelength), group delay, etc. Some exemplary classifiers that can be employed using CAD-Modes can include deterministic classifiers, stochastic classifiers (e.g., Bayesian classifiers), and neural network classifiers.

[0111] The disclosed systems and methods can employ the use of spread-spectrum, wide instantaneous bandwidth (up to 100% or more fractional bandwidth), coherent, colored-noise-like coded waveforms. There are limitless embodiments of such waveforms.

[0112] Noise-like waveforms can enable higher-order cross-range focusing techniques to be employed that can improve the lateral resolution of size-limited ultrasound transducer arrays, e.g., medical ultrasound transducer arrays.31184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00

[0113] For example, each biological tissue type and each diseased tissue types may exhibit their own unique ultrasound echo return as a function of frequency and spatial morphology. Using conventional Elastograph-Mode (E-Mode) modalities, it can be difficult to take advantage of such properties to classify tissues, e.g., due to measurement errors such as the inability to accurately characterize the ultrasound wave propagation through overlaying inhomogeneous media. Exemplary waveforms produced by the exemplary system 100, e.g.. wide instantaneous bandwidth, colored-noise-like waveforms, can enable tissue differentiation by simultaneously determining the propagation delay for each acoustic ray through intervening tissue layers and accurately determining the spatial echo features of the target volume under investigation. Classifiers, one example being Bayesian inference Classifiers among others, can be applied to the feature data obtained from the measured characteristics of the received echo to automatically classify tissue types observed in the target volume providing a Computer Aided Diagnostic-Mode (CAD-Mode).

[0114] Unlike conventional E-Modes, which inherently have significantly reduced image quality and rely on individual operator technique, the noise-like waveforms synthesized by the exemplary methods can inherently provide improved image quality while simultaneously colorizing the resultant image by tissue type in the ATS and / or CAD-Modes. With this advantage, user technique can be mitigated and the margins of a lesion are discernible thus permitting improved diagnoses.

[0115] Several applications and uses of the disclosed technology can be implemented to exploit the described features of the aforementioned systems, methods, and devices. Some examples are described for clinical use of the disclosed technology.

[0116] In one exemplary application, the resultant image quality, the ATS and CAD modes of an exemplary noise-like ultrasound device can enable the primary care physician to incorporate this modality into a routine examination screening protocol to locate early-stage malignancies (e.g., Stage 0 or 1), as well as later stage cancers. As a result of this application, the device can potentially, for example, enhance the survival rate of hard-to-diagnose asymptomatic patients suffering from malignancies such as stomach, pancreatic, and bladder cancers, etc.

[0117] In another exemplary application, the resultant image quality, ATS, and CAD modes of an exemplary noise-like ultrasound device can permit board-certified radiologists to diagnose neoplasms as benign or malignant prior to any surgical biopsy or resection intervention. As a32184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00result of this application, the ability of radiologists to locate and diagnose early-stage malignancies (e.g., Stage 0 or 1) can potentially improve patient survival rates. Additionally, unnecessary biopsies can potentially be avoided, along with their attendant risk of hard-to-treat or even lethal complications such as for example, methicillin-resistant staphylococcus aureus (MRSA staph) infections.

[0118] In another exemplary application, the resultant 3D image quality of an exemplary noise-like ultrasound device and its 4D imaging capability can be used in fine needle biopsy and other medical procedures. For example, the exemplary spread-spectrum ultrasound device can be integrated into an exemplary fine needle biopsy instrument (e.g., with the device’s transducer probe), which can permit the fine needle biopsy of very small, early-stage (e.g., Stage 0 or 1) neoplasms to confirm noninvasive diagnoses. As a result of this application, the ability of surgeons to avoid open biopsies and the potential for hard-to-treat and lethal complications that may result is clearly beneficial to the patient.

[0119] In another exemplary application, the integration of the ultrasound imagery produced by this noise-like waveform ultrasound device with high-definition video commonly used by minimally invasive, robotic surgical devices can permit the fusing of optical and ultrasound imagery to improve surgical outcomes. Such fused video and ultrasound images can give surgeons the added ability to locate and surgically excise diseased tissue without excising excessive healthy tissue.

[0120] In another exemplary application, given the improved 3D image quality of this spreadspectrum, noise-like waveform ultrasound device, its 4D imaging capability, and its ATS modes, an exemplary spread- spectrum ultrasound device can reduce the amount of time for the brachytherapy treatment of malignant neoplasms by precisely guiding the insertion of catheters and sealed radioactive sources into the proper location. The application of this spread-spectrum, noise-like ultrasound device to brachytherapy can be especially useful for the treatment of small, hard-to-locate neoplasms and their margins.

[0121] In another exemplary application, given the improved 3D image quality of this spreadspectrum. noise-like ultrasound device, its 4D imaging capability, and its ATS modes, an exemplary spread-spectrum ultrasound device can enable the effective insertion of high dose, localized pharmaceutical treatments of diseases by precisely guiding the insertion of catheters and pharmaceuticals into the proper location. The application of this spread-spectrum, colored -noise-33184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00like ultrasound device to brachytherapy can be especially useful for the treatment of small, hard- to-locate neoplasms.Example Embodiments and Implementations of the Code Multivariate Optimizer

[0122] Example embodiments and implementations are described below for the code multivariate optimizer, in accordance with the present technology.

[0123] FIGS. 8A-8F show data plots depicting the exemplary resultant colored-noise-like coded waveform performance parameters for code lengths of 32 (FIG. 8A), 48 (FIG. 8B), 64 (FIG. 8C), 96 (FIG. 8D), 128 (FIG. 8E), and 256 (FIG. 8F) and for 4096 total seeds. In each, four- scenarios of the example candidate optimized colored-noise- like waveforms are optimized (e.g., 1-Bit, Bi-polar; 2-Bit, 3-Level: 3-Bit, 5-Level, and 3-Bit, 7-Level) resulting in four Optimized Colored-Noise-Like Waveform Sets 205 produced by the Waveform Multivariate Optimizer 204. Note that each data plot in FIGS, 8A-8F, for each scenario, shows a mean value of the performance parameters that is depicted by a circle and a standard deviation depicted by an ellipse surrounding the circle.

[0124] Table 2 shows a statistical summary of the example data plotted in FIGS. 8A-8F.34184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00Table 2Code Bit / Mean Std Dev Mean Std Dev Unique Length Level SLL SLL NG (dB) NG (dB) Codes (dB) (dB)32 1-Bit Bi-polar -50.12 0.04 -13.84 0.00 2 32 2-Bit 3-Level -52.01 1.75 - 12.97 0.44 25 32 3-Bit 5-Level -69.03 1.46 -12.75 0.30 30 32 3-Bit 7-Level -71.06 3.83 -12.49 0.45 108 48 1-Bit Bi-polar -50.82 1.92 -14.63 0.59 216 48 2-Bit 3-Level -54.46 1.77 -14.65 0.40 1419 48 3-Bit 5-Level -66.20 4.29 -14.57 0.35 1109 48 3-Bit 7-Level -71.29 5.04 -14.6 0.31 1323 64 1-Bit Bi-polar -50.18 2.4 -16.05 0.52 3103 64 2-Bit 3-Level -55.58 1.95 -15.84 0.36 4023 64 3-Bit 5-Level -64.39 4.09 -15.81 0.32 4017 64 3-Bit 7-Level -68.24 4.24 -15.78 0.29 4060 96 1-Bit Bi-polar -47.92 1.76 - 17.96 0.37 4096 96 2-Bit 3-Level -55.60 1.82 -17.56 0.29 4096 96 3-Bit 5-Level -62.11 2.53 -17.53 0.26 4096 96 3-Bit 7-Level -65.13 2.86 -17.47 0.25 4096 128 1-Bit Bi-polar -47.55 1.31 -19.23 0.29 4096 128 2-Bit 3-Level -55.64 1.63 -18.81 0.26 4096 128 3-Bit 5-Level -61.67 2.21 -18.77 0.23 4096 128 3-Bit 7-Level -64.47 2.44 -18.68 0.21 4096 256 1-Bit Bi-polar -47.73 0.82 -22.19 0.19 4096 256 2-Bit 3-Level -55.61 1.19 -21.69 0.19 4096 256 3-Bit 5-Level -60.97 1.49 -21.67 0.17 4096256 3-Bit 7-Level -62.76 1.33 -21.59 0.16 4096

[0125] A Code Multivariate Optimizer 304 was employed to search the space of unique codes listed in Table 2 for lengths 32, 48, 64, 96, 128, and 256. Table 2, above, is a statistical summary of the scatter plots of optimized codes shown in FIG. 8A-8F. In each case, the pseudo-random number generator used in the optimizer was initialized with 4096 different random seeds in order to produce completely independent search patterns through the respective search spaces. The goal of the optimizer, for example, was to minimize peak side-lobe- level (SLL) and minimize noise gain (NG). In the objective function, each property was given approximately equal weighting to balance both objectives. The resulting optimized codes were filtered to remove non-unique versions including negated, reversed, and negated and reversed codes. Codes that belong in lower sets, e.g., including, but not limited to, bipolar codes in 3-, 5-, or 7-level codes, were also filtered out. All unique codes generated were plotted in FIG. 8A-8F.35184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00

[0126] As shown in Table 2, for example, the improved performance of the 3-bit, multi-level waveforms remains markedly superior to the others. In particular, achievable SLL values of less than -60 dB for 3-bit, 5- and 7-level codes for many hundreds of candidates across many code lengths are adequate for ultrasound imaging with minimal or imperceivable artifacts and minimal sacrifice in noise gain.

[0127] Also shown in Table 2 is the behavior of the Code Multivariate Optimizer 304 across search spaces of varying size and complexity. For shorter code lengths (32 and 48), independent initializations of the optimizer frequently converged to identical unique codes, providing statistical evidence that these solutions represent local minima in the search space. The repeated discovery of the same codes from different starting points suggests the optimizer is reliably identifying stable optima. For intermediate code length 64, convergence behavior shows a transition with some repeated solutions but increasing diversity. In contrast, for code lengths 96 and above, no convergence to identical solutions was observed across the 4096 independent runs. The absence of any repeated solutions in these larger search spaces suggests that the identified codes may not represent local minima, but rather reflect the vast size of the search space relative to the optimizer's exploration capability. As computational technology evolves and more sophisticated optimization techniques are developed, the identification and characterization of local and global minima in these nearly infinite search spaces will become increasingly feasible.

[0128] The Waveform Selector 208 can leverage the optimized code sets identified by the Code Multivariate Optimizer 304 to enhance ultrasound imaging performance across various operational scenarios. In some embodiments, the Waveform Selector 208 maintains a library of pre-optimized codes organized by code length and bit-level configuration, allowing selection based on specific imaging requirements such as penetration depth, frame rate, and image quality constraints. For example, when imaging shallow structures such as vascular anatomy or superficial tissues, the Waveform Selector 208 may select shorter codes (e.g., length 32 or 48) to minimize artifacts caused by simultaneous transmission and reception, wherein the shorter pulse duration reduces temporal overlap between the transmitted signal and received echoes from near-field structures, whereas when imaging deeper structures such as cardiac chambers or abdominal organs where simultaneous transmission and reception is not a concern, the Waveform Selector 208 may select longer codes (e.g., length 128 or 256) to maximize pulse compression gain and improve penetration depth and signal-to-noise ratio. The Waveform Selector 208 may select codes from36184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00the 3-bit, 5-level or 7-level sets when superior sidelobe suppression (SLL < -60 dB) is required for high-contrast imaging applications, while selecting codes from 2 -bit, 3-level sets when a balance between performance and implementation complexity is desired. In another embodiment, the Waveform Selector 208 may cycle through multiple different Candidate Optimized Colored-Noise-Like Waveforms from the optimized code sets across sequential transmit events, wherein corresponding images beamformed from each transmit event are combined using coherent addition (e.g., complex averaging) and / or incoherent combination (e.g., magnitude averaging or compounding) to suppress coherent artifacts, reduce speckle noise, and improve overall image quality. The use of different codes with varying sidelobes across transmit events causes coherent artifacts to appear at different locations in the respective beamformed images, such that coherent and / or incoherent combination causes these artifacts to average out or cancel while preserving the desired signal. The availability of hundreds to thousands of unique optimized codes within each set provides the Waveform Selector 208 with substantial flexibility to adapt waveform characteristics to specific clinical applications, imaging depths, or artifact suppression strategies without requiring real-time optimization.Example Embodiments and Implementations of the Machine Learning Techniques for Enhancing Waveform Optimization

[0129] Example embodiments and implementations are described below of the disclosed systems and methods employing machine learning techniques for enhancing waveform optimization, in accordance with the present technology.

[0130] In some implementations, machine learning techniques can be employed to enhance the waveform optimization process described in methods 150, 200 and 300. The disclosed machine learning approaches can learn from collections of optimized waveforms to generate new waveforms with desirable properties, such as low range sidelobe levels (SLLs) and low noise gain (NG), potentially reducing the computational burden of iterative optimization while maintaining or improving waveform quality.

[0131] In some embodiments, a machine learning-based system can be configured to train on datasets of optimized colored-noise-like waveforms that have been produced by the Waveform Multivariate Optimizer 204 (as described in method 200) and / or the Code Multivariate Optimizer 304 (as described in method 300). The training dataset can include waveforms with their associated performance metrics, including but not limited to: sidelobe level (SLL), peak-to-37184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00average-power ratio (PAPR), noise gain (NG), signal-to-noise gain (SNG), ambiguity function characteristics, and other relevant parameters described in optimization parameter(s) 203 and selection criteria 207. The machine learning-based system can be embodied as various embodiments of the system 1000, system 100, and system 100C described herein.

[0132] FIG. 9 shows a diagram of an example embodiment of a machine learning-based method 900 for generating optimized coded waveforms for ultrasound imaging. The method 900 includes a process 910 to train a machine learning model on a dataset that comprises waveform parameters and associated performance metrics, including at least one of a sidelobe level (SLL) or a noise gain (NG). The method 900 includes a process 920 to receive optimization parameters specifying target performance characteristics for a coded waveform. The method 900 includes a process 930 to generate, using the trained machine learning model, a candidate coded waveform predicted to meet the target performance characteristics. The method 900 includes a process 940 to output the candidate coded waveform for use in an ultrasound transmission.

[0133] Example embodiments and implementations of the machine learning-based system and machine learning-based methods, in accordance with the present technology, such as the method 900, are described below. Example features of the system 1000, system 100, and system 100C and of the method 150, method 200, method 250. and method 300 can be employed by the example machine learning-based system and example machine learning-based methods, such as the method 900, are described below.

[0134] For example, in some implementations, a neural network architecture can be employed as a generative model to learn the mapping between waveform parameters (such as the number of amplitude levels, pulse widths, pulse spacing, and waveform duration) and the resulting waveform performance characteristics. The neural network can be trained using supervised learning, where the input features include desired performance specifications (e.g., target SLL of -60 dB, target NG of -12 dB), and the output is a candidate waveform or waveform code that is predicted to meet those specifications.

[0135] In some implementations, for example, the machine learning model can comprise one or more of the following architectures: (1) a feedforward neural network that learns to predict waveform amplitude sequences and timing parameters based on desired performance metrics, wherein the network can include multiple hidden layers with nonlinear activation functions to capture complex relationships between input specifications and output waveform characteristics;38184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00(2) a recurrent neural network (RNN) or long short-term memory (LSTM) network that can capture temporal dependencies in waveform sequences, wherein these architectures are particularly suited for modeling the sequential nature of waveforms where the value at each time step may depend on previous time steps; (3) a convolutional neural network (CNN) that can learn spatial patterns in waveform representations, wherein CNNs can be effective at identifying features in the frequency domain representation of waveforms or in the ambiguity function characteristics; (4) a generative adversarial network (GAN) where a generator network learns to produce waveforms and a discriminator network learns to distinguish between optimized and sub-optimal waveforms, wherein the adversarial training process can drive the generator to produce increasingly high-quality waveforms that are indistinguishable from those produced by traditional optimization methods; (5) a variational autoencoder (VAE) that learns a latent representation of high-performing waveforms and can generate new waveforms by sampling from the learned latent space, wherein the VAE can be particularly useful for exploring the space of possible waveforms and discovering novel designs with unexpected performance characteristics; (6) a transformer¬ based architecture that can model complex relationships between waveform parameters and performance characteristics, wherein transformers can capture long-range dependencies and have shown strong performance in sequence generation tasks; (7) a large language model (LLM) or LLM-derived architecture that uses attention mechanisms and autoregressive generation to produce waveform sequences, wherein LLMs can be adapted for waveform generation by treating waveform amplitude levels and timing parameters as discrete tokens or continuous values in a sequence generation framework, and wherein pre-trained LLMs can potentially be fine-tuned for ultrasound waveform generation tasks or used as foundation models that leverage knowledge of sequential patterns learned from large-scale pre-training to improve waveform optimization performance with limited domain- specific training data; (8) a deep belief network (DBN) comprising stacked Restricted Boltzmann Machines that can learn hierarchical representations of waveform features through layer-wise unsupervised pretraining followed by supervised fine- tuning, wherein DBNs can be effective for learning complex feature hierarchies from waveform data and can serve as generative models or as feature extractors for subsequent classification or regression tasks; (9) an autoencoder network that learns compressed representations of waveforms through an encoder-decoder architecture, wherein autoencoders can be used for dimensionality reduction, feature extraction, anomaly detection in waveform characteristics, denoising of39184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00waveform parameters, or as building blocks for more complex architectures; ( 10) a self-organizing map (SOM) that provides unsupervised learning and visualization of high-dimensional waveform parameter spaces by mapping them to lower-dimensional topological representations, wherein SOMs can be useful for exploring the relationships between different waveform designs, identifying clusters of similar waveforms, and visualizing the distribution of optimized waveforms in parameter space: (11) a graph neural network (GNN) that can process graph-structured representations of waveform relationships, transducer array connectivity patterns, or acoustic propagation paths, wherein GNNs can capture spatial relationships between array elements or temporal relationships between waveform segments and can be particularly useful for optimizing waveforms for complex array geometries or for modeling interactions between multiple transmitted waveforms in multi-element systems; (12) a residual neural network (ResNet) architecture that uses skip connections to enable training of very deep networks, wherein ResNets can learn complex hierarchical representations of waveform characteristics and can be adapted for either classification tasks (evaluating waveform quality) or generation tasks (producing waveform sequences); (13) an attention-based neural network that uses attention mechanisms to focus on relevant parts of input waveform specifications or historical waveform data when generating new waveforms, wherein attention mechanisms can help the model identify which input features or which parts of the waveform sequence are most important for achieving desired performance characteristics; and / or (14) a diffusion model that generates waveforms through an iterative denoising process, starting from random noise and gradually refining it into a high-quality waveform that meets specified performance criteria, wherein diffusion models have shown strong performance in generative tasks and can produce diverse, high-quality outputs.

[0136] The machine learning models can be trained using one or more learning paradigms selected based on the availability of labeled training data, the specific optimization objectives, and the computational resources available. In supervised learning approaches, for example, the model learns from labeled examples where each input (desired waveform specifications) has a known target output (optimized waveform parameters and associated performance metrics), wherein supervised learning algorithms can include linear regression and logistic regression for modeling relationships between input specifications and output parameters, support vector machines (SVM) for classification of waveform quality categories or regression of performance metrics, k-nearest neighbors (k-NN) for predicting waveform characteristics based on similarity to known examples,40184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00decision trees and random forests for learning hierarchical decision rules that map specifications to waveforms, gradient boosting methods such as XGBoost. LightGBM, AdaBoost, and CatBoost for iteratively improving predictions by combining multiple weak learners, and neural networks and deep learning models as described above for learning complex nonlinear mappings. In unsupervised learning approaches, for example, the model discovers structure in unlabeled waveform data without explicit performance targets, wherein unsupervised learning algorithms can include clustering methods such as k-means clustering for grouping similar waveforms, hierarchical clustering for building dendrograms of waveform relationships, DBSCAN for density-based clustering that can identify outlier waveforms, and Gaussian Mixture Models for probabilistic clustering; dimensionality reduction methods such as Principal Component Analysis (PCA) for identifying principal directions of variation in waveform parameter space, t-SNE for visualizing high-dimensional waveform data in two or three dimensions, UMAP for scalable dimensionality reduction that preserves both local and global structure, and Independent Component Analysis (ICA) for separating waveform characteristics into independent components; and association rule learning methods such as Apriori, FP-Growth, and Eclat for discovering relationships between different waveform parameter combinations. In semi-supervised learning approaches, for example, the model leverages both a small set of labeled examples (optimized waveforms with known performance metrics) and a large set of unlabeled examples (waveform candidates without evaluated performance) to improve learning efficiency, wherein semisupervised learning techniques can include self-training and pseudo-labeling where the model iteratively labels unlabeled waveforms based on confident predictions and adds them to the training set, graph-based methods that propagate labels over a similarity graph connecting labeled and unlabeled waveforms based on their parameter similarity, and semi-supervised variants of deep neural networks that use consistency regularization (encouraging similar predictions for perturbed versions of the same waveform) or entropy minimization (encouraging confident predictions on unlabeled data). In reinforcement learning approaches, for example, the model learns to generate waveforms through interaction with an environment (which may be simulated or may involve actual ultrasound imaging) by receiving rewards based on waveform performance and learning a policy that maximizes cumulative reward, wherein reinforcement learning algorithms can include value-based methods such as Q-Learning, Deep Q-Networks (DQN), SARSA, Monte Carlo methods, and Temporal-Difference learning that learn to estimate the value41184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00of different waveform generation actions, and policy-gradient and actor-critic methods such as REINFORCE, Proximal Policy Optimization (PPO), Deep Deterministic Policy Gradient (DDPG), and Trust Region Policy Optimization (TRPO) that directly learn a policy for waveform generation. In ensemble learning approaches, for example, multiple base learners are combined to form a stronger overall model that often achieves better accuracy and robustness than individual models, wherein ensemble strategies can include bagging methods such as Random Forests that reduce variance by averaging predictions from many models trained on resampled data, boosting methods such as AdaBoost, Gradient Boosting. XGBoost, LightGBM, and CatBoost that sequentially focus on difficult examples to reduce bias, and stacking methods that learn a metamodel to optimally combine outputs from several base models trained with different algorithms or architectures.

[0137] In some implementations, the training process for the machine learning model can be performed offline, for example, using historical data from the Waveform Multivariate Optimizer 204 or Code Multivariate Optimizer 304. For example, a training dataset can be assembled by¬ running the multivariate optimizer multiple times with varying optimization parameters 203 and storing the resulting optimized waveforms along with their measured performance characteristics. The dataset can include thousands or millions of waveform examples spanning different hardware configurations (e.g., 2-bit 3-level, 3-bit 5-level, 3-bit 7-level DACs as described in Tables 1 and 2 and FIGS. 6A-6C and 8A-8F) and different optimization criteria.

[0138] In some implementations, the training dataset can be structured to include: input features comprising desired performance specifications (target SLL, target NG, target PAPR, hardware constraints such as number of amplitude levels, waveform duration, etc.); output targets comprising waveform amplitude sequences, pulse width sequences, pulse spacing sequences; and metadata comprising actual achieved performance metrics (measured SLL, measured NG, measured PAPR, etc.).

[0139] In some embodiments, for example, the training process can employ standard deep learning techniques including batch normalization to stabilize training, dropout regularization to prevent overfitting, learning rate scheduling to improve convergence, data augmentation techniques such as time shifting, amplitude scaling, or adding controlled noise to increase dataset diversity, and cross-validation to assess generalization performance.

[0140] In some implementations, the neural network training process can employ one or more42184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00optimization algorithms to minimize the loss function and update the network weights. These optimization algorithms can include, but are not limited to: (i) adaptive moment estimation (ADAM), which is an adaptive learning rate optimization algorithm that computes individual learning rates for different parameters based on estimates of first and second moments of the gradients, wherein ADAM is particularly effective for training deep neural networks and can handle sparse gradients and non-stationary objectives: (ii) stochastic gradient descent (SGD), which is a fundamental optimization method that updates weights using gradients computed on mini-batches of training data, wherein SGD can include momentum terms to accelerate convergence and escape local minima: (iii) RMSprop, which is an adaptive learning rate method that maintains a moving average of squared gradients to normalize the gradient descent steps and can be effective for recurrent neural networks: (iv) AdaGrad. which is an optimizer that adapts the learning rate based on the historical gradient information, giving frequently occurring features lower learning rates and infrequent features higher learning rates; (v) AdaDelta, which is an extension of AdaGrad that seeks to reduce its aggressive, monotonically decreasing learning rate by restricting the window of accumulated past gradients; and / or (vi) Nadam, which is a combination of ADAM: and Nesterov momentum that can provide faster convergence in some scenarios.

[0141] In some implementations, the optimization algorithms can further include evolutionary and population-based optimizers, including but not limited to: genetic algorithms, which are population-based optimization methods that use mechanisms inspired by biological evolution, including selection, crossover, and mutation operations, wherein genetic algorithms can be particularly useful for optimizing hyperparameters of the neural network or for evolving network architectures themselves (neuroevolution); particle swarm optimization (PSO), which is a population-based stochastic optimization technique where candidate solutions (particles) move through the parameter space influenced by their own best known position and the swarm's best known position; differential evolution, which is an evolutionary algorithm that optimizes a problem by iteratively improving candidate solutions with regard to a given measure of quality, using operations of mutation, crossover, and selection; and covariance matrix adaptation evolution strategy (CMA-ES), which is an evolutionary algorithm for difficult non-linear non-convex optimization problems that adapts the covariance matrix of a multivariate normal distribution to guide the search.43184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00

[0142] In some implementations, the optimization algorithms can further include hybrid and specialized optimizers, including but not limited to: Bayesian optimization, which is a strategy for optimizing hyperparameters or neural network architectures that builds a probabilistic model of the objective function and uses it to select the most promising hyperparameters to evaluate; limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS), which is a quasi-Newton method that approximates the Broyden-Fletcher-Goldfarb-Shanno algorithm using a limited amount of computer memory and can be effective for certain neural network training scenarios; and / or proximal policy optimization (PPO), which can be employed when using reinforcement learning approaches for waveform generation to optimize policies that generate waveforms based on reward signals derived from waveform performance metrics.

[0143] In some implementations, different optimization algorithms can be used for different stages of the training process. For example, ADAM may be used for initial training due to its fast convergence properties, followed by SGD with momentum for fine-tuning to achieve better generalization. Genetic algorithms may be employed for architecture search or hyperparameter optimization, while gradient-based methods like ADAM are used for weight optimization within a given architecture. The selection of optimization algorithm can depend on factors including the neural network architecture, the size and characteristics of the training dataset, computational resources available, and the specific requirements of the waveform generation task. In some embodiments, the system can automatically select or adapt the optimization algorithm based on training progress and performance metrics.

[0144] In some embodiments, for example, the machine learning model can be integrated into the Colored-Noise-Like Waveform Synthesizer 122 to work in conjunction with or as an alternative to the traditional optimization methods. For example, when a user specifies desired waveform characteristics through the optimization parameters 203, the trained machine learning model can rapidly generate one or more candidate waveforms predicted to meet those specifications.

[0145] These machine learning-generated candidates can then be utilized in several ways: (1) the candidates can be used directly as the set of colored-noise-iike coded waveforms 209 if they meet the selection criteria 207, wherein this approach provides the fastest waveform generation but requires high confidence in the machine learning model's predictions; (2) the candidates can be used as improved seed waveforms for the Random Seed Waveform Generator 202, providing44184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00better starting points for the Waveform Multivariate Optimizer 204 and potentially reducing the number of optimization iterations required, wherein this hybrid approach combines the speed of machine learning with the refinement capability of iterative optimization; or (3) the candidates can be refined through a reduced number of iterations by the Waveform Multivariate Optimizer 204 to fine-tune the waveform properties, wherein by starting from a high-quality machine learning¬ generated waveform, the optimizer can converge more quickly to an optimal solution.

[0146] Idle use of machine learning can provide several advantages over purely iterative optimization methods. The following are just some example advantages. First, once trained, the machine learning model can generate candidate waveforms in significantly less time than running the full iterative optimization process, which may be important for real-time or near-real-time waveform synthesis applications, wherein while the Waveform Multivariate Optimizer 204 may require hundreds or thousands of iterations to converge, a trained neural network can produce a candidate waveform in a single forward pass, typically requiring only milliseconds of computation time. Second, the machine learning model can learn non-obvious relationships between waveform parameters and performance characteristics that might not be captured by traditional objective functions, potentially discovering novel waveform designs with superior properties, wherein for example, the model might identify subtle patterns in pulse spacing that lead to unexpectedly low sidelobe levels, or discover amplitude sequences that provide better PAPR characteristics than those found through conventional optimization. Third, the model can generalize across different hardware configurations and optimization criteria, providing a flexible tool that adapts to varying system requirements without requiring separate optimization processes for each configuration, wherein a single trained model can potentially generate appropriate waveforms for 2-bit, 3-bit, or higher-bit DAC configurations by conditioning on the hardware specifications as input features. Fourth, by reducing or eliminating the need for iterative optimization, the machine learning approach can reduce the computational resources required by the system 100, potentially enabling waveform synthesis on less powerful computing hardware or reducing power consumption in portable ultrasound systems.

[0147] In some implementations, the machine learning model can employ transfer learning, where a model pre- trained on a large dataset of waveforms for one hardware configuration (e.g., 3-bit 5-level DACs) can be fine-tuned with a smaller dataset for a different configuration (e.g., 3- bit 7-level DACs), reducing the amount of training data and computational resources needed for45184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00each new system configuration. The transfer learning process can proceed as follows: (1 ) pre-train a base model on a large dataset of waveforms from one or more hardware configurations; (2) for a new hardware configuration, freeze the early layers of the network which learn general features; (3) fine-tune the later layers of the network using a smaller dataset specific to the new configuration; and (4) optionally, gradually unfreeze and fine-tune earlier layers if sufficient data is available. This approach is particularly valuable when adapting the system to new hardware configurations for which limited optimization data is initially available.

[0148] The machine learning approach can also incorporate active learning strategies, where the model identifies regions of the waveform parameter space where it has high uncertainty and requests additional optimization runs by the Waveform Multivariate Optimizer 204 in those regions. This targeted data collection can improve the model’s performance while minimizing the computational cost of generating training data. In some implementations, active learning can be implemented as follows: (1) train an initial model on available data; (2) generate predictions with uncertainty estimates (e.g., using ensemble methods or Bayesian neural networks); (3) identify parameter combinations where the model has high uncertainty; (4) run the Waveform Multivariate Optimizer 204 for these specific parameter combinations; (5) add the new optimized waveforms to the training dataset; (6) retrain or update the model; and (7) repeat the process iteratively. This approach ensures that computational resources for generating training data are focused on the most informative examples, leading to more efficient learning.

[0149] In some embodiments, the machine learning model can be continuously updated and improved through online learning, where new waveforms generated during system operation are added to the training dataset and used to periodically retrain or update the model. This allows the system to adapt to changing requirements or to incorporate improved waveform designs discovered through operation. For example, the system 100 can be configured to: (1) monitor the performance of all waveforms used in clinical imaging; (2) identify waveforms that perform particularly well or poorly in practice; (3) add successful waveforms to a growing training dataset; (4) periodically retrain the machine learning model (e.g., nightly or weekly); and (5) deploy the updated model to improve future waveform generation. This continuous learning approach enables the system to adapt to specific use cases, patient populations, or imaging protocols over time, potentially discovering domain- specific optimizations that were not apparent during initial training.46184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00

[0150] In some implementations, multiple machine learning models can be trained using different architectures, different subsets of training data, or different random initializations, and their predictions can be combined through ensemble methods. For example, a system could employ an ensemble of different neural network architectures (e.g., combining predictions from an LSTM, a CNN, and a transformer), bootstrap aggregating (bagging) where multiple models are trained on different random samples of the training data, or boosting approaches where models are trained sequentially to correct errors made by previous models. Ensemble methods can provide more robust predictions and better uncertainty estimates than single models, which is particularly valuable in the context of ultrasound imaging where waveform quality directly impacts patient care.

[0151] The performance of machine learning-generated waveforms can be validated by computing their actual performance metrics and comparing them to the predicted performance. Waveforms that meet or exceed the selection criteria 207 can be accepted for use in the system 100, while those that fall short can be used as feedback to improve the machine learning model. In some implementations, a validation framework can include automated performance testing wherein each machine learning-generated waveform can be automatically evaluated by computing its SLL. NG. PAPR, and other relevant metrics, and these measured values can be compared against both the target specifications and the model's predictions. The validation framework can further include statistical analysis wherein the distribution of performance metrics across many generated waveforms can be analyzed to assess the model's reliability and identify systematic biases or failure modes. The validation framework can further include A / B testing wherein machine learning-generated waveforms can be compared against conventionally optimized waveforms in actual imaging scenarios to assess their real-world performance. The validation framework can further include a feedback loop wherein waveforms that perform differently than predicted (either better or worse) can be added to the training dataset with their actual performance metrics, creating a feedback loop that continuously improves model accuracy.

[0152] In some implementations, the machine learning model can incorporate image quality metrics into the supervised learning feedback loop, wherein actual ultrasound images generated using the machine learning-generated waveforms are analyzed to quantify image quality characteristics that are then used to refine the model's predictions and improve future waveform generation. The image quality metrics can be computed by comparing images produced using the47184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00colored-noise-like coded waveforms against reference images, wherein the reference images can be produced using non-coded waveform transmissions (e.g., conventional pulsed waveforms) or other baseline waveforms. The image quality assessment can include quantification of artifacts, wherein artifacts such as range sidelobes, grating lobes, or other imaging artifacts can be detected and measured in the reconstructed ultrasound images, and the severity and spatial distribution of these artifacts can be quantified and associated with the specific waveform parameters that generated them. The image quality assessment can further include signal- to-noise ratio (SNR) measurements and generalized signal-to-noise ratio (gSNR) measurements, wherein the SNR can be computed as the ratio of mean signal intensity to the standard deviation of noise in a region of interest, and gSNR can be computed to provide a more robust measure that accounts for signal variability, and wherein these metrics can be computed for different regions of the image or for the entire image, and deviations from expected or target values can be identified. The image quality assessment can further include contrast (C) measurements, contrast-to-noise ratio (CNR) measurements, and generalized contrast-to-noise ratio (gCNR) measurements, wherein contrast quantifies the difference in echo intensity between different regions (such as between a target and background tissue), CNR quantifies the detectability of a target by accounting for both contrast and noise levels, and gCNR provides a generalized metric that is less sensitive to assumptions about signal statistics, and wherein these metrics are particularly applicable to assessment of anechoic cysts and other low-echogenicity targets where contrast detectability is critical for diagnostic performance. The image quality assessment can further include spatial resolution metrics, wherein axial resolution, lateral resolution, and / or elevational resolution can be measured using point targets, line targets, or other test objects in phantom studies or characterized from clinical images using edge sharpness analysis, speckle decorrelation, or other techniques, wherein speckle decorrelation measures the rate at which speckle patterns decorrelate with spatial displacement and provides a measure of the effective resolution of the imaging system. The image quality assessment can further include temporal consistency metrics, wherein the stability of image features across multiple frames or imaging sequences can be evaluated to assess waveform-induced variability.

[0153] The image-based feedback can be integrated into the machine learning training process through supervised learning, wherein the training dataset is augmented with image quality metrics as additional output targets or as components of a multi-objective loss function. For example, the48184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00neural network can be trained to predict not only the waveform characteristics (amplitude sequences, pulse widths, pulse spacing) but also the expected image quality metrics (predicted SNR, predicted artifact levels, predicted contrast ratio) that will result from using those waveforms. During training, the loss function can include terms that penalize deviations between predicted image quality metrics and actual measured image quality metrics, thereby encouraging the model to generate waveforms that produce high-quality images in practice. In some implementations, the image quality metrics can be weighted according to their clinical importance, wherein for example, artifact reduction may be weighted more heavily than minor SNR variations, or contrast enhancement may be prioritized for specific imaging applications such as tumor detection.

[0154] In some implementations, the system 100 can be configured to automatically acquire reference images using non-coded waveform transmissions and compare them to images acquired using machine learning-generated coded waveforms to compute image quality deviations. The comparison can include pixel-by-pixel or region-by-region analysis to identify spatial locations where the coded waveform produces superior or inferior image quality relative to the reference. Statistical measures of deviation can be computed, including but not limited to mean squared error (MSE), peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), or other image quality metrics. These deviation metrics can be stored along with the corresponding waveform parameters in the training dataset, enabling the machine learning model to learn the relationship between waveform design choices and their impact on actual image quality.

[0155] In some embodiments, image similarity metrics can be employed as quantitative measures for assessing the quality of images produced by machine learning-generated waveforms relative to reference images. The image similarity metrics can include, but are not limited to: structural similarity index (SSIM), which assesses the perceptual similarity between two images by comparing luminance, contrast, and structure components and provides a value between -1 and 1, where 1 indicates perfect similarity; normalized cross-correlation (NCC), which measures the correlation between two images normalized by their respective standard deviations and is particularly robust to linear intensity variations; sum of absolute differences (SAD), which computes the sum of the absolute differences between corresponding pixels in two images and provides a measure of dissimilarity that is computationally efficient; mean absolute error (MAE), which computes the average absolute difference between corresponding pixels; root mean squared49184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00error (RMSE), which computes the square root of the average of squared differences between corresponding pixels and is sensitive to large errors; correlation coefficient, which measures the linear relationship between pixel intensities in two images; mutual information (MI), which quantifies the amount of information that one image contains about another and is particularly useful for comparing images that may have nonlinear intensity relationships; feature similarity index (FSIM), which evaluates image similarity based on low-level feature maps such as phase congruency and gradient magnitude; visual information fidelity (VIF), which quantifies the information shared between reference and test images; and / or gradient magnitude similarity deviation (GMSD), which uses pixel-wise gradient magnitude similarity to assess image quality. In some implementations, multiple image similarity metrics can be computed simultaneously and combined (e.g., through weighted averaging or as components of a multi-dimensional quality vector) to provide a comprehensive assessment of image quality that captures different aspects of similarity and can be used as training targets or loss function components for the machine learning model.

[0156] In some implementations, the training data for the machine learning model can comprise completely simulated radiofrequency (RF) imaging data generated through ultrasound simulation software or computational models. The simulated RF data can provide precise control over noise characteristics, wherein specific noise levels, noise distributions, and noise correlations can be introduced in a controlled manner to train the model to handle various noise conditions. The simulation can model the complete imaging chain including transmit beamforming, acoustic wave propagation through tissue with specified acoustic properties (e.g., sound speed, attenuation, scattering), interaction with target structures, receive signal acquisition, and digital signal processing. The simulated data can be generated for diverse anatomical models, phantom geometries, and acoustic conditions to provide a comprehensive training dataset. The simulated RF data can be processed through a beamformer (either simulated or using actual beamforming algorithms that would be employed in the system 100) to form images that can be analyzed for quali ty metrics. The use of simulated data enables generation of large training datasets with known ground truth, perfect reproducibility, and systematic exploration of parameter spaces that may be difficult or impossi ble to achieve with physical experiments.

[0157] In some embodiments, noise-free simulated RF data can be generated and processed to produce noise-free reference images that enable quantification of artifacts caused specifically by50184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00the waveform characteristics, independent of system noise or other error sources. The noise-free simulation approach allows isolation of waveform-induced artifacts such as range sidelobes, off- axis scattering artifacts, or other image degradations that result from the autocorrelation properties of the transmitted waveform and its interaction with the imaging target geometry. By comparing images generated from coded waveforms against the noise-free reference image generated from an ideal impulse or non-coded reference waveform under identical noise-free conditions, the artifact contribution attributable solely to the waveform design can be precisely quantified. This quantification can include measurement of artifact amplitude, artifact spatial distribution, artifact texture characteristics, and other metrics that characterize how the waveform deviates from ideal imaging performance. These waveform- specific artifact metrics can be used as training targets or loss function components for the machine learning model, enabling the model to learn waveform designs that minimize artifacts independent of noise considerations, or to learn the trade-offs between noise performance and artifact performance.

[0158] The simulated RF data can be generated using known ultrasound simulation tools and methods, including but not limited to: Field II, which uses spatial impulse response methods to simulate ultrasound fields and is particularly efficient for simulating linear array transducers and can model arbitrary transducer geometries, apodization, focusing, and pulse characteristics; FOCUS (Fast Object-oriented C++ Ultrasound Simulator), which uses the fast-nearfield method for efficient calculation of ultrasound pressure fields and is well-suited for large-scale simulations with complex transducer arrays; k-Wave, which uses k-space pseudospectral methods to solve acoustic wave propagation equations and can model nonlinear propagation, heterogeneous media with spatially varying acoustic properties, and complex boundary conditions; analytical methods based on Helmholtz equations, which provide solutions for time-harmonic acoustic fields and can be used for frequency-domain analysis of ultrasound propagation; Rayleigh-Sommerfeld diffraction methods, which provide solutions for acoustic field propagation based on diffraction theory and are applicable to modeling transmit and receive beams; finite element methods (FEM), which discretize the acoustic domain and solve wave equations numerically, allowing modeling of complex geometries and material properties; finite difference time domain (FDTD) methods, which discretize both space and time to solve acoustic wave equations and can model transient phenomena and complex propagation conditions; and / or angular spectrum methods, which decompose acoustic fields into plane wave components and propagate each component according51184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00to its spatial frequency, enabling efficient modeling of beam propagation through layered media. The choice of simulation tool can depend on factors including computational efficiency requirements, the complexity of the acoustic medium being modeled, the accuracy requirements for the simulation, and the specific phenomena being studied (e.g., nonlinear propagation, attenuation, scattering).

[0159] In some implementations, the noise-free simulated data can be generated for a library of canonical imaging targets with known scattering properties, including but not limited to point targets at various depths and lateral positions, line targets at various orientations, resolution targets comprising arrays of point scatterers at known spacings, contrast-detail phantoms with targets of varying size and contrast, anechoic cyst phantoms with spherical or cylindrical void regions, hyperechoic lesion phantoms with regions of increased scattering, speckle-generating phantoms with random scatterer distributions that produce fully-developed speckle patterns, and / or anatomically realistic phantoms based on segmented medical image data. For each canonical target, noise-free images can be generated using ideal reference waveforms (e.g., Dirac delta impulses, short Gaussian pulses) and using candidate machine learning-generated coded waveforms. The differences between these images quantify the artifacts introduced by the coded waveform, and these artifact measurements can be associated with the waveform parameters in the training dataset. By training on a diverse library of target geometries, the machine learning model can learn to generate waveforms that produce minimal artifacts across a wide range of imaging scenarios.

[0160] In some embodiments, the machine learning model can be trained on many diverse image datasets to prevent overfitting to the content of any particular image and to ensure that the learned waveform generation strategies generalize across different imaging scenarios, anatomical structures, and acoustic conditions. The training approach can include image datasets from multiple distinct sources, wherein each dataset represents different imaging content, different acoustic properties, different scatterer distributions, and different target characteristics. The diverse datasets can include, but are not limited to: multiple phantom datasets with different phantom designs, scatterer concentrations, acoustic properties, and target arrangements; multiple anatomical region datasets representing different organs and tissue types (e.g,, liver, kidney, breast, thyroid, cardiac, vascular, musculoskeletal, obstetric), each with distinct echogenicity patterns, scattering characteristics, and structural features; multiple pathology datasets representing52184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00different disease states, lesion types, and abnormalities (e.g., cysts, solid masses, calcifications, fibrosis, inflammation) to ensure waveforms perform well for diagnostic imaging; multiple patient or subject datasets representing different body habitus, tissue compositions, ages, and physiological conditions to capture biological variability; multiple depth range datasets with targets at different depths to ensure waveform performance across the imaging field of view; multiple contrast level datasets with varying echogenicity differences to ensure waveforms maintain contrast resolution across different target conspicuity levels; and / or multiple speckle pattern datasets with different speckle textures, speckle signal-to-noise ratios, and speckle correlation properties. By training on this diverse collection of image datasets, the machine learning model learns waveform characteristics that are robust and generalizable rather than optimized for specific image content, thereby avoiding overfitting and ensuring reliable performance across the wide variety of imaging conditions encountered in clinical practice.

[0161] In some implementations, the training process can employ data augmentation techniques to artificially expand the diversity of datasets, wherein existing RF data or simulated data are systematically modified to create variations that expose the model to additional imaging scenarios without requiring collection of entirely new datasets. Data augmentation techniques applicable to RF data can include, but are not limited to; noise injection at various levels and with various statistical properties (e.g., Gaussian, Rayleigh, speckle noise, thermal noise) added to the RF signals to simulate different SNR conditions; synthetic target insertion wherein simulated targets with controlled properties are added to RF data through acoustic simulation to create additional training examples with known ground truth; random spatial cropping and subsampling of RF data to create training examples at different scales, aperture sizes, and field-of-view configurations; time-gating or windowing of RF signals to simulate different depth ranges or focal zones; channel dropout or element failure simulation wherein signals from selected transducer elements are removed or attenuated to simulate array malfunctions and train robust models: acoustic parameter variations in the simulation process such as sound speed perturbations, attenuation coefficient changes, or scattering strength variations to simulate different tissue types and imaging conditions; and / or transmit-receive aperture variations wherein different subsets of array elements are used for transmission and reception to simulate different imaging configurations. These augmentation techniques ensure that the machine learning model is exposed to a rich variety of RF signal conditions during training, further reducing the risk of overfitting to53184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00specific signal characteristics and improving the model's ability to generate high-quality waveforms across diverse clinical applications.

[0162] In some implementations, the training data can alternatively or additionally comprise stored RF imaging data acquired from actual ultrasound scans. The stored RF data can be obtained from in vivo imaging, wherein RF data is collected during clinical imaging procedures on human or animal subjects, providing realistic data that captures the complexity and variability of actual clinical conditions including tissue heterogeneity, motion artifacts, acoustic clutter, and patient-to-patient variations. The stored RF data can alternatively or additionally be obtained from ex vivo imaging, wherein RF data is collected from excised tissue samples or organs, allowing for controlled experimental conditions while maintaining realistic tissue acoustic properties. The stored RF data can alternatively or additionally be obtained from phantom imaging, wherein RF data is collected from tissue-mimicking phantoms with known acoustic properties and target geometries, enabling controlled and reproducible experiments with characterized image quality. The stored RF data can alternatively or additionally be obtained from calibrated imaging targets, wherein RF data is collected from precision test targets with known geometry, acoustic impedance, and scattering properties (e.g., wire targets, point targets, resolution targets, contrast detail targets) that enable quantitative assessment of imaging performance. The stored RF data can be collected using non-coded reference waveforms and machine learning-generated coded waveforms under identical imaging conditions to enable direct comparison and computation of image quality metrics.

[0163] In some implementations, the training data can comprise real-time RF imaging data that is acquired during operation of the system 100 and processed by a separate processor for machine learning training purposes. The separate processor can be configured to receive RF data from the system 100 (e.g., from the A / D Converters 116 or from the Digital Signal Processor 104) and perform beamforming, image formation, and image quality analysis in parallel with or independently from the primary imaging pipeline. The separate processor can be implemented as a dedicated computing device, a cloud-based computing resource, or as additional processing capability within the Computing Device 104 that is allocated specifically for machine learning training tasks. The real-time processing approach enables continuous learning and model improvement based on actual clinical data as it is acquired, allowing the system to adapt to the specific imaging conditions, patient populations, and clinical applications encountered during use.54184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00The separate processor can perform beamforming using the same algorithms employed in the primary imaging pipeline to ensure consistency, or can employ alternative beamforming strategies to evaluate different image formation approaches. In some implementations, the separate processor can generate multiple reconstructed images from the same RF data using different beamforming parameters or image formation techniques, enabling comparative analysis of how different processing choices interact with the waveform characteristics to affect final image quality.

[0164] In some implementations, the machine learning model training can be specific to a particular beamformer algorithm or beamformer architecture, wherein the training process accounts for the characteristics, limitations, and processing operations of the specific beamformer that will be used to form images from the transmitted waveforms. The beamformer- specific training enables the machine learning model to generate waveforms that are optimized for the particular signal processing and image formation methodology employed by the system 100, potentially discovering waveform characteristics that exploit strengths or compensate for limitations of the specific beamformer. The beamformer algorithms and architectures for which the training can be customized include, but are not limited to: analog beamformers, including phased analog beamformers with full dynamic focusing and steering capabilities that use analog delay lines or phase shifters to coherently combine signals from multiple transducer elements, and sequentially scanned or line-by-line analog beamformers with limited delay range that scan the beam across the field of view in a sequential manner with constraints on available delay values; digital delay-and-sum (DAS) beamformers, including time-delay beamformers that apply true time delays or interpolated delays to align signals from different elements before summation, and phaseshift beamformers that apply phase shifts in the frequency domain or use narrowband approximations to achieve beam steering and focusing; synthetic aperture beamformers, including simple synthetic aperture (SA) beamformers that synthesize a large effective aperture by combining data from multiple transmit-receive events, synthetic transmit aperture (STA) beamformers that use a single element or small subset for transmission and all elements for reception, and multi-line transmit (MLT) and multi-line receive (MLR) variants that simultaneously form multiple beams to increase frame rates; adaptive and coherence-based beamformers, including minimum variance beamformers (also known as Capon beamformers) that adaptively adjust weights to minimize output variance while maintaining unit gain toward the55184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00desired direction, general adaptive beamforming techniques such as linearly constrained minimum variance (LCMV) variants that impose multiple constraints on the beam pattern, coherence-factor- weighted DAS beamformers such as coherence factor (CF), generalized coherence factor (GCF), and phase coherence factor (PCF) that weight the beamformed signal based on spatial or temporal coherence of the received signals, and short-lag spatial coherence (SLSC) beamforming that forms images based on the spatial coherence of backscattered signals at short spatial lags: coded-excitation and matched-filter beamformers, including pulse-compression beamformers that use chirps, phase codes, or other coded waveforms with matched filtering to achieve improved SNR and range resolution, and coded synthetic aperture beamformers that combine synthetic aperture techniques with matched filtering for coded transmissions; plane-wave and diverging-wave beamformers, including beamformers for coherent compounding of multiple steered plane waves where images from different plane wave angles are coherently or incoherently combined to improve image quality, and diverging-wave or spherical-wave compounding beamformers that use unfocused or diverging transmissions followed by receive focusing and compounding: three- dimensional (3D) and matrix-array beamformers, including 2D matrix-array DAS beamformers that perform full 3D focusing with 2D transducer arrays, row-column or sparse-array beamformers that use reduced-element-count architectures with specialized addressing schemes, and 1.25D, 1.5D, or 1.75D elevation-steered beamformers that provide limited elevation focusing or steering capabilities; and specialized beamformer architectures, including parallel receive beamformers that form multiple receive beams simultaneously for each transmit event to increase frame rate or field of view, beamformers for capacitive micromachined ultrasonic transducer (CMUT) or piezoelectric micromachined ultrasonic transducer (PMUT) arrays that are often digital and designed for wideband operation, and sub-aperture or micro-beamforming architectures that perform partial beamforming in front-end application-specific integrated circuits (ASICs) to reduce channel count and data bandwidth requirements.

[0165] By training the machine learning model specifically for the beamformer architecture that will be used in the system 100, the model can learn waveform characteristics that are optimally matched to that beamformer's signal processing operations. For example, when training for a coherence-based beamformer, the model can learn to generate waveforms that maximize spatial coherence of desired signals while maintaining low coherence for clutter and noise, thereby improving the coherence-weighting effectiveness. When training for a synthetic aperture56184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00beamformer, the model can learn waveforms that account for the specific transmit-receive sequences and aperture synthesis operations employed. When training for plane-wave or diverging-wave beamformers, the model can learn waveforms that optimize performance for the unfocused or minimally focused transmission geometry. When training for adaptive beamformers, the model can learn waveforms that provide sufficient signal diversity and covariance matrix conditioning for effective adaptive weight computation. This beamformer-specific optimization enables the machine learning approach to achieve superior performance compared to generic waveform optimization that does not account for the specific image formation methodology.

[0166] In some implementations, the machine learning model can be configured to optimize multiple waveforms as an ensemble, wherein a set of two or more different waveforms are jointly optimized to work together in an imaging sequence to achieve better performance in image quality metrics than could be achieved using a single waveform or using multiple identical waveforms. This ensemble waveform optimization approach is particularly applicable to synthetic aperture imaging, multi-angle plane-wave imaging, coded excitation sequences, or other imaging modalities where multiple transmit events are used to interrogate overlapping or similar regions of the imaging space and the received signals are combined to form a final image. In the ensemble optimization approach, the machine learning model learns to generate complementary waveforms wherein artifacts, sidelobes, or other image degradations produced by at least one waveform in the ensemble are suppressed, cancelled, or reduced by at least one other waveform in the ensemble that interrogates a similar or overlapping image space. For example, in synthetic aperture imaging where multiple transmit events with different transmit aperture positions or configurations are used sequentially, the machine learning model can generate a sequence of waveforms wherein the range sidelobes or off-axis artifacts from one transmission are positioned differently in the image space compared to other transmissions, such that when the received signals are coherently combined, the artifacts from different transmissions interfere destructively while the main lobe signals interfere constructively, resulting in improved overall image quality with reduced artifacts and improved contrast.

[0167] The ensemble waveform optimization can account for spatial relationships between transmit events, wherein waveforms for neighboring transmit positions or adjacent transmit angles are designed to have complementary properties that enhance cancellation of artifacts when the signals are combined. For example, in a synthetic aperture sequence where transmit events57184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00sequentially scan across the array, the machine learning model can generate waveforms for adjacent transmit positions that have intentionally different sidelobe structures such that when signals from multiple transmit events illuminate the same scatterer from slightly different angles and are coherently summed, the sidelobe contributions average out or cancel while the main lobe contributions add coherently. Similarly, in multi-angle plane-wave imaging where steered plane waves at different angles are transmitted and the resulting images are compounded, the machine learning model can generate waveforms for different steering angles that have complementary artifact patterns, wherein artifacts that are prominent at one steering angle are minimized at other angles, such that the compounded image exhibits reduced artifacts compared to any individual plane-wave transmission. The ensemble optimization approach enables the machine learning model to exploit the diversity available in multi-transmission imaging sequences to achieve image quality improvements that would not be possible with single-waveform optimization.

[0168] In some implementations, the machine learning model can be trained using image quality metrics computed from the compounded or synthesized images formed by combining signals from multiple waveforms in the ensemble, rather than evaluating each waveform independently. The training objective can include minimizing artifacts in the final compounded image, maximizing contrast-to-noise ratio or generalized contrast-to-noise ratio in the final image, or optimizing other image quality metrics that characterize the final diagnostic image quality. During training, the machine learning model receives as input the specifications for the complete imaging sequence (number of transmit events, transmit positions or angles, aperture configurations, etc.) and generates as output a set of complementary waveforms optimized for that specific sequence. The model learns through supervised learning how different combinations of waveforms interact through the beamforming and compounding process to affect final image quality, enabling it to discover non-obvious relationships between waveform ensemble design and image quality that may not be apparent from single-waveform analysis. This training approach ensures that the generated waveforms are optimized not in isolation but as an integrated ensemble that works together to achieve superior imaging performance,

[0169] In some implementations, the training dataset can comprise a hybrid combination of simulated data, stored data from various sources (in vivo, ex vivo, phantom, calibrated targets), and real-time data, wherein each data source provides complementary information that contributes to training a robust machine learning model. For example, simulated data can be used to pre-train58184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00the model with a large diverse dataset under controlled conditions, stored phantom data can be used to validate the model's predictions against known ground truth, stored in vivo data can be used to expose the model to realistic clinical variability, and real-time data can be used for continuous adaptation and refinement. The training process can employ domain adaptation techniques to account for differences between simulated and real data or between different data sources, ensuring that the model generalizes effectively across all conditions. In some implementations, the training dataset can be stratified by data source, imaging application, anatomical region, or other relevant factors, and the machine learning model can be trained to condition its predictions on these factors, enabling application-specific or anatomy-specific waveform optimization.

[0170] The image-based feedback loop can enable the machine learning model to discover subtle relationships between waveform parameters and image quality that may not be apparent from waveform-level metrics alone. For example, the model may learn that certain pulse spacing patterns that appear optimal based on SLL and NG metrics actually produce undesirable artifacts in reconstructed images due to interactions with the beamforming process, acoustic propagation effects, or receive signal processing. Conversely, the model may discover that some waveforms with slightly suboptimal waveform-level metrics produce superior images due to favorable interactions with the complete imaging chain. This end-to-end learning approach, wherein the machine learning model is trained based on final image quality rather than intermediate waveform metrics alone, can lead to more clinically relevant waveform designs.

[0171] In some implementations, the image-based feedback can be collected during clinical use of the system 100, wherein images acquired during routine patient examinations are automatically analyzed to assess image quality, and this information is used to continuously improve the machine learning model through online learning. Privacy- preserving techniques can be employed to ensure patient data protection, wherein only anonymized image quality metrics (rather than the actual patient images) are retained for training purposes. The system can be configured to identify imaging scenarios where waveform performance is suboptimal (e.g., high noise, poor contrast, significant artifacts) and flag these cases for additional analysis or for generating new optimized waveforms specifically tailored to those challenging conditions.

[0172] The integration of image quality metrics into the machine learning feedback loop can also enable adaptive waveform selection, wherein the system 100 uses real-time image quality59184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00assessment to dynamically select among multiple candidate waveforms during an imaging session. For example, after transmitting a first coded waveform and analyzing the resulting image quality, the machine learning model can rapidly generate and evaluate alternative waveforms predicted to improve specific image quality deficiencies observed in the initial image. This closed-loop optimization approach can converge to optimal waveforms for the specific patient, anatomy, and imaging conditions within a small number of iterations (e.g., 2 to 5 frames), providing personalized and adaptive imaging performance.

[0173] In some implementations, a hybrid approach can be employed where the machine learning model generates an initial population of candidate waveforms, which are then refined through a smaller number of optimization iterations, combining the speed of machine learning with the precision of iterative optimization. For example, a hybrid method can proceed as follows: (1) in an ML generation phase, the machine learning model generates N candidate waveforms (e.g., N=10 to 100) based on the specified optimization parameters 203, wherein these candidates represent diverse starting points that the model predicts will lead to high-quality solutions: (2) in a preliminary evaluation phase, each candidate is evaluated to compute its actual performance metrics, wherein candidates that already meet the selection criteria 207 can be immediately accepted; (3) in a selective refinement phase, the top K candidates (e.g., K=5 to 20) that show the most promise but do not yet meet all criteria are selected for further optimization; (4) in a limited iteration optimization phase, each selected candidate is refined through a limited number of iterations (e.g., 10 to 50 iterations) by the Waveform Multivariate Optimizer 204, which is substantially fewer than the hundreds or thousands of iterations that might be required starting from random seeds; and (5) in a final selection phase, the refined waveforms are evaluated against the selection criteria 207, and those that meet the criteria are added to the set of colored-noise-like coded waveforms 209. This hybrid approach provides a balance between the computational efficiency of machine learning and the guarantee of meeting precise specifications provided by iterative optimization, and can be particularly effective when the machine learning model consistently generates waveforms that are "close" to optimal but require fine-tuning to meet strict performance requirements.

[0174] In some embodiments, the machine learning components can be implemented as part of the Colored-Noise-Like Waveform Synthesizer 122 within the Computing Device 104. The implementation can include trained neural network models stored in memory, along with60184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00associated preprocessing and postprocessing parameters; an inference engine comprising software or hardware accelerated inference engine (e.g.. using graphics processing units (GPUs), tensor processing units (TPUs), application-specific integrated circuits ( ASICs), field-programmable gate arrays (FPGAs), or specialized Al accelerators) to rapidly generate waveform predictions; optional connection to cloud-based or local training infrastructure for model updates and retraining; logging and monitoring systems to track the performance of machine learning-generated waveforms and identify opportunities for model improvement; and a management system for different model versions, allowing rollback to previous versions if a new model shows degraded performance. The computing device 104 can be configured with sufficient memory and processing capability to store and execute the machine learning models, which may range from relatively small feedforward networks with thousands of parameters to large transformer models with millions or billions of parameters, depending on the complexity of the waveform generation task and the available computational resources.

[0175] The integration of machine learning into the waveform synthesis process represents an advancement in adaptive and intelligent ultrasound imaging systems that can learn from experience and continuously improve their performance over time. This capability can provide several benefits for clinical applications including rapid protocol customization wherein clinicians can quickly generate optimized waveforms for new imaging protocols or specialized applications without waiting for lengthy optimization processes; personalized imaging wherein the system can potentially learn to generate waveforms optimized for specific patient characteristics (e.g., body habitus, tissue composition) based on feedback from previous imaging sessions; adaptive optimization wherein the system can adapt to changes in hardware performance over time (e.g., transducer aging) by continuously learning from operational data; reduced setup time wherein new ultrasound systems can be configured more quickly by leveraging machine learning models trained on data from similar systems, rather than requiring extensive per-system optimization; and enhanced image quality wherein by discovering non-obvious waveform optimizations, the machine learning approach can potentially achieve better image quality than conventional optimization methods, leading to improved diagnostic accuracy.

[0176] In some implementations, the system 100 can use machine learning to adapt waveforms in real-time during an imaging procedure. For example, if the system detects that current image quality is suboptimal (e.g., high noise levels or poor contrast), the machine learning model can61184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00rapidly generate alternative waveforms predicted to perform better under the current imaging conditions. This adaptation can occur within the time between successive image frames, enabling dynamic optimization without interrupting the imaging workflow.

[0177] When deploying the system 100 with a new transducer array or DAC configuration, the machine learning model can be fine-tuned using transfer learning from a base model trained on similar hardware. A relatively small number of optimization runs on the new hardware (e.g., 100 to 1000) can be sufficient to adapt the model, after which it can generate high-quality waveforms specific to the new configuration without further optimization.

[0178] In some applications, multiple conflicting objectives must be balanced (e.g., minimizing SLL while minimizing NG and maintaining acceptable PAPR). The machine learning model can learn to navigate these trade-offs by training on a diverse dataset that spans the Pareto front of optimal solutions. When a user specifies relative priorities for different objectives, the model can generate waveforms that appropriately balance these competing goals.

[0179] In some implementations, quantum computing can be employed to optimize waveform codes and generate optimized coded waveforms, wherein quantum algorithms leverage quantum mechanical phenomena such as superposition, entanglement, and quantum tunneling to explore the waveform parameter space more efficiently than classical computing approaches. Quantum computers operate on quantum bits (qubits), which are two-level quantum systems that can exist in superpositions of basis states |0 and |1), and quantum optimization algorithms encode the waveform code optimization problem in terms of qubit states and qubit interactions. For multi¬ level waveform codes (such as 2-bit 3-level, 3-bit 5-level, or 3-bit 7-level codes as described in Table 1 and FIGS. 6A-6C), the discrete amplitude levels at each time sample can be encoded using multiple qubits, wherein for example a ternary (3-level) amplitude can be encoded using 2 qubits, a 5-level amplitude can be encoded using 3 qubits, and a 7-level amplitude can be encoded using 3 qubits with appropriate state mapping. Alternatively, quantum systems with more than two levels, known as qudits (quantum digits), including qutrits (3-level quantum systems) and higher¬ dimensional qudits, can be employed where available to directly represent multi-level amplitude values without requiring multiple qubits per amplitude level. Quantum computers can be particularly advantageous for the discrete optimization problems inherent in code design, where the goal is to find optimal sequences of amplitude levels, pulse widths, and pulse spacings from a finite set of possibilities, wherein the combinatorial nature of these problems can lead to62184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00exponentially large search spaces that are computationally intractable for classical exhaustive search but may be amenable to quantum optimization algorithms. The quantum computing approach can be applied to optimize codes for colored-noise-like coded waveforms as described in methods 200 and 300, wherein the quantum computer searches for qubit configurations (representing code sequences) that minimize objective functions based on peak sidelobe level (SLL), noise gain (NG), peak-to-average- power ratio (PAPR). or other performance metrics, and wherein these objective functions are formulated in terms of the qubit states that encode the waveform parameters.

[0180] Quantum optimization algorithms that can be employed for waveform code generation include, but are not limited to: Quantum Annealing, which uses quantum tunneling to escape local minima and find global optima, wherein the waveform code optimization problem is encoded in a problem Hamiltonian expressed in terms of qubit interactions and biases, and wherein quantum annealers such as those based on superconducting flux qubits find low-energy qubit configurations that correspond to high-quality code sequences, and wherein the optimization problem can be formulated as a quadratic unconstrained binary optimization (QUBO) problem where binary decision variables are directly mapped to individual qubits; Quantum Approximate Optimization Algorithm (QAOA). which is a hybrid quantum-classical algorithm that uses parametrized quantum circuits operating on qubits to prepare candidate solution states and classical optimization to update circuit parameters, wherein QAOA applies a sequence of quantum gates to evolve the qubit state toward configurations that minimize the objective function encoded in a cost Hamiltonian, and wherein QAOA can be applied to combinatorial optimization problems including code sequence selection; Variational Quantum Eigensolver (VQE). which is a hybrid algorithm that uses quantum circuits to prepare trial qubit states and evaluate expectation values of cost operators, while classical optimization updates variational parameters, wherein VQE can be adapted for waveform optimization by encoding code quality metrics in the cost operator that acts on the qubits; Grover's Algorithm, which provides quadratic speedup for unstructured search problems by applying a sequence of quantum operations to amplify the amplitude of target qubit states, and wherein Grover's Algorithm can be applied to search through code sequences (encoded as qubit states) to find those meeting specified performance criteria; and Quantum Machine Learning algorithms that combine quantum computing with machine learning techniques, wherein quantum neural networks process information encoded in qubit states through parametrized63184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00quantum circuits, or quantum kernel methods compute similarities between data points using quantum feature maps that embed classical data into quantum state space, and wherein these algorithms can be trained to predict waveform performance from qubit-encoded parameters or generate optimized codes with potential computational advantages over classical approaches.

[0181] In some implementations, the quantum computing approach can be integrated with the classical machine learning techniques described herein through a hybrid quantum-classical framework. For example, a quantum computer can be used to perform the combinatorial optimization of discrete code parameters, wherein the amplitude level sequences for multi-level codes or binary phase sequences for phase-coded waveforms are encoded as configurations of multiple qubits, and the quantum optimizer searches for qubit states that minimize the objective function. Classical computers then decode the optimized qubit configurations back into waveform parameters, perform continuous parameter optimization (such as pulse width optimization or timing parameter adjustment), and execute the machine learning models for waveform performance prediction. The quantum optimizer generates candidate code sequences by finding low-energy qubit configurations, these sequences are decoded and evaluated using classical simulation or actual ultrasound imaging to compute performance metrics, and these metrics can be fed back to refine the quantum optimization process by updating the problem Hamiltonian or variational parameters. In some implementations, quantum annealing can be used to solve the Code Multivariate Optimizer 304 described in method 300, wherein the code optimization problem is mapped to a QUBO formulation where each binary decision variable (such as whether a particular amplitude level is selected at a particular time) is assigned to a qubit, the objective function is expressed as quadratic interactions between qubits plus linear biases on individual qubits, and the quantum annealer finds low-energy qubit states corresponding to high-quality code sequences that are then decoded into the actual waveform codes.

[0182] The use of quantum computing for waveform code optimization can provide several potential advantages. First, quantum algorithms may be able to explore larger code spaces or find better solutions in less time compared to classical optimization methods, particularly for problems where the objective function landscape has many local minima that trap classical optimizers, wherein quantum tunneling allows the quantum system to escape local minima by transitioning between qubit configurations through quantum mechanical processes that have no classical analog. Second, quantum optimization may discover novel code structures or unexpected code sequences64184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00that classical algorithms overlook due to getting trapped in local optima, potentially leading to waveforms with superior performance characteristics, wherein the quantum superposition of multiple qubit configurations enables simultaneous exploration of many candidate solutions. Third, as quantum computing hardware continues to advance, with increasing qubit counts, improved qubit coherence times, reduced error rates, and the potential development of qudit-based systems that can directly represent multi-level states, the integration of quantum optimization into waveform synthesis systems can provide a competitive advantage in ultrasound imaging performance. The quantum-optimized codes generated through this approach, obtained by¬ decoding the optimal qubit configurations found by the quantum computer, can be used directly in the system 100 as part of the set of colored-noise-like coded waveforms 209, or can be used as high-quality seed codes for further refinement through classical optimization or machine learning techniques.

[0183] In some implementations, the Quantum Fourier Transform (QFT) can be employed as part of the quantum algorithm for evaluating or optimizing autocorrelation properties of waveform codes. The QFT is the quantum analog of the discrete Fourier transform and can be implemented efficiently on a quantum computer with O(log²N) quantum gates for an N-point transform, compared to O(N log N) operations for the classical Fast Fourier Transform (FFT). The autocorrelation function and power spectrum of a sequence are related through the Fourier transform, wherein the power spectrum is the Fourier transform of the autocorrelation function, and conversely the autocorrelation function is the inverse Fourier transform of the power spectrum, and wherein for a code sequence, the power spectrum is |X(ω)|² where X(ro) is the Fourier transform of the sequence. In a quantum algorithm, the code sequence can be encoded in quantum state amplitudes as Σn xn|n, the QFT can be applied to compute the spectrum in quantum superposition, quantum operations can manipulate the amplitudes to effectively compute |X(ω)|² representing the power spectrum, and an inverse QFT can transform back to obtain autocorrelation information. Quantum amplitude estimation or other quantum measurement protocols can then be used to extract autocorrelation values at selected lags or to estimate global properties such as the total autocorrelation energy E = Σk Ck² with fewer quantum measurements than would be required by classical sampling methods. This QFT-based approach is particularly valuable when only certain lags or aggregate autocorrelation metrics are needed rather than the complete dense autocorrelation function, or when the autocorrelation energy or spectrum-based cost function must65184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00be evaluated many times during an iterative optimization process. The QFT can be integrated into variational quantum algorithms such as QAOA or VQE, wherein the cost function evaluation involves applying QFT to candidate code sequences encoded in quantum states, computing spectrum-related quantities, and measuring expectation values that guide the classical optimization of circuit parameters. For multi-level codes, the QFT can be generalized to operate on qudit states or on multi-qubit encodings of the code elements, enabling quantum-accelerated Fourier analysis of the more complex amplitude sequences used in the 3-level, 5-level, or 7-level coded waveforms described in this disclosure.

[0184] In some implementations, the quantum optimization objective can be formulated to achieve a flat or constant power spectrum, wherein a code sequence with a perfectly flat power spectrum X(ω)² = constant across all frequencies corresponds to an ideal impulse-like autocorrelation function with zero sidelobes at all non-zero lags, representing optimal coded waveform performance. The optimization cost function can be expressed as the variance or deviation of the power spectrum from a constant value, wherein minimizing the spectrum flatness metric Σω (|X(ω)|² - μ)² where it is the mean spectrum value is equivalent to minimizing the autocorrelation energy Σk Ck² through the Fourier transform relationship between autocorrelation and power spectrum. The quantum algorithm using QFT can evaluate the spectrum flatness by computing |X(ω)|² in quantum superposition and measuring the variance or other flatness metrics through quantum expectation value estimation, providing a direct assessment of how close the candidate code sequence is to the ideal constant-spectrum condition. For applications where the received waveform is processed with a filter other than the matched filter (which by definition matches the transmitted code), the optimization objective can be extended to consider crosscorrelation properties between the transmitted code and the receive filter impulse response, wherein the cross-correlation Rxy(k) = Σn xnyn+k between the code sequence x and the filter sequence y determines the range sidelobe structure after filtering. The quantum computer can encode both the code sequence and the filter response in quantum states, use QFT-based techniques to compute the cross-spectrum X(ω)Y*(ω) where Y(co) is the Fourier transform of the filter, and evaluate metrics based on the inverse Fourier transform of this cross-spectrum to assess the filtered output sidelobe characteristics. The optimization can then search for code sequences that produce flat or minimally-varying cross-spectra with the specified receive filter, or equivalently, minimal cross¬ correlation energy Σk Rxy(k)² for non-zero lags k. enabling the quantum algorithm to find codes66184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00optimized for specific receiver processing configurations beyond the standard matched filter case. This cross-correlation optimization is particularly relevant for systems where receive filtering must be constrained by hardware limitations, noise characteristics, or other system considerations that prevent use of the ideal matched filter, and wherein the quantum optimization can account for these practical filtering constraints in the code design process.Examples

[0185] The following examples are illustrative of several embodiments of the present technology. Other exemplary embodiments of the present technology may be presented prior to the following listed examples, or after the following listed examples.

[0186] In some embodiments in accordance with the present technology (example Al), a method for synthesizing noise-like waveforms for ultrasound imaging includes producing a set of one or more candidate seed waveforms based on at least one waveform parameter; producing an optimized noise-like waveform set by iterating a set of trial waveforms using the set of one or more candidate seed waveforms and based on one or more optimization parameters; and selecting one or more output waveforms from the optimized noise-like waveform set based on one or more selection criteria to produce a set of noise-like waveforms.

[0187] Example A2 includes the method of example Al or any of examples A 1-Al 5, wherein the set of noise-like waveforms includes coded waveforms that are coded based on at least one of colored-noise-like characteristics and / or white-noise-like characteristics.

[0188] Example A3 includes the method of example Al or any of examples A1-A15, further comprising: determining the at least one waveform parameter based on criteria including one or more hardware specifications.

[0189] Example A4 includes the method of example A3 or any of examples A1-A15, wherein the one or more hardware specifications includes a number of signed amplitude levels of waveforms to be synthesized.

[0190] Example A5 includes the method of example A4 or any of examples A1-A15, wherein the number of signed amplitude levels is in a range of 3 to 4096 levels for the waveforms to be synthesized.

[0191] Example A6 includes the method of example A4 or any of examples A1-A15, wherein number of signed amplitude levels is in a range of 5 to 4096 levels for the waveforms to be synthesized.67184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00

[0192] Example A7 includes the method of example A4 or any of examples A 1 -A 15, wherein the one or more hardware specifications includes a 3-bit digital-to-analog converter (DAC), and wherein the number of signed amplitude levels is in a range of 4 to 7 levels.

[0193] Example A8 includes the method of example A4 or any of examples Al -A 15, wherein the one or more hardware specifications includes fixed or variable durations of the signed amplitude levels of waveforms to be synthesized.

[0194] Example A9 includes the method of example A8 or any of examples A1-A15, wherein the variable durations of the signed amplitude levels include a random duration.

[0195] Example 10 includes the method of example Al or any of examples Al -A 15, wherein the producing the set of one or more candidate seed waveforms includes generating a set of random numbers, where each random number in the set of random numbers is generated by creating a nonzero output that spans a range based on an amplitude level and / or based on a time duration, to create one or more seed waveforms.

[0196] Example All includes the method of example Al or any of examples Al -Al 5, wherein the producing the set of one or more candidate seed waveforms includes selecting one or more seed waveforms from a data store.

[0197] Example A12 includes the method of example All or any of examples A1-A15, wherein the selecting the one or more seed waveforms includes accounting for whether a seed code to be selected has been used or not.

[0198] Example A13 includes the method of example A1 or any of examples A1-A15, wherein the producing the optimized noise-like waveform set includes iterating the set of trial waveforms utilizing the set of one or more candidate seed waveforms with respect to the one or more optimization parameters until a candidate waveform is determined to be within an allowable distance from an optimal objective function defined by the one or more optimization parameters.

[0199] Example A 14 includes the method of example A 1 or any of examples A 1-A 15, further comprising: using the set of noise-like waveforms in tomographic ultrasound imaging or B-Mode ultrasound imaging.

[0200] Example A15 includes the method of any of examples A1-A14, wherein the method is implemented by a system comprising an array of transducer elements to transmit an ultrasound signal at a target volume, one or more digital-to-analog converters (DACs) in communication with the array of transducer elements, a computing device that comprises a processor and a memory68184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00and is in communication with the one or more DACs, and one or more analog-to-digital converters (ADCs) in communication with the array of transducer elements and the computing device,

[0201] In some embodiments in accordance with the present technology (example A16), a method for synthesizing noise-like waveforms for ultrasound imaging includes producing a set of one or more candidate seed codes based on at least one code parameter; producing an optimized noise-like code set by iterating a set of trial codes using the set of one or more candidate seed codes and based on one or more code optimization parameters; selecting one or more codes from the optimized noise-like code set based on one or more code selection criteria to form a selected set of waveform-formation codes; and generating a set of noise-like waveforms by processing the selected set of waveform-formation codes with respect to one or more waveform parameters.

[0202] Example A17 includes the method of example A 16 or any of examples A16-A29, wherein the set of noise-like waveforms includes coded waveforms that are coded based on at least one of colored-noise-like characteristics and / or white-noise-like characteristics.

[0203] Example A18 includes the method of example A16 or any of examples A16-A29, further comprising: determining the at least one waveform parameter based on criteria including one or more hardware specifications.

[0204] Example A19 includes the method of example Al 8 or any of examples A16-A29, wherein the one or more hardware specifications includes a number of signed amplitude levels of waveforms to be synthesized.

[0205] Example A20 includes the method of example A19 or any of examples A16-A29. wherein the number of signed amplitude levels is in a range of 3 to 4096 levels for the waveforms to be synthesized.

[0206] Example A21 includes the method of example A19 or any of examples A16-A29, wherein number of signed amplitude levels is in a range of 5 to 4096 levels for the waveforms to be synthesized.

[0207] Example A22 includes the method of example A19 or any of examples A16-A29, wherein the one or more hardware specifications includes a 3-bit digital-to-analog converter (DAC), and wherein the number of signed amplitude levels is in a range of 4 to 7 levels.

[0208] Example A23 includes the method of example A18 or any of examples A16-A29, wherein the one or more hardware specifications includes fixed or variable durations of the signed amplitude levels of waveforms to be synthesized.69184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00

[0209] Example A24 includes the method of example A23 or any of examples A16-A29, wherein the variable durations of the signed amplitude level s include a random duration.

[0210] Example A25 includes the method of example A16 or any of examples A16-A29, wherein the set of noise-like waveforms are generated by, at least, convolving a code impulse train with a waveform chip function defined at least in part by a period length (T).

[0211] Example A26 includes the method of example A25 or any of examples A16-A29, wherein the waveform chip function is defined as having a value of 1 for time (t) < T / 2 and having a value of 0 for T / 2 < t < T.

[0212] Example A27 includes the method of example Al 6 or any of examples A16-A29, wherein the producing the optimized noise-like code set includes iterating the set of trial codes utilizing the set of one or more candidate seed codes with respect to the one or more optimization parameters until a candidate code is determined to be within an allowable distance from an optimal objective function defined by the one or more code optimization parameters.

[0213] Example A28 includes the method of example Al 6 or any of examples A16-A29, further comprising: using the set of noise-like waveforms in tomographic ultrasound imaging or B-Mode ultrasound imaging.

[0214] Example A29 includes the method of any of examples A16-A28, wherein the method is implemented by a system comprising an array of transducer elements to transmit an ultrasound signal at a target volume, one or more digital-to-analog converters (DACs) in communication with the array of transducer elements, a computing device that comprises a processor and a memory and is in communication with the one or more DACs, and one or more analog-to-digital converters (ADCs) in communication with the array of transducer elements and the computing device.

[0215] Example A29b includes the method of example A 16 or any of examples A16-A29, wherein the one or more code optimization parameters includes a constraint requiring a discrete code sequence to have a non-zero DC component, wherein the non-zero DC component is a non¬ zero sum or non-zero mean value across all samples in the discrete code sequence.

[0216] Example A30 includes the method of any of examples A16-A29 and A29b, wherein the method is implemented by a system comprising an array of transducer elements to transmit an ultrasound signal at a target volume, one or more digital-to-analog converters (DACs) in communication with the array of transducer elements, a computing device that comprises a processor and a memory' and is in communication with the one or more DACs, and one or more70184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00analog-to-digital converters (ADCs) in communication with the array of transducer elements and the computing device.

[0217] In some embodiments in accordance with the present technology (example A31), a system for synthesizing noise-like waveforms for ultrasound imaging includes an array of transducer elements to transmit an ultrasound signal at a target volume; one or more digital-to-analog converters (DACs) in communication with the array of transducer elements: a computing device, comprising a processor and a memory, in communication with the one or more DACs; and one or more analog-to-digital converters (ADCs) in communication with the array of transducer elements and the computing device, wherein the computing device is configured to generate one or more noise-like coded waveforms that possess a number of signed amplitude levels in a range of 3 to 4,096 levels.

[0218] Example A32 includes the system of example A31 or any of examples A31-A41, wherein the computing device is configured to: produce a set of one or more candidate seed waveforms based on at least one waveform parameter; produce an optimized noise-like waveform set by iterating a set of trial waveforms using the set of one or more candidate seed waveforms and based on one or more optimization parameters; and select one or more output waveforms from the optimized noise-like waveform set based on one or more selection criteria to produce a set of the one or more noise-like coded waveforms.

[0219] Example A33 includes the system of example A31 or any of examples A31-A41, wherein the computing device is configured to: produce a set of one or more candidate seed codes based on at least one code parameter; produce an optimized noise-like code set by iterating a set of trial codes using the set of one or more candidate seed codes and based on one or more code optimization parameters; select one or more codes from the optimized noise-like code set based on one or more code selection criteria to form a selected set of waveform-formation codes; and generate a set of the one or more noise-like coded waveforms by processing the selected set of waveform-formation codes with respect to one or more waveform parameters.

[0220] Example A34 includes the system of example A31 or any of examples A31-A41, wherein the one or more DACs includes one or more low-power DACs or a one or more high- power DACs.

[0221] Example A35 includes the system of example A31 or any of examples A31-A41, further comprising: one or more output amplifiers in communication with the array of transducer71184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00elements and the one or more DACs, wherein the one more output is configured to amplify an analog signal output of the one or more DACs.

[0222] Example A36 includes the system of example A35 or any of examples A31-A41, wherein the one or more output amplifiers includes one or more linear power amplifiers (LPAs).

[0223] Example A37 includes the system of example A31 or any of examples A31-A41, further comprising: one or more pre-amplifiers in communication with the array of transducer elements and the one or more ADCs, wherein the one more pre-amplifiers is configured to amplify a received ultrasonic signal that is returned from the volume of interest based on a transmitted ultrasound signal.

[0224] Example A38 includes the system of example A31 or any of examples A31-A41, wherein the one or more DACs are configured as a colored-noise-like waveform transmitter module, comprising: a plurality of individual transmitter beamformer and transmitter controller (TxB-TxC) modules configured to receive digitized instruction signals from the computing device; and a plurality of individual colored-noise-like waveform transmitter modules, each in communication with a corresponding individual TxB-TxC module, and configured to produce colored-noise-like, analog waveforms for transduction by the array of transducer elements based on a control from the plurality of individual TxB-TxC modules.

[0225] Example A39 includes the system of example A38 or any of examples A31-A41, further comprising: a plurality of analog receiving (Rx) modules, each comprising one or more preamplifiers and / or signal conditioning circuitry, configured to signal process received returned acoustic signals transduced by the array of transducer elements.

[0226] Example A40 includes the system of example A39 or any of examples A31-A41, further comprising: a receiving (Rx) beamformer & ADC controller module in communication with and configured to control the plurality of analog receiving (Rx) modules.

[0227] Example A41 includes the system of example A31 or any of examples A31-A41, wherein the system is configured to implement the method according to any of example A1-A15 or the method according to any of examples A16-A30.

[0228] In some embodiments in accordance with the present technology (example B42), a machine learning-based method for generating optimized coded waveforms for ultrasound imaging includes training a machine learning model on a dataset comprising waveform parameters and associated performance metrics including at least one of a sidelobe level (SLL) or a noise gain72184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00(NG); receiving optimization parameters specifying target performance characteristics for a coded waveform; generating, using the trained machine learning model, a candidate coded waveform predicted to meet the target performance characteristics; and outputting the candidate coded waveform for use in an ultrasound transmission.

[0229] Example B43 includes the method of example B42 or any of examples B42-B60), wherein the machine learning model comprises a neural network architecture selected from the group consisting of a feedforward neural network, a recurrent neural network, a convolutional neural network, a generative adversarial network, a variational autoencoder, and a transformerbased architecture.

[0230] Example B44 includes the method of example B43 or any of examples B42-B60, wherein the neural network architecture is a feedforward neural network comprising multiple hidden layers with nonlinear activation functions.

[0231] Example B45 includes the method of example B42 or any of examples B42-B60, wherein the dataset further comprises waveform parameters for multi-level coded waveforms including 2-bit 3-level codes, 3-bit 5-level codes, and 3-bit 7-level codes.

[0232] Example B46 includes the method of example B45 or any of examples B42-B60, wherein the performance metrics further include peak-to-average-power ratio and signal-to-noise gain.

[0233] Example B47 includes the method of example B42 or any of examples B42-B60, wherein training the machine learning model comprises using supervised learning with labeled examples where each input comprises desired waveform specifications and each target output comprises optimized waveform parameters and associated performance metrics.

[0234] Example B48 includes the method of example B47 or any of examples B42-B60), wherein the supervised learning employs an optimization algorithm selected from the group consisting of adaptive moment estimation (ADAM), stochastic gradient descent (SGD), root mean square propagation (RMSprop), adaptive gradient algorithm. (AdaGrad), and AdaDelta.

[0235] Example B49 includes the method of example B42 or any of examples B42-B60, wherein the dataset comprises simulated radiofrequency imaging data generated through ultrasound simulation software.73184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00

[0236] Example B50 includes the method of example B49 or any of examples B42-B60, wherein the simulated radiofrequency imaging data includes noise-free reference data enabling quantification of artifacts caused specifically by waveform characteristics.

[0237] Example B51 includes the method of example B42 or any of examples B42-B60, wherein the dataset comprises stored radiofrequency imaging data acquired from actual ultrasound scans.

[0238] Example B52 includes the method of example B51 or any of examples B42-B60, wherein the stored radiofrequency imaging data is obtained from in vivo imaging, ex vivo imaging, phantom imaging, or calibrated imaging targets.

[0239] Example B53 includes the method of example B42 or any of examples B42-B60, further comprising evaluating the candidate coded waveform by computing actual performance metrics and comparing the actual performance metrics to predicted performance metrics.

[0240] Example B54 includes the method of example B53 or any of examples B42-B60, further comprising refining the candidate coded waveform through a limited number of optimization iterations when the actual performance metrics do not meet selection criteria.

[0241] Example B55 includes the method of example B42 or any of examples B42-B60, wherein generating the candidate coded waveform comprises generating multiple candidate coded waveforms as an ensemble optimized to work together in an imaging sequence.

[0242] Example B56 includes the method of example B55 or any of examples B42-B60, wherein the ensemble comprises complementary waveforms wherein artifacts produced by at least one waveform in the ensemble are suppressed by at least one other waveform in the ensemble.

[0243] Example B57 includes the method of example B42 or any of examples B42-B60, wherein the SLL includes a maximum magnitude of first range sidelobe, also known as a peak sidelobe level, or an average sidelobe level.

[0244] Example B58 includes the method of example B42 or any of examples B42-B60, wherein the peak sidelobe level is a first range sidelobe, and wherein the average sidelobe level includes an integrated side-lobe level by a scale factor.

[0245] Example B59 includes the method of example B42 or any of examples B42-B60, wherein the NG is an amplification of noise relative to signal.74184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00

[0246] Example B60 includes the method of example B42 or any of examples B42-B59, wherein the associated performance metrics further includes a peak-to-average-power ratio (PAPR).Conclusion

[0247] Implementations of the subject matter and the functional operations described in this specification, such as various modules, can be implemented in digital electronic circuitry, or in computer software, firmware, or hardware, including the structures disclosed in this specification and their structural equivalents, or in combinations of one or more of them. Implementations of the subject matter described in this specification can be implemented as one or more computer program products, e.g., one or more modules of computer program instructions encoded on a tangible and non-transitory computer readable medium for execution by, or to control the operation of, one or more data processing apparatuses. The computer readable medium can be a machine- readable storage device, a machine-readable storage substrate, a memory device, a composition of matter affecting a machine-readable propagated signal, or a combination of one or more of them. The terms “data processing apparatus” or “data processing units” encompass all apparatus, devices, and machines for processing data, including by way of example a programmable processor, a computer, or multiple processors or computers. The apparatus can include, in addition to hardware, code that creates an execution environment for the computer program in question, e.g,, code that constitutes processor firmware, a protocol stack, a database management system, an operating system, or a combination of one or more of them.

[0248] A computer program (also known as a program, software, software application, script, or code) can be written in any form of programming language, including compiled or interpreted languages, and it can be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment. A computer program does not necessarily correspond to a file in a file system. A program can be stored in a portion of a file that holds other programs or data (e.g., one or more scripts stored in a markup language document), in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files that store one or more modules, sub programs, or portions of code). A computer program can be deployed to be executed on one computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected by a communication network.75184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00

[0249] The processes and logic flows described in this specification can be performed by one or more programmable processors executing one or more computer programs to perform functions by operating on input data and generating output. The processes and logic flows can also be performed by, and apparatus can also be implemented as, special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application specific integrated circuit).

[0250] Processors suitable for the execution of a computer program include, by way of example, both general and special purpose microprocessors, such as, for example, digital signal processors (DSP), and any one or more processors of any kind of digital computer. Generally, a processor will receive instructions and data from a read only memory or a random access memory or both. The essential elements of a computer are a processor for performing instructions and one or more memory devices for storing instructions and data. Generally, a computer will also include, or be operatively coupled to receive data from or transfer data to, or both, one or more mass storage devices for storing data, e.g., magnetic, magneto optical disks, or optical disks. However, a computer need not have such devices. Computer readable media suitable for storing computer program instructions and data include all forms of non volatile memory, media and memory devices, including by way of example semiconductor memory devices, e.g., EPROM, EEPROM, and flash memory devices. The processor and the memory can be supplemented by. or incorporated in, special purpose logic circuitry,

[0251] While this patent document contains many specifics, these should not be construed as limitations on the scope of any invention or of what may be claimed, but rather as descriptions of features that may be specific to particular embodiments of particular inventions. Certain features that are described in this patent document in the context of separate embodiments can also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments separately or in any suitable subcombination. Moreover, although features may be described above as acting in certain combinations and even initially claimed as such, one or more features from a claimed combination can in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.

[0252] Similarly, while operations are depicted in the drawings in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, to achieve desirable76184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO00results. In certain circumstances, multitasking and parallel processing may be advantageous. Moreover, the separation of various system components in the embodiments described above should not be understood as requiring such separation in all embodiments.

[0253] Only a few implementations and examples are described and other implementations, enhancements and variations can be made based on what is described and illustrated in this patent document.184667792.1

Claims

1. PCT Patent Application Attorney Docket No. 116291.8025. WO002.CLAIMS3.What is claimed is:

1. A method for synthesizing noise-like waveforms for ultrasound imaging, comprising: producing a set of one or more candidate seed waveforms based on at least one waveform parameter;5.producing an optimized noise-like waveform set by iterating a set of trial waveforms using the set of one or more candidate seed waveforms and based on one or more optimization parameters; and6.selecting one or more output waveforms from the optimized noise-like waveform set based on one or more selection criteria to produce a set of noise-like waveforms.

2. The method of claim 1, wherein the set of noise-like waveforms includes coded waveforms that are coded based on at least one of colored-noise-like characteristics and / or white-noise-like characteristics.

3. The method of claim 1, further comprising:9.determining the at least one w'aveform parameter based on criteria including one or more hardware specifications.

4. The method of claim 3, wherein the one or more hardware specifications includes a number of signed amplitude levels of waveforms to be synthesized.

5. The method of claim 4, wherein the number of signed amplitude levels is in a range of 3 to 4096 levels for the waveforms to be synthesized.

6. The method of claim 4, wherein number of signed amplitude levels is in a range of 5 to 4096 levels for the waveforms to be synthesized.

7. The method of claim 4, wherein the one or more hardware specifications includes a 3-bit digital-to-analog converter (DAC), and wherein the number of signed amplitude levels is in a range of 4 to 7 levels.14.78184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO008. The method of claim 4, wherein the one or more hardware specifications includes fixed or variable durations of the signed amplitude levels of waveforms to be synthesized.

9. The method of claim 8, wherein the variable durations of the signed amplitude levels include a random duration.

10. The method of claim 1, wherein the producing the set of one or more candidate seed waveforms includes generating a set of random numbers, where each random number in the set of random numbers is generated by creating a non-zero output that spans a range based on an amplitude level and / or based on a time duration, to create one or more seed waveforms.

11. The method of claim 1, wherein the producing the set of one or more candidate seed waveforms includes selecting one or more seed waveforms from a data store.

12. The method of claim 11, wherein the selecting the one or more seed waveforms includes accounting for whether a seed code to be selected has been used or not.

13. The method of claim 1, wherein the producing the optimized noise-like waveform set includes iterating the set of trial waveforms utilizing the set of one or more candidate seed waveforms with respect to the one or more optimization parameters until a candidate waveform is determined to be within an allowable distance from an optimal objective function defined by the one or more optimization parameters.

14. The method of claim 1, further comprising:23.using the set of noise-like waveforms in tomographic ultrasound imaging or B-Mode ultrasound imaging.

15. The method of any of claims 1-14, wherein the method is implemented by a system comprising an array of transducer elements to transmit an ultrasound signal at a target volume, one or more digital-to-analog converters (DACs) in communication with the array of transducer elements, a computing device that comprises a processor and a memory and is in communication with the one or more DACs, and one or more analog-to-digital converters (ADCs) in communication with the array of transducer elements and the computing device.25.79184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO0016. A method for synthesizing noise-like waveforms for ultrasound imaging, comprising: producing a set of one or more candidate seed codes based on at least one code parameter;28.producing an optimized noise-like code set by iterating a set of trial codes using the set of one or more candidate seed codes and based on one or more code optimization parameters; selecting one or more codes from the optimized noise-like code set based on one or more code selection criteria to form a selected set of waveform-formation codes; and29.generating a set of noise-like waveforms by processing the selected set of waveform¬ formation codes with respect to one or more waveform parameters.

17. The method of claim 16, wherein the set of noise-like waveforms includes coded waveforms that are coded based on at least one of colored-noise-like characteristics and / or white-noise-like characteristics.

18. The method of claim 16, further comprising:32.determining the at least one waveform parameter based on criteria including one or more hardware specifications.

19. The method of claim 18, wherein the one or more hardware specifications includes a number of signed amplitude levels of waveforms to be synthesized.

20. The method of claim 19, wherein the number of signed amplitude levels is in a range of 3 to 4096 levels for the waveforms to be synthesized.

21. The method of claim 19, wherein number of signed amplitude levels is in a range of 5 to 4096 levels for the waveforms to be synthesized.

22. The method of claim 19, wherein the one or more hardware specifications includes a 3-bit digital-to-analog converter (DAC), and wherein the number of signed amplitude levels is in a range of 4 to 7 levels.

23. The method of claim 19, wherein the one or more hardware specifications includes fixed or variable durations of the signed amplitude levels of waveforms to be synthesized.38.80184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO0024. The method of claim 23, wherein the variable durations of the signed amplitude levels include a random duration.

25. The method of claim 16, wherein the set of noise-like waveforms are generated by, at least, convolving a code impulse train with a waveform chip function defined at least in part by a period length (T).

26. The method of claim 25, wherein the waveform chip function is defined as having a value of 1 for time (t) < T / 2 and having a value of 0 for T / 2 < t < T.

27. The method of claim 16, wherein the producing the optimized noise-like code set includes iterating the set of trial codes utilizing the set of one or more candidate seed codes with respect to the one or more optimization parameters until a candidate code is determined to be within an allowable distance from an optimal objective function defined by the one or more code optimization parameters.

28. The method of claim 16, further comprising:45.using the set of noise-like waveforms in tomographic ultrasound imaging or B-Mode ultrasound imaging.

29. The method of claim 16, wherein the one or more code optimization parameters includes a constraint requiring a discrete code sequence to have a non-zero DC component, wherein the non-zero DC component is a non-zero sum or non-zero mean value across all samples in the discrete code sequence.

30. The method of any of claims 16-29, wherein the method is implemented by a system comprising an array of transducer elements to transmit an ultrasound signal at a target volume, one or more digital-to-analog converters (DACs) in communication with the array of transducer elements, a computing device that comprises a processor and a memory and is in communication with the one or more DACs, and one or more analog-to-digital converters (ADCs) in communication with the array of transducer elements and the computing device.48.81184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO0031. A system for synthesizing noise-like waveforms for ultrasound imaging, comprising: an array of transducer elements to transmit an ultrasound signal at a target volume; one or more digital-to-analog converters (DACs) in communication with the array of transducer elements;51.a computing device, comprising a processor and a memory, in communication with the one or more DACs; and52.one or more analog-to-digital converters (ADCs) in communication with the array of transducer elements and the computing device,53.wherein the computing device is configured to generate one or more noise-like coded waveforms that possess a number of signed amplitude levels in a range of 3 to 4,096 levels.

32. The system of claim 31, wherein the computing device is configured to:55.produce a set of one or more candidate seed waveforms based on at least one waveform parameter;56.produce an optimized noise-like waveform set by iterating a set of trial waveforms using the set of one or more candidate seed waveforms and based on one or more optimization parameters; and57.select one or more output waveforms from the optimized noise-like waveform set based on one or more selection criteria to produce a set of the one or more noise-like coded waveforms.

33. The system of claim 31, wherein the computing device is configured to:59.produce a set of one or more candidate seed codes based on at least one code parameter: produce an optimized noise-like code set by iterating a set of trial codes using the set of one or more candidate seed codes and based on one or more code optimization parameters;60.select one or more codes from the optimized noise-like code set based on one or more code selection criteria to form a selected set of waveform-formation codes; and61.generate a set of the one or more noise-like coded waveforms by processing the selected set of waveform-formation codes with respect to one or more waveform parameters.

34. The system of claim 31, wherein the one or more DACs includes one or more low-power DACs or a one or more high-power DACs.63.82184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO0035. The system of claim 31, further comprising:66.one or more output amplifiers in communication with the array of transducer elements and the one or more DACs, wherein the one more output amplifiers is configured to amplify an analog signal output of the one or more DACs.

36. The system of claim 35. wherein the one or more output amplifiers includes one or more linear power amplifiers (LPAs).

37. The system of claim 31, further comprising:69.one or more pre-amplifiers in communication with the array of transducer elements and the one or more ADCs, wherein the one more pre-amplifiers is configured to amplify a received ultrasonic signal that is returned from the volume of interest based on a transmitted ultrasound signal.

38. The system of claim 31, wherein the one or more DACs are configured as a colorednoise-like waveform transmitter module, comprising:71.a plurality of individual transmitter beamformer and transmitter controller (TxB-TxC) modules configured to receive digitized instruction signals from the computing device; and a plurality of individual colored-noise-like waveform transmitter modules, each in communication with a corresponding individual TxB-TxC module, and configured to produce colored-noise-like, analog waveforms for transduction by the array of transducer elements based on a control from the plurality of individual TxB-TxC modules.

39. The system of claim 38, further comprising:73.a plurality of analog receiving (Rx) modules, each comprising one or more preamplifiers and / or signal conditioning circuitry, configured to signal process received returned acoustic signals transduced by the array of transducer elements.

40. The system of claim 39, further comprising:75.a receiving (Rx) beamformer & ADC controller module in communication with and configured to control the plurality of analog receiving (Rx) modules.

41. The system of any of claims 31-40, wherein the system is configured to implement the method according to any of claims 1-15 or the method according to any of claims 16-30.77.83184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO0042. A machine learning-based method for generating optimized coded waveforms for ultrasound imaging, comprising:80.training a machine learning model on a dataset comprising waveform parameters and associated performance metrics including at least one of a sidelobe level (SLL) or a noise gain (NG);81.receiving optimization parameters specifying target performance characteristics for a coded waveform;82.generating, using the trained machine learning model, a candidate coded waveform predicted to meet the target performance characteristics; and83.outputting the candidate coded waveform for use in an ultrasound transmission.

43. The method of claim 42, wherein the machine learning model comprises a neural network architecture selected from the group consisting of a feedforward neural network, a recurrent neural network, a convolutional neural network, a generative adversarial network, a variational autoencoder, and a transformer-based architecture.

44. The method of claim 43, wherein the neural network architecture is a feedforward neural network comprising multiple hidden layers with nonlinear activation functions.

45. The method of claim 42, wherein the dataset further comprises waveform parameters for multi-level coded waveforms including 2-bit 3-level codes, 3-bit 5-level codes, and 3-bit 7-level codes.

46. The method of claim 45, wherein the performance metrics further include peak-to-average-power ratio and signal-to-noise gain.

47. The method of claim 42, wherein training the machine learning model comprises using supervised learning with labeled examples where each input comprises desired waveform specifications and each target output comprises optimized waveform parameters and associated performance metrics.

48. The method of claim 47, wherein the supervised learning employs an optimization algorithm selected from the group consisting of adaptive moment estimation (ADAM), stochastic90.84184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO0092.gradient descent (SGD), root mean square propagation (RMSprop), adaptive gradient algorithm, (AdaGrad), and AdaDelta.

49. The method of claim 42, wherein the dataset comprises simulated radiofrequency imaging data generated through ultrasound simulation software.

50. The method of claim 49, wherein the simulated radiofrequency imaging data includes noise-free reference data enabling quantification of artifacts caused specifically by waveform characteristics.

51. The method of claim 42, wherein the dataset comprises stored radiofrequency imaging data acquired from actual ultrasound scans.

52. The method of claim 51, wherein the stored radiofrequency imaging data is obtained from in vivo imaging, ex vivo imaging, phantom imaging, or calibrated imaging targets.

53. The method of claim 42, further comprising evaluating the candidate coded waveform by computing actual performance metrics and comparing the actual performance metrics to predicted performance metrics.

54. The method of claim 53, further comprising refining the candidate coded waveform through a limited number of optimization iterations when the actual performance metrics do not meet selection criteria.

55. The method of claim 42, wherein generating the candidate coded waveform comprises generating multiple candidate coded waveforms as an ensemble optimized to work together in an imaging sequence.

56. The method of claim 55, wherein the ensemble comprises complementary waveforms wherein artifacts produced by at least one waveform in the ensemble are suppressed by at least one other waveform in the ensemble.

57. The method of claim 42, wherein the SLL includes a maximum magnitude of first range sidelobe, also known as a peak sidelobe level, or an average sidelobe level.102.85184667792.1PCT Patent Application Attorney Docket No. 116291.8025. WO0058. The method of claim 57, wherein the peak sidelobe level is a first range sidelobe, and wherein the average sidelobe level includes an integrated side-lobe level by a scale factor.

59. The method of claim 42, wherein the NG is an amplification of noise relative to signal.

60. The method of claim 42, wherein the associated performance metrics further includes a peak-to-average-power ratio (PAPR).107.86184667792.1

Images