System and method for generating color Doppler images from short undersampled signal groups

Abstract

An ultrasound imaging system can acquire short and / or undersampled radio frequency (RF) signal groups for generating color Doppler images. The ultrasound imaging system can process short and / or undersampled signal groups to simulate color Doppler images acquired from long RF signal groups. In some examples, the ultrasound imaging system may include a neural network to process the signal groups. In some examples, the neural network may include two cascaded neural networks. In some examples, a power Doppler-based flux mask may be used on the output of the neural network during training. In some examples, adversarial loss may be used on the output of the neural network during training.

Description

Technical Field

[0001] This disclosure relates to imaging systems and methods for generating color Doppler images. In particular, it relates to imaging systems and methods for generating color Doppler images based on short, undersampled radio frequency (RF) signal ensembles. Background Technology

[0002] Traditionally, color Doppler (CD) has been used to diagnose atherosclerotic disease in extracranial arteries and to examine the hemodynamic properties of the heart. However, examining blood flow in the heart and narrowed arteries is often challenging due to the high and turbulent blood velocities.

[0003] CD images are generated by estimating the average phase shift within a group of ultrasound pulse signals (radio frequency (RF) signals) from the same sample volume. The most widely used phase shift estimation technique is lag-1 autocorrelation of slow-time RF signal pulses. Power Doppler (PD) is extracted from the lag-0 autocorrelation. CD provides information related to blood flow velocity (e.g., speed and direction). PD provides more sensitive blood flow detection than CD, but does not provide directional information. By applying a high-pass filter (e.g., a wall filter) along the slow-time direction, it is possible to separate the moving blood echo from the tissue background.

[0004] One determining factor of CD signal quality is the number of pulse / slow observations in the signal group (RF group size) used to generate a single CD image. A larger RF group size provides greater accuracy and sensitivity for CD blood flow velocity estimation compared to a smaller RF group size. Another CD quality factor is the pulse repetition frequency (PRF) of the signal group pulses. The PRF determines the time sampling rate of the Doppler signal and the highest blood flow velocity that can be recovered. Summary of the Invention

[0005] As disclosed herein, artificial intelligence (e.g., deep learning) can be used to correlate short / decimated signal groups in a color Doppler image with longer signal groups and / or signal groups with higher PRF. In this way, a high-quality color Doppler image can be generated based on the short / decimated signal groups, which mimics the appearance of a color Doppler image generated based on the longer signal groups and / or signal groups with higher PRF.

[0006] An example ultrasound imaging system according to this disclosure may include a processor configured to: receive an ultrasound signal corresponding to a first radio frequency (RF) signal group, the first RF signal group including a first length, a first pulse repetition frequency (PRF), and a first sensitivity; estimate a second RF signal group based at least in part on a reference ultrasound signal corresponding to at least one reference RF signal group, the at least one reference RF signal group including at least one of a second pulse repetition rate, a second length, or a second sensitivity different from the first RF signal group; and use the second RF signal group to generate a color Doppler image.

[0007] An example method according to this disclosure may include: receiving an ultrasound signal corresponding to a first radio frequency (RF) signal group, the first RF signal group including a first length, a first pulse repetition frequency (PRF), and a first sensitivity; estimating a second RF signal group based at least in part on a reference ultrasound signal corresponding to at least one reference RF signal group, the at least one reference RF signal group including at least one of a second pulse repetition rate, a second length, or a second sensitivity different from the first RF signal group; and using the second RF signal group to generate a color Doppler image.

[0008] According to an example of this disclosure, a non-transient computer-readable medium may contain instructions that, when executed, cause an imaging system to: receive an ultrasound signal corresponding to a first radio frequency (RF) signal group, the first RF signal group including a first length, a first pulse repetition frequency (PRF), and a first sensitivity; estimate a second RF signal group based at least in part on a reference ultrasound signal corresponding to at least one reference RF signal group, the at least one reference RF signal group including at least one of a second pulse repetition rate, a second length, or a second sensitivity different from the first RF signal group; and use the second RF signal group to generate a color Doppler image. Attached Figure Description

[0009] Figure 1 Example CD images of long signal groups, short signal groups, and decimated signal groups are shown.

[0010] Figure 2 This is a block diagram of an ultrasound system based on the principles of this disclosure.

[0011] Figure 3 This is a block diagram illustrating an example processor based on the principles of this disclosure.

[0012] Figure 4 This is a block diagram illustrating the process of training and deploying a neural network based on the principles of this disclosure.

[0013] Figure 5 An overview of the training and deployment phases of a neural network based on the principles of this disclosure is shown.

[0014] Figure 6 The illustration shows an example workflow for generating a training dataset based on the principles of this disclosure.

[0015] Figure 7 It is a diagram of a neural network based on the principles of this disclosure.

[0016] Figure 8 This is a flowchart illustrating a method for using optional masking loss based on the principles of this disclosure.

[0017] Figure 9 This is a flowchart of a method based on the principles of this disclosure.

[0018] Figure 10 Example extracted signal group color Doppler images and example enhanced color Doppler images are shown in accordance with the principles of this disclosure.

[0019] Figure 11 Examples of short signal groups, long signal groups, and color Doppler images of enhanced original phase shift results are shown in accordance with the principles of this disclosure.

[0020] Figure 12 Examples of short signal groups, long signal groups, and color Doppler images of enhanced original phase shift results are shown in accordance with the principles of this disclosure.

[0021] Figure 13 Examples of short signal groups, long signal groups, and enhanced post-processed color Doppler images based on the principles of this disclosure are shown.

[0022] Figure 14 Examples of short signal groups, long signal groups, and enhanced post-processed color Doppler images based on the principles of this disclosure are shown. Detailed Implementation

[0023] The following description of certain embodiments is merely exemplary in nature and is in no way intended to limit the invention or its application or use. In the following detailed description of embodiments of the system and method, reference is made to the accompanying drawings, which form part of this description and illustrate specific embodiments in which the described system and method can be practiced. These embodiments have been described in sufficient detail to enable those skilled in the art to practice the currently disclosed systems and methods, and it should be understood that other embodiments can be utilized and structural and logical changes can be made without departing from the spirit and scope of the system. Furthermore, for clarity, detailed descriptions of certain features will not be discussed where they are obvious to those skilled in the art, so as not to obscure the description of the system. Therefore, the following detailed description is not restrictive, and the scope of the system is defined only by the claims.

[0024] Color Doppler (CD) images are generated by estimating the average phase shift within a burst of ultrasound pulses (radio frequency (RF) signals) from the same sample volume. In some examples, each pulse may correspond to one frame (e.g., an RF frame). Longer bursts provide high-quality CD images with sensitive and close estimations of blood flow velocity. However, longer bursts can reduce the frame rate, sometimes significantly, which can hinder diagnosis. The short burst lengths commonly used in clinical scanners extend clutter bandwidth, making wall filtering more challenging and reducing sensitivity. Additionally, obtaining the average Doppler shift from shorter RF bursts can lead to increased variability, thus reducing measurement accuracy.

[0025] Compared to short signal clusters, undersampled signal cluster lengths (e.g., decimated signal clusters) offer improved sensitivity. Undersampled signal clusters also provide greater opportunities for interleaved acquisition (e.g., acquiring alternating ultrasound frames in different modes (e.g., Doppler and B-mode)). However, undersampled signal clusters may suffer from insufficient sampling time of the Doppler signal. More specifically, when the detected Doppler frequency exceeds half the pulse repetition frequency (PRF), blood velocity can no longer be definitively recovered and aliasing may occur.

[0026] Figure 1Example CD images 100, 102, and 104 are shown, representing long (or complete) signal groups, short signal groups, and extracted signal groups, respectively. All CD images 100, 102, and 104 were acquired from a flow phantom. Long signal group CD image 100 was generated from a signal group consisting of 14 ultrasound pulses with a pulse repetition frequency (PRF) of 4000 Hz. Short signal group CD image 102 was generated from a signal group consisting of 7 ultrasound pulses with a PRF of 4000 Hz. As noted above, the short signal group reduces sensitivity compared to the long signal group. For example, in the circled area, CD image 102 failed to detect all the flow detected in CD image 100. Extracted signal group CD image 104 was generated from a signal group consisting of 7 ultrasound pulses with a PRF of 2000 Hz. Although CD image 104 shows higher sensitivity to flow than CD image 102, aliasing of velocity values ​​can be observed, for example, in region 108.

[0027] Therefore, it is desirable to obtain the advantages of long signal groups with high PRF (e.g., high CD image quality as shown in CD image 100) while maintaining the flexibility of short / undersampled signal groups (e.g., higher imaging frame rate, interleaving capability).

[0028] As disclosed herein, artificial intelligence (e.g., deep learning) can be used to concatenate short / undersampled signal groups to CD images generated using longer signal groups and / or signal groups with higher PRF (Proportional Rate of Flow). More specifically, the deep learning framework includes one or more convolutional neural networks (CNNs) where short / undersampled signal groups serve as input data and long and / or high PRF signal group CD images serve as output data. Once trained, the deep learning framework can provide CD images from short / undersampled signal groups that exhibit higher quality (e.g., higher sensitivity, more accurate velocity estimation) compared to typical CD images generated from short / undersampled signal groups. In other words, the CD images provided by the deep learning framework can provide CD images of quality closer to those generated from long and / or high PRF signal groups.

[0029] Figure 2A block diagram of an ultrasound imaging system 200 constructed according to the principles of this disclosure is shown. The ultrasound imaging system 200 according to this disclosure may include a transducer array 214, which may be included in an ultrasound probe 212 (e.g., an external probe or an internal probe (e.g., an intracardiac echo recording (ICE) probe or a transesophageal echo recording (TEE) probe)). In other embodiments, the transducer array 214 may be in the form of a flexible array configured to conformally apply to the surface of the object to be imaged (e.g., a patient). The transducer array 214 is configured to emit ultrasound signals (e.g., beams, waves) and receive echoes in response to the ultrasound signals. Various transducer arrays can be used, such as linear arrays, curved arrays, or phased arrays. For example, the transducer array 214 can include a two-dimensional array of transducer elements (as shown), which is capable of scanning in both elevation and azimuth dimensions for 2D and / or 3D imaging. As is well known, the axial direction is perpendicular to the array plane (in the case of a curved array, the axial direction spreads out in a fan shape), the azimuth direction is usually defined by the longitudinal dimension of the array, and the elevation direction is transverse to the azimuth direction.

[0030] In some embodiments, the transducer array 214 may be coupled to a microwave beamformer 116, which may be located within the ultrasonic probe 212, and the transmission and reception of signals may be controlled by the transducer elements in the array 214. In some embodiments, the microwave beamformer 216 may be controlled by active elements in the array 214 (e.g., an active subset of array elements that defines the active aperture at any given time) to control the transmission and reception of signals.

[0031] In some embodiments, the microwave beamformer 216 may be coupled, for example, via a probe cable or wirelessly, to a transmit / receive (T / R) switch 218, which switches between transmitting and receiving and protects the main beamformer 222 from high-energy transmitted signals. In some embodiments, such as in a portable ultrasound system, the T / R switch 218 and other components of the system may be included in the ultrasound probe 212, rather than in an ultrasound system base that may house image processing electronics. An ultrasound system base typically includes software and hardware components, including circuitry for signal processing and image data generation, and executable instructions (e.g., processing circuitry 250 and user interface 224) for providing a user interface.

[0032] Under the control of microwave beamformer 216, the transmission operation of ultrasonic signals emitted from transducer array 214 is guided by transmission controller 220, which can be coupled to T / R switch 218 and main beamformer 222. Transmission controller 220 can control the direction in which the beam is steered. The beam can travel in a straight line from transducer array 214 (perpendicular to transducer array 214) or be steered at different angles to obtain a wider field of view. Transmission controller 220 can also be coupled to user interface 224 and receive input from user operation of user control keys. User interface 224 may include one or more input devices, such as control panel 252, which may include one or more mechanical controls (e.g., buttons, encoders, etc.), touch-sensitive controls (e.g., trackpads, touchscreens, etc.) and / or other known input devices.

[0033] In some embodiments, the partially beamformed signal generated by microwave beamformer 216 can be coupled to main beamformer 222, in which partially beamformed signals from individual patches of transducer elements can be combined into a fully beamformed signal. In some embodiments, microwave beamformer 216 is omitted, and transducer array 214 is under the control of main beamformer 222, which performs beamforming of all signals. In embodiments with and without microwave beamformer 216, the beamformed signals from main beamformer 222 are coupled to processing circuitry 250, which may include one or more processors (e.g., signal processor 226, mode-B processor 228, Doppler processor 260, and one or more image generation and processing units 268) configured to generate ultrasound images based on the beamformed signals (e.g., beamformed RF data).

[0034] Signal processor 226 can be configured to process received beamformed RF data in various ways, such as bandpass filtering, decimation, I and Q component separation, and harmonic signal separation. Signal processor 226 can also perform additional signal enhancement, such as speckle suppression, signal recombination, and noise cancellation. The processed signals (also referred to as I and Q components or IQ signals) can be coupled to additional downstream signal processing circuitry for image generation. The IQ signals can be coupled to multiple signal paths within the system, each of which can be associated with a specific arrangement of signal processing components suitable for generating different types of image data, such as B-mode image data and Doppler image data. For example, the system may include a B-mode signal path 258 that couples signals from signal processor 226 to B-mode processor 228 for generating B-mode image data.

[0035] The B-mode processor is capable of using amplitude detection to image structures within the body. The signal generated by the B-mode processor 228 can be coupled to a scan converter 230 and / or a multi-plane reformer 232. The scan converter 230 can be configured to arrange the echo signals into a desired image format according to their spatial relationship when received. For example, the scan converter 230 can arrange the echo signals into a two-dimensional (2D) fan-shaped format, or a pyramidal or other shaped three-dimensional (3D) format. The multi-plane reformer 232 is capable of converting echoes received from points in a common plane within a volumetric region of the body into an ultrasound image (e.g., a B-mode image) of that plane, as described, for example, in U.S. Patent US 6,443,896 (Detmer). In some embodiments, the scan converter 230 and the multi-plane reformer 232 can be implemented as one or more processors.

[0036] Volumetric renderer 234 can generate an image of a 3D dataset as viewed from a given reference point (also referred to as a projection, drawing, or rendering result), such as as described in U.S. Patent 6,530,885 (Entrekin et al.). In some embodiments, volumetric renderer 234 can be implemented as one or more processors. Volumetric renderer 234 can generate drawings (e.g., positive or negative drawings) using any known or future known techniques (e.g., surface drawing and maximum intensity drawing).

[0037] In some embodiments, the system may include a Doppler signal path 262 that couples the output from signal processor 226 to Doppler processor 260. Doppler processor 260 may be configured to estimate the Doppler frequency shift and generate Doppler image data. The Doppler image data may include color data, which is then overlaid with B-mode (i.e., grayscale) image data for display. Doppler processor 260 may be configured, for example, to use a wall filter to filter out unwanted signals (i.e., noise or clutter associated with non-moving tissue). Doppler processor 260 may also be configured to estimate velocity and power using known techniques. For example, the Doppler processor may include a Doppler estimator such as an autocorrelation function, where the velocity (Doppler frequency) estimate is based on the independent variable of a lag-1 autocorrelation function, while the Doppler power estimate is based on the magnitude of a lag-0 autocorrelation function. Motion can also be estimated using known phase-domain (e.g., parametric frequency estimators such as MUSIC, ESPRIT, etc.) or time-domain (e.g., cross-correlation) signal processing techniques. Instead of a velocity estimator, or in addition to a velocity estimator, other estimators relating to the temporal or spatial distribution of velocity can be used, such as estimators of acceleration or temporal and / or spatial velocity derivatives. In some embodiments, the velocity and / or power estimates may undergo further thresholding to further reduce noise, and also undergo segmentation and post-processing (e.g., filling and smoothing). The velocity and / or power estimates can then be mapped to a desired display color gamut based on the color map. The color data (also referred to as Doppler image data) can then be coupled to a scan converter 230, in which the Doppler image data can be converted to a desired image format and superimposed on a B-mode image of the tissue structure to form a color Doppler or power Doppler image. In some examples, the power estimates (e.g., lag-0 autocorrelation) can be used to mask or segment the flow (e.g., the velocity estimates) in the color Doppler image before superimposing the color Doppler image onto the B-mode image.

[0038] According to embodiments of this disclosure, an ultrasonic probe 212 can emit a signal group of ultrasonic pulses (e.g., an RF signal group or a simple signal group). In some embodiments, the signal group can be a short and / or undersampled signal group. The ultrasonic probe 212 can receive ultrasonic signals in response to the emitted signal group. These signals can be provided to a Doppler processor 260 via a beamformer 222 and a signal processor 226. The Doppler processor 260 can generate an enhanced CD image based on the received signals. In some embodiments, the enhanced CD image generated by the Doppler processor 260 based on signals from short and / or undersampled signal groups can have higher sensitivity to flow and / or more accurate velocity estimation (e.g., reduced aliasing) compared to a typical CD image generated based on short and / or undersampled signal groups. In some examples, the enhanced CD image generated by the Doppler processor 260 can have sensitivity and accuracy closer to or equal to that of a CD image generated by long and / or high PRF signal groups. In some examples, the signals used to generate the enhanced CD image can be wall-filtered. In other examples, the signals used to generate the enhanced CD image may not be wall-filtered.

[0039] In some embodiments, the Doppler processor 260 may be implemented by one or more processors and / or application-specific integrated circuits (ASICs). In some embodiments, the Doppler processor 260 may include any one or more machine learning, artificial intelligence algorithms and / or multiple neural networks. In some examples, the Doppler processor 260 may include deep neural networks (DNNs), convolutional neural networks (CNNs), recurrent neural networks (RNNs), autoencoder neural networks, etc., to generate enhanced CD images based on short and / or undersampled signal groups. The neural network may be implemented in hardware components (e.g., neurons are represented by physical components) and / or software components (e.g., neurons and paths are implemented in a software application). Neural networks implemented according to this disclosure can be trained using a variety of topologies and learning algorithms to produce desired outputs. For example, a software-based neural network may be implemented using a processor configured to execute instructions (e.g., a single-core or multi-core CPU, a single GPU or GPU cluster, or multiple processors arranged for parallel processing), which may be stored in a computer-readable medium and, when run, cause the processor to execute trained algorithms for generating enhanced CD images based on short and / or undersampled signal groups. In some embodiments, the Doppler processor 260 may combine other image processing methods (e.g., segmentation, histogram analysis) to implement a neural network.

[0040] In various embodiments, any of a variety of currently known or later-developed learning techniques can be used to train one or more neural networks to obtain a neural network (e.g., a trained algorithm or a hardware-based node system) configured to analyze input data in the form of ultrasound images, measurements, and / or statistical data. In some embodiments, the neural network can be trained statically. That is, the neural network can be trained with a dataset and deployed on the Doppler processor 260. In some embodiments, the neural network can be trained dynamically. In these embodiments, the neural network can be trained with an initial dataset and deployed on the Doppler processor 260. However, after the neural network is deployed on the Doppler processor 260, the neural network can be further trained and modified based on ultrasound images acquired by the system 200.

[0041] Output from the scan converter 230, the multiplane reformer 232, and / or the volume plotter 234 can be coupled to the image processor 236 for further enhancement, buffering, and temporary storage before display on the image display 238. The graphics processor 240 can generate graphic overlays to be displayed along with the images. These overlays can include, for example, standard identification information (e.g., patient name), the date and time of the image, imaging parameters, etc. For these purposes, the graphics processor can be configured to receive input from the user interface 224, such as a typed patient name or other annotation. The user interface 224 can also be coupled to the multiplane reformer 232 for selecting and controlling the display of multiple multiplane reformatted (MPR) images.

[0042] System 200 may include local memory 242. Local memory 242 may be implemented as any suitable non-transient computer-readable medium (e.g., flash drive, disk drive). Local memory 242 may store data generated by system 200, including ultrasound images, executable instructions, imaging parameters, training datasets, or any other information required for the operation of system 200.

[0043] As mentioned above, system 200 includes a user interface 224. User interface 224 may include a display 238 and a control panel 252. Display 238 may include a display device implemented using various known display technologies (e.g., LCD, LED, OLED, or plasma display technologies). In some embodiments, display 238 may include multiple displays. Control panel 252 may be configured to receive user input (e.g., signal group length, imaging mode). Control panel 252 may include one or more hardware controls (e.g., buttons, knobs, dial pads, encoders, mice, trackballs, etc.). In some embodiments, control panel 252 may additionally or alternatively include software controls (e.g., GUI control elements or simply GUI controls) provided on a touch-sensitive display. In some embodiments, display 238 may be a touch-sensitive display that includes one or more software controls of control panel 252.

[0044] In some embodiments, Figure 2 The various components shown can be combined. For example, image processor 236 and graphics processor 240 can be implemented as a single processor. In some embodiments, Figure 2 The various components shown can be implemented as separate components. For example, signal processor 226 can be implemented as a separate signal processor for each imaging mode (e.g., B-mode, Doppler). In some embodiments, Figure 2 One or more of the processors shown may be implemented by a general-purpose processor and / or a microprocessor configured to perform a specific task. In some embodiments, one or more of the processors may be implemented as dedicated circuitry. In some embodiments, one or more of the processors (e.g., image processor 236) may be implemented using one or more graphics processing units (GPUs).

[0045] Figure 3 This is a block diagram illustrating an example processor 300 according to the principles of this disclosure. Processor 300 can be used to implement one or more processors and / or controllers described herein, for example, Figure 2 The image processor 236 and / or shown Figure 2 Any other processor or controller shown. Processor 300 can be any suitable processor type, including but not limited to microprocessors, microcontrollers, digital signal processors (DSPs), field-programmable arrays (FPGAs) (wherein the FPGA has been programmed to form a processor), graphics processing units (GPUs), application-specific integrated circuits (ASICs) (wherein the ASIC has been designed to form a processor), or combinations thereof.

[0046] Processor 300 may include one or more cores 302. Core 302 may include one or more arithmetic logic units (ALUs) 304. In some embodiments, in addition to or in place of ALU 304, core 302 may include a floating-point logic unit (FPLU) 306 and / or a digital signal processing unit (DSPU) 308.

[0047] Processor 300 may include one or more registers 312 communicatively coupled to core 302. Registers 312 may be implemented using dedicated logic gates (e.g., flip-flops) and / or any memory technology. In some embodiments, registers 312 may be implemented using static memory. Registers may provide data, instructions, and addresses to core 302.

[0048] In some embodiments, processor 300 may include one or more levels of cache memory 310 communicatively coupled to core 302. Cache memory 310 may provide computer-readable instructions to core 302 for execution. Cache memory 310 may provide data processed by core 302. In some embodiments, the computer-readable instructions may have already been provided to cache memory 310 by local memory (e.g., local memory attached to external bus 316). Cache memory 310 may be implemented using any suitable cache memory type, such as metal-oxide-semiconductor (MOS) memory (e.g., static random access memory (SRAM), dynamic random access memory (DRAM)) and / or any other suitable memory technology.

[0049] Processor 300 may include controller 314, which can control other processors and / or components included in the system (e.g., Figure 2 The control panel 252 and scan converter 230 shown are used for input to the processor 300 and / or from the processor 300 to other processors and / or components included in the system (e.g., Figure 2 The outputs of the display 238 and volume plotter 234 are shown. The controller 314 can control the data paths in the ALU 304, FPLU 306, and / or DSPU 308. The controller 314 can be implemented as one or more state machines, data paths, and / or dedicated control logic units. The gates of the controller 314 can be implemented as stand-alone gates, FPGAs, ASICs, or any other suitable technology.

[0050] Register 312 and cache memory 310 can communicate with controller 314 and core 302 via internal connections 320A, 320B, 320C and 320D. The internal connections can be implemented as a bus, multiplexer, crossbar switch and / or any other suitable connection technology.

[0051] Input and output to processor 300 can be provided via bus 316, which may include one or more wires. Bus 316 may be communicatively coupled to one or more components of processor 300, such as controller 314, cache memory 310, and / or register 312. Bus 316 may be coupled to one or more components of the system, such as the previously mentioned display 238 and control panel 252.

[0052] Bus 316 may be coupled to one or more external memories. The external memory may include read-only memory (ROM) 332. ROM 332 may be a masked ROM, electronically programmable read-only memory (EPROM), or any other suitable technology. The external memory may include random access memory (RAM) 333. RAM 333 may be static RAM, battery-backed static RAM, dynamic RAM (DRAM), or any other suitable technology. The external memory may include electrically erasable programmable read-only memory (EEPROM) 335. The external memory may include flash memory 334. The external memory may include a magnetic storage device, such as a disk 336. In some embodiments, external memory (e.g., local memory 242) may be included in the system (e.g., Figure 2 The ultrasound imaging system 200 shown is used in this system.

[0053] Figure 4 A block diagram illustrating the process for training and deploying a neural network based on the principles of this disclosure is shown. Figure 4 The process shown can be used to train a neural network included in the Doppler processor 260. Figure 4The left-hand side (Stage 1) illustrates the training of the neural network. To train the neural network, a training set comprising multiple instances of the input array and output classification can be presented to one or more training algorithms of the neural network (e.g., the AlexNet training algorithm, as described in "ImageNet Classification with Deep Convolutional Neural Networks" by Krizhevsky, A., Sutskever, I., and Hinton, GE, NIPS 2012 or its successors). Training may involve selecting a starting network architecture 412 and preparing training data 414. The starting network architecture 412 can be a blank architecture (e.g., an architecture with a defined layer and node arrangement but without any previously trained weights) or a partially trained network (e.g., a pre-trained network), which can then be further tailored for the classification of ultrasound images. The starting architecture 412 (e.g., blank weights) and training data 414 are provided to a training engine 410 (e.g., an ADAM optimizer) for training the model. After a sufficient number of iterations (e.g., when the model performs consistently within an acceptable error range), model 420 is considered trained and ready for deployment. Figure 4 The intermediate (stage 2) diagram illustrates this situation. Figure 4 On the right-hand side (or in stage 3), (via inference engine 430) a trained model 420 is applied to analyze new data 432, which is data not presented to the model during the initial training (in stage 1). For example, the new data 432 may include unknown images, such as live ultrasound images acquired during patient scanning (e.g., cardiac images during echocardiography). The trained model 420, implemented via engine 430, is used to classify the unknown images based on the training of model 420 to provide output 434 (e.g., generating enhanced CD images based on short and / or undersampled signal clusters). The system can then use the output 434 for subsequent processes 440 (e.g., displaying the enhanced CD images, performing additional image processing on the enhanced CD images).

[0054] In embodiments that implement the neural network of the Doppler processor 260 using a trained model 420, the initial architecture may be a convolutional neural network or a deep convolutional neural network initial architecture, which may be trained to generate enhanced CD images based on short and / or undersampled signal groups. Training data 414 may include multiple (hundreds, often thousands or even more) annotated / labeled images (also referred to as training images). It should be understood that training images do not need to include complete images generated by the imaging system (e.g., representing the complete field of view of an ultrasound probe or the entire MRI volume), but may include mosaics or portions of images, such as those including flowing portions.

[0055] In various embodiments, the trained neural network may be implemented at least in part in a computer-readable medium comprising executable instructions run by a processor (e.g., Doppler processor 260).

[0056] Figure 5 An overview of the training and deployment phases of a neural network according to the principles of this disclosure is shown. During the training phase, an untrained neural network 500 receives a training set as input 504, comprising a portion and / or the entire image signal including signals from short and / or undersampled RF signal clusters. In some examples, wall filtering may be applied to the signal. In other examples, wall filtering may not be applied to the signal. The portion of the signal may include a portion including flow (e.g., blood flow in a blood vessel). In some examples, all signals from short and / or undersampled RF signal clusters may be provided as input (e.g., blood flow in a blood vessel and signals from surrounding tissue). However, this would train the neural network 500 on noise, which would result in artifacts being generated in the output (e.g., false indications of flow in non-moving tissue). The training set may also include a desired output set 506 associated with each input in input 504. Output 506 may include a CD image generated based on signals from long and / or high PRF signal clusters. In some examples, output 506 may include components of the CD image, such as the phase of lag-1 autocorrelation or log-compressed lag-0 autocorrelation. See reference Figure 4 The untrained neural network 500 discussed here can adjust itself (e.g., adjust weights, number of layers, number of nodes in each layer, etc.) until the neural network 500 can reliably generate an output that is exactly the same or sufficiently similar to the expected output 506 based on the corresponding input 504.

[0057] Following the training phase, a trained neural network 502 (e.g., in an ultrasound imaging system, such as ultrasound imaging system 200) can be deployed. The trained neural network 502 can receive previously unseen (e.g., new) signal portions from short / undersampled RF signal groups as input 508. In some examples, the entire signal can be fed to the trained neural network 502. Wall filtering may or may not be applied to the signal. The signal may have already been acquired by an ultrasound probe (e.g., ultrasound probe 212). The trained neural network 502 can perform operations (e.g., processing) on ​​the input 508 to generate an enhanced CD image as output 510. As previously discussed, the enhanced CD image can more closely resemble an image generated from long and / or high PRF signal groups compared to a CD image generated from short / undersampled RF signal groups without using the trained neural network 502.

[0058] Figure 6 The illustration depicts an example process or workflow for generating a training dataset based on the principles of this disclosure. In a first step 600, signals from long and / or high PRF (Pulse Resonance Factor) RF signal clusters may be acquired. These signals may be signals from one or more blood vessels, the heart or portions thereof, or other in vivo regions with flow acquired by an ultrasound probe. In some examples, the ultrasound probe may acquire signals from a flow phantom (e.g., the Gammex flow phantom from Sun Nuclear Corporation). In some examples, signals may be acquired using ultrasound simulation software (e.g., software developed by the Technical University of Denmark). The Field II algorithm developed by Arendt Jensen is used to simulate and acquire the signal. The signal acquired at step 600 may correspond to multiple frames as shown at step 602. For example, in some applications, long signal groups may generate 10-20 frames. At step 602, a subset of frames from long and / or high PRF signal groups can be used to provide a set of frames corresponding to short and / or undersampled signal groups at step 604.

[0059] At step 606, a CD image can be generated using frames from long and / or high PRF signal groups. The CD image generated at step 606 can be a high-quality CD image (e.g., high sensitivity and accurate velocity estimation). At step 608, a CD image can also be generated using frames from short and / or undersampled signal groups. However, step 608 is not required to generate a training dataset, and step 608 is provided only to illustrate the quality difference between the CD image generated at step 606 based on long and / or high PRF signal groups and the CD image generated based on short signal groups. For example, aliasing of velocity values ​​can be observed in the image at step 608 in region 610.

[0060] Finally, at step 612, the short and / or undersampled signal clusters generated at step 604 are concatenated to the corresponding CD image generated at step 606 based on the long and / or high PRF signal clusters. These input-output pairs are then fed to a neural network (e.g., neural network 500) for training.

[0061] This document provides two examples for generating training data. However, these examples are provided for illustrative purposes only, and this disclosure is not limited to the examples for generating training data disclosed herein.

[0062] In a first example according to this disclosure, simulated data is used to train the CNN. A flow phantom is generated by simulating scatterers within the boundaries of assumed blood vessels (e.g., 18 simulated blood vessels with radii of 1.5–7.5 mm). The motion of the scatterers is simulated based on parabolic flow profiles. The angle and radius of the assumed blood vessels, as well as the peak flow velocity along the assumed blood vessels, vary randomly for each phantom.

[0063] Doppler imaging of the simulated fluid phantom was performed using Field II simulation. An S5-1 transducer (center frequency 2.5 MHz, 50% bandwidth) was simulated, and a signal group of 2N RF frames was generated for each phantom in the phantom at an assumed PRF (e.g., 2N = 10, PRF = 2000 Hz). The RF signal for each channel was then sampled, beamformed, and QBP filtered.

[0064] Therefore, the resulting 3D dataset has the following dimensions: samples × receiver lines × signal group size. The RF signal group (CNN input, e.g., N = 5, PRF = 1000 Hz) is then factored down by 2. A 1D autocorrelation method is used to estimate the Doppler phase shift for the original RF signal group (CNN target). Pieces are extracted from each pair of training input / target (e.g., 800 pieces, size 100 samples × 20 receiver lines (× 5 RF frames, for the RF signal group)). Since only flow is simulated (e.g., no tissue scattering is simulated), wall filtering is not applied to the simulated dataset, essentially simulating an ideal wall-filtered dataset.

[0065] In a second example according to this disclosure, a Gammex fluid phantom (available from Philips Healthcare) is scanned using an EPIQ ultrasound scanner to generate training data for a CNN. The velocity of the fluid phantom is different each time it is acquired. Using an S5-1 transducer (center frequency 2.5 MHz, 50% bandwidth), a signal group of 14 pulses at a specific PRF is obtained and QBP filtered (long signal group, Nle = 14, PRF = 4000 Hz). In this example, only the first four pulses of the long signal group are extracted for use as CNN input for the short signal group (Nse = 4 pulses, PRF = 4000 Hz). For each case, wall filtering is performed on both the long and short signal groups using appropriate wall filters (e.g., high-pass filters). Wall filtering results in the loss of a small number (N1) of the RF signal group frames (e.g., Nle' = Nle - N1 = 12, Nse' = Nse - N1 = 2).

[0066] A 1D autocorrelation method is employed to estimate the average Doppler phase shift of long RF signal clusters (e.g., the desired output of a CNN). Pieces are extracted from each pair of training input / desired output (e.g., 500 pieces, sized at 200 samples × 20 receiver lines (× 2 RF frames, for the RF signal cluster)). Doppler-specific enhancements are used to facilitate training, involving selecting training pieces only around points where the power of the Doppler signal (lag-0 autocorrelation of the RF signal cluster) exceeds a user-defined depth correlation threshold. This reduces the noise level during CNN training.

[0067] Each RF signal group patch in the RF signal group patch is centered at zero and normalized to unit power. The normalized RF signal group patches and their corresponding CD phase image patches are shuffled to facilitate generalization and are split into training and validation datasets.

[0068] Figure 7 This is a diagram illustrating an example of a neural network based on the principles of this disclosure. The neural network 700 is a convolutional neural network with two sequential uNets. To the left of line 701, a schematic diagram 702 of the neural network 700 is shown, illustrating the first uNet 704 and the second uNet 706. To the right of line 701 is a block diagram 708 of the corresponding architecture of the neural network 700. Horizontal dashed lines 703 and 705 indicate which parts of block diagram 708 correspond to which uNets 704 and 706 in schematic diagram 702.

[0069] exist Figure 7In the example shown, the first uNet 704 focuses on extracting information from each individual frame of a short and / or undersampled RF signal group while maintaining the same signal group size. A convolutional layer then compresses the processed short and / or undersampled RF signal group into a single average Doppler phase image. The result of this layer is cascaded with the average Doppler phase of the short and / or undersampled signal group (auxiliary input #1), which facilitates average Doppler phase estimation (e.g., using features of the short R1 phase). This result is then fed into the second uNet 706, which focuses on processing the estimated average Doppler phase to match it with the estimated Doppler phase of a long and / or high PRF signal group, thereby generating an enhanced CD image. In other words, the first uNet 704 receives a short and / or undersampled RF signal group and outputs an average Doppler phase image and the Doppler phase of the short and / or undersampled signal group. The second uNet 706 receives the output of the first uNet as input and outputs the enhanced CD image. In some examples, the second uNet 706 can output a generated RF signal group, based on which an enhanced CD image can be generated.

[0070] Optionally, in some examples (e.g., such as...) Figure 7 In the example shown, Doppler-specific enhancements to the neural network 700 can be included during the training phase. These Doppler-specific enhancements can include a custom masking mean squared error (MSE) loss. More specifically, after high-pass wall filtering, non-flowing regions exhibit high-magnitude, noisy Doppler shift values ​​that may dominate the MSE used during training, thus biasing the MSE and causing artifacts to appear in the predicted image. To minimize this problem, a mask based on long signal group power Doppler can be implemented. The long signal group power Doppler image can be thresholded according to a depth-dependent threshold and then convolved with a Gaussian kernel to obtain a smooth mask of Doppler flow. This mask can be concatenated with the CNN output during training and used to estimate a custom weighted MSE loss function, as indicated by arrow 710 and the box labeled “Auxiliary Input 2”. Using the masking loss function during training allows the neural network 700 to focus on flowing regions and prevent or reduce bias due to high background values. During testing, since no error metric was calculated, the last connection layer could be removed.

[0071] Optionally, in some examples, adversarial loss can be used during training. Adversarial loss can make the resulting enhanced CD images look more natural. Adversarial loss can be added by using a Conditional Generative Adversarial Network (cGAN) or a similar model (collectively referred to as a discriminator network). In these examples, the discriminator network can be trained simultaneously with a neural network 700 to distinguish the enhanced CD images from CD images generated based on long and / or high PRF signal clusters, based on a binary cross-entropy loss provided by the following formula:

[0072] Binary_cross_entropy--∑(label i ·log(p(label i ))+(1-label,)·log(1-p(label i )))

[0073] Formula 1

[0074] Where p is a probability function, and label i It is the label of the i-th input frame entering the adversarial network. For example, if the input frame is an augmented CD image (e.g., "fake") generated by a neural network 700, then the label is... i The label can be 1, but if the input box is a CD image generated based on a long and / or high PRF signal group (e.g., "real"), then the label... i It can be equal to 0. At the output of the discriminator network, the probability that the i-th input frame is "fake" is calculated. Therefore, p(label) i The label (p) can have values ​​between 0 and 1, including both 0 and 1. i When it corresponds to the correct label (e.g., when label...) i =1->p(label) i ) = 1 and label i =0->p(label) i When p(label) = 0, the binary cross-entropy loss is minimized. i When incorrect (e.g., when the label) i =0->p(label) i ) = 1 and label i =1->p(label) i When ) = 0, the binary cross-entropy loss is maximized.

[0075] The discriminator network is trained to distinguish the output of neural network 700 (e.g., distinguishing an enhanced CD image from a CD image from a long and / or high PRF signal cluster) to minimize the loss in Equation 1. Once the discriminator network has been trained to minimize this loss, the discriminator network is frozen, and neural network 700 continues to be trained to minimize the total loss. In some examples, the total adversarial loss used to train neural network 700 can be determined by averaging the weighted MSE and the binary cross-entropy loss.

[0076] Modifying neural network 700 using MSE loss and / or adversarial loss can impart a higher value to the background noise, which is closer to the value expected by conventional Doppler processing. This can facilitate image processing because it may not be necessary to modify the signal and image processing (SIP) and / or background noise filtering of the ultrasound system (e.g., signal processor 226, image processor 236).

[0077] Figure 7 The neural network 700 shown is provided as an example only. Other neural networks can be used, such as a regular serial CNN.

[0078] Figure 8 This is a flowchart illustrating a method 800 using optional masking loss based on the principles of this disclosure. (See reference...) Figure 7 The masking mean square loss discussed can help ensure that a neural network (e.g., a neural network 700) learns to reconstruct meaningful flow regions while avoiding bias from high-magnitude Doppler phase noise (e.g., from surrounding tissue without flow). Equation 2 provides the MSE loss:

[0079]

[0080] Where E is the expected value / average value, and R1 is the lag-1 autocorrelation function of the RF signal group (e.g., Doppler phase). It is a Doppler phase image predicted by a neural network, and It is a high-quality Doppler phase image generated from a long and / or high PRF RF signal group. The mask is a mask generated using the R0 (lag-0) power Doppler image from a long and / or high PRF RF signal group.

[0081] Figure 8 The illustration shows the steps for generating the mask in Equation 2 when the neural network is trained on back-wall filtered data. The long and / or high PRF signal groups 802 of N frames of back-wall filtered data are acquired, and the Doppler power (R0) image 804 is estimated from the long and / or high PRF signal groups 802 as provided in the following formula:

[0082]

[0083] Among them, RF wallfiltered (x,z,i) is the signal group of the longwall filter, where x is the azimuth coordinate, z is the depth, and i is the index in the RF signal group.

[0084] The linear depth dependence threshold of 806 for the mask is also estimated using Equation 4 given below:

[0085]

[0086] Where a and K are user-defined parameters defining the threshold of attack, and E[R0] is the average value of the long and / or high PRF signal cluster R0. Finally, the estimated (e.g., the original) mask 808 and the smoothed mask 810 are generated by the following formula:

[0087]

[0088] Mask(x,z) = G(x,z) * Mask raw Formula 6 (x,z)

[0089] Here, G(x,z) indicates the Gaussian kernel, and the asterisk (*) indicates convolution. The methods shown in Equations 5 and 6 are merely examples. Other more advanced / complex thresholding strategies can be followed to obtain masks and other filtering kernels, or other methods can be used.

[0090] During training, a mask such as mask 808 or mask 810, or a mask generated using another appropriate process, is provided to a neural network such as neural network 700. Figure 7 In the example shown, as indicated by arrow 710, a mask 808 or mask 810 can be provided to the neural network 700. (See reference...) Figure 7 The subject of discussion, such as Figure 8 The mask shown is used only during training to perform loss estimation when long and / or high PRF signal clusters are available. During deployment, the neural network has already been trained. Therefore, the mask is no longer needed because no loss estimation or training is performed (except in embodiments employing dynamic training). In some examples, the same mask can be used to detect flow regions and generate image mosaics of the flow that include the training dataset used to train the neural network. This is similar to obtaining pixel-centered image mosaics where mask(x, z) > 0.

[0091] When still using reference Figure 7 When adversarial loss occurs from the discriminator network as discussed in Equation 1, the total loss is given by the following equation:

[0092]

[0093] The R1 loss is provided by Equation 2. Therefore, the neural network can be trained to minimize the masking loss and maximize the discriminator loss. In other words, the neural network can be trained to ignore noise and make the discriminator network unable to distinguish between a CD image with a long RF signal cluster and an enhanced CD image generated by the neural network.

[0094] While examples of this disclosure have been described with reference to CD images, at least some of the principles of this disclosure can be applied to power Doppler (PD) images (e.g., R0, lag-0 autocorrelation). For example, a neural network such as Neural Network 700, or an additional branch of a neural network, can be used to generate enhanced PD images based on short and / or undersampled RF signal clusters to mimic PD images generated based on long and / or high PRF signal clusters. In these examples, when training the neural network to generate enhanced PD images, standard losses (e.g., mean squared error losses) and / or masking losses (e.g., losses provided in Equation 2) can be used. In some examples, logarithmic losses can be applied, which can facilitate optimization of the neural network by compressing the arithmetic values ​​involved. The logarithmic loss can be given by:

[0095]

[0096] Where E indicates the expected value / average value, and R0 is the lag-0 autocorrelation function of the RF signal group (e.g., power Doppler information). It is a Doppler phase image predicted by a neural network, and It is a high-quality Doppler phase image generated from long and / or high PRF signal groups.

[0097] Figure 9 This is a flowchart of method 800 based on the principles of this disclosure. In some examples, method 800 may be performed by an ultrasound imaging system (e.g., Figure 2 The system 200 shown is used to execute this method. In some examples, method 900 may be executed by a Doppler processor (e.g., Doppler processor 260).

[0098] At box 902, the action of "receiving an ultrasound signal corresponding to a group of RF signals" can be performed. This reception can be performed by a processor (e.g., Doppler processor 260). The group of RF signals can have a first length (e.g., multiple pulses, RF frames), a first pulse repetition frequency (PRF), and / or a first sensitivity. The group of RF signals may have already been transmitted by an ultrasound probe (e.g., ultrasound probe 212). The ultrasound signal generated by the group of RF signals may have already been received by the ultrasound probe, and this signal is provided to the processor.

[0099] At box 904, "processing ultrasound signals to generate a color Doppler image" can be performed. This processing can be performed by a processor. In some examples, the processing may include estimating a second RF signal group based at least in part on a reference ultrasound signal corresponding to at least one reference RF signal group. The reference RF signal group may have at least one of a second pulse repetition rate, a second length, or a second sensitivity that differs from the first RF signal group. The processing may also include using the second RF signal group to generate a color Doppler image. In some examples, the enhanced color Doppler image may have at least one of the sensitivity or velocity estimation results (e.g., blood flow velocity estimation results) that is closer to the color Doppler image generated from an RF signal group having similar characteristics to the reference RF signal group than the color Doppler image generated from the unprocessed RF signal group. In some examples, the length of the reference RF signal group may be greater than the length of the RF signal group. That is, the second length may be greater than the first length. For example, the second length may be the length corresponding to a long signal group (e.g., ~14 frames), while the first length corresponds to a short and / or undersampled signal group (e.g., ~4 frames). In some examples, this processing can be performed by a neural network. In some examples, the neural network can be a series of convolutional networks (e.g., Figure 7 The two uNets shown).

[0100] Optionally, at box 906, in an example where processing is performed by a neural network, "training the neural network" may be performed. In some examples, training may include: providing multiple signals corresponding to a group of RF signals of a first length as input, and providing multiple corresponding color Doppler images generated from the corresponding multiple signals corresponding to a group of RF signals of a second length as the desired output. In some examples, the input may be generated from a subset of the group of RF signals of the second length. The training set may be obtained from simulations, fluid phantoms, and / or bodies. In some examples, masking and / or adversarial losses may be applied during training. In some examples, the neural network is trained before it is implemented (e.g., deployed) on a processor.

[0101] Figure 10 An example extracted signal group color Doppler image 1000 and an example enhanced color Doppler image 1002 are shown, based on the principles of this disclosure. Image 1000 was acquired from a Gammex fluid phantom using an EPIQ system (Philips Healthcare). In this example, the neural network used to generate the enhanced CD image 1002 was trained only on a simulated dataset (e.g., from field II). Although the neural network was trained only on simulated data, the enhanced CD image 1002 of the fluid phantom also exhibited partial compensation for the aliasing of velocity estimation results present in the extracted signal group CD image 1000.

[0102] Figure 11 Example short signal groups, long signal groups, and enhanced original phase-shifted color Doppler images are shown according to the principles of this disclosure. Column 1100 includes CD images of extracted signal groups from four frames. Column 1102 includes CD images of long signal groups from 14 frames. Column 1104 includes enhanced CD images generated by a series of conventional CNNs (without masking applied during training). Column 1106 includes enhanced CD images generated by a series of conventional CNNs (with masking applied during training). Column 1108 includes enhanced CD images generated by... Figure 7 The enhanced CD image generated by the neural network 700 shown (masking was applied during training). For this example, all neural networks were trained solely on phantom data. The image in row 1110 was acquired from the first Gammex flow phantom. The image in row 1112 was acquired from the second Gammex flow phantom. The image in row 1114 is an in vivo image of the heart. Moving right from column 1104 to column 1108, for each imaged medium (e.g., the first Gammex flow phantom, the second Gammex flow phantom, and the in vivo object) in each of the three rows 1110, 1112, and 1114 respectively, a gradual improvement in the enhanced CD image (e.g., noise suppression, resolution) can be observed.

[0103] Figure 12 Example short signal groups, long signal groups, and enhanced original phase-shift results color Doppler images are shown according to the principles of this disclosure. Column 1200 includes CD images of decimated signal groups from four frames. Column 1202 includes CD images of long signal groups from 14 frames. Column 1204 includes... Figure 7 The enhanced CD image generated by the neural network 700 shown (masking was applied during training). Column 1206 includes images generated by... Figure 7 The enhanced CD image generated by the neural network 700 shown (with masking and adversarial loss applied during training). For this example, all neural networks were trained solely on phantom data. The image in row 1210 was acquired from the first Gammex flow phantom. The image in row 1212 was acquired from the second Gammex flow phantom. The image in row 1214 is an in vivo image of the heart. Moving right from column 1204 to column 1206, improvements in the enhanced CD image (e.g., noise suppression, resolution) can be seen.

[0104] Figure 13Example short signal groups, long signal groups, and enhanced post-processed color Doppler images are shown according to the principles of this disclosure. Column 1300 includes CD images of signal groups from four frames for four different Gammex flow models. Column 1302 includes CD images of long signal groups from 14 frames for four different Gammex flow models. Column 1304 includes images of long signal groups from fourteen frames for four different Gammex flow models. Figure 7 The enhanced CD image generated by the neural network 700 shown is illustrated (masking was applied during training on four different Gammex flow phantoms, but without adversarial loss). For this example, the neural network 700 was trained solely on the phantom data. The image shown in column 1304 is more similar to the image in column 1302 than the image shown in column 1300, indicating that the enhanced CD image provided by the neural network 700 based on short and / or undersampled signal groups has improved sensitivity, improved velocity estimation, and / or improved resolution compared to unprocessed short and / or undersampled signal groups.

[0105] Figure 14 Example short signal clusters, long signal clusters, and enhanced post-processed color Doppler images are shown according to the principles of this disclosure. Column 1400 includes CD images of signal clusters from four frames of data from three different cardiac test cases acquired in vivo. Column 1402 includes CD images of long signal clusters from 14 frames of data from three different cardiac test cases acquired in vivo. Column 1204 includes images of long signal clusters from three different cardiac test cases acquired in vivo. Figure 7 The enhanced CD images generated by the neural network 700 shown are illustrated (masking was applied during training on three different cardiac test cases acquired in vivo, but without adversarial loss). For this example, the neural network 700 was trained solely on phantom data. Compared to the images shown in column 1400, the images shown in column 1404 exhibit reduced flow velocity variance and greater noise suppression, indicating that the enhanced CD images provided by the neural network 700 based on short and / or undersampled signal clusters have improved sensitivity, improved velocity estimation, and / or improved resolution compared to unprocessed short and / or undersampled signal clusters. Further improvements can be seen by training the neural network 700 on the in vivo training dataset instead of the phantom dataset, or by increasing the flow velocity in the phantom dataset to match the velocity of cardiac blood flow.

[0106] Figure 9-13 All example results shown use the power Doppler (R0) of the full signal group to ensure a fair comparison of the estimated flow velocities. This means that the mask used to determine whether to show a color Doppler (R1) image or a B-mode image at a particular pixel is derived from the full signal group.

[0107] As disclosed herein, artificial intelligence (e.g., deep learning) can be used to link short and / or undersampled signal groups to CD images generated using longer signal groups. More specifically, a deep learning framework comprising one or more convolutional neural networks (CNNs) (e.g., two serial uNets) can be used to process short and / or undersampled signal groups to generate CD images that mimic CD images generated from longer signal groups. During training, the deep learning framework can receive short and / or undersampled signal groups as input data and CD images generated from long and / or high PRF signal groups as desired output data. Once trained, the deep learning framework can provide enhanced CD images from short and / or undersampled signal groups of higher quality (e.g., higher sensitivity, more accurate velocity estimation results) compared to typical CD images generated from short and / or undersampled signal groups. In other words, the enhanced CD images provided by the deep learning framework can provide CD images of quality closer to those generated from long and / or high PRF signal groups. In some applications, the systems and methods disclosed herein can provide improved CD images with less loss in frame rate and / or interleaving capability.

[0108] While the examples described herein discuss the processing of ultrasound image data, it should be understood that the principles of this disclosure are not limited to ultrasound and can be applied to image data from other modalities such as magnetic resonance imaging and computed tomography.

[0109] In various embodiments that implement components, systems, and / or methods using programmable devices (e.g., computer-based systems or programmable logic units), it should be understood that the aforementioned systems and methods can be implemented using various known or later-developed programming languages ​​(e.g., "C", "C++", "C#", "Java", "Python", etc.). Therefore, various storage media (e.g., computer disks, optical disks, electronic storage devices, etc., capable of containing information that can instruct devices such as computers) can be prepared to implement the aforementioned systems and / or methods. Once a suitable device accesses the information and programs contained on the storage medium, the storage medium can provide the information and programs to the device, thereby enabling the device to perform the functions of the systems and / or methods described herein. For example, if a computer disk containing appropriate materials (e.g., source files, object files, executable files, etc.) is provided to a computer, the computer can receive this information, appropriately configure itself, and perform the functions of the various systems and methods outlined in the diagrams and flowcharts above, thereby implementing various functions. That is, the computer can receive portions of various information from the disk relating to different elements of the aforementioned systems and / or methods, implement the respective systems and / or methods, and coordinate the functions of the aforementioned systems and / or methods.

[0110] In light of this disclosure, it should be noted that the various methods and apparatuses described herein can be implemented in hardware, software, and firmware. Furthermore, the various methods and parameters are included by way of example only and are not intended to be limiting. In view of this disclosure, those skilled in the art can implement the teachings to influence these techniques while determining their own techniques and desired instrumentation, while remaining within the scope of this invention. The functionality of one or more processors in the processors described herein can be incorporated into a smaller number or a single processing unit (e.g., a CPU) and can be implemented using application-specific integrated circuits (ASICs) or general-purpose processing circuitry programmed to perform the functions described herein in response to executable instructions.

[0111] While this system has been specifically described with reference to ultrasound imaging systems, it is conceivable that it can be extended to other medical imaging systems to systematically acquire one or more images. Therefore, this system can be used to acquire and / or record image information relating to, but not limited to, the kidneys, testes, breasts, ovaries, uterus, thyroid gland, liver, lungs, musculoskeletal system, spleen, heart, arteries and vascular system, as well as other imaging applications related to ultrasound-guided interventions. Additionally, this system may include one or more procedures that can be used with conventional imaging systems, such that said one or more procedures can provide the features and advantages of this system. Upon studying this disclosure, those skilled in the art will readily perceive certain additional advantages and features of this disclosure, or will experience certain additional advantages and features of this disclosure upon employing the novel systems and methods of this disclosure. Another advantage of this system and method is the ability to easily upgrade conventional medical imaging systems to incorporate the features and advantages of this system, device, and method.

[0112] Of course, it should be understood that any of the examples, embodiments, or processes described herein may be combined with one or more other examples, embodiments, and / or processes, or may be separated in an apparatus or apparatus portion of the system, apparatus, and method, and / or performed in an apparatus or apparatus portion of the system, apparatus, and method.

[0113] Finally, the foregoing discussion is intended to illustrate the system only and should not be construed as limiting the claims to any particular embodiment or group of embodiments. Therefore, while the system has been described in particular and in detail with reference to exemplary embodiments, it should be understood that those skilled in the art can devise many modifications and alternative embodiments without departing from the broader and contemplated spirit and scope of the invention as set forth in the claims. Thus, the specification and drawings should be considered illustrative and not intended to limit the scope of the claims.

Claims

1. An ultrasound imaging system comprising: a processor configured to: receive ultrasound signals corresponding to a first radio frequency signal group, the first radio frequency signal group comprising a first length, a first pulse repetition frequency, and a first sensitivity; and input the first radio frequency signal group into one or more artificial intelligence algorithms configured to output a second radio frequency signal group, a color Doppler image, or a power Doppler image, wherein the one or more artificial intelligence algorithms are trained using: reference ultrasound signals corresponding to at least one reference radio frequency signal group comprising at least one of a second higher pulse repetition rate, a second longer length, or a second higher sensitivity than the first radio frequency signal group.

2. The ultrasound imaging system of claim 1, wherein, the one or more artificial intelligence algorithms comprise one or more machine learning algorithms.

3. The ultrasound imaging system of claim 2, wherein, the one or more machine learning algorithms comprise one or more neural networks.

4. The ultrasound imaging system of claim 1, wherein, the processor is further configured to wall filter the ultrasound signals corresponding to the first radio frequency signal group.

5. The ultrasound imaging system of claim 1, wherein, the processor is configured to input the first radio frequency signal group into a neural network, wherein the neural network comprises a series of convolutional neural networks.

6. The ultrasound imaging system of claim 5, wherein, the series of convolutional neural networks is a uNet.

7. The ultrasound imaging system of claim 5, wherein, a first neural network of the series of convolutional neural networks receives the ultrasound signals corresponding to the RF signal group and provides an average Doppler phase image and a Doppler phase as a first output, wherein a second neural network of the series of convolutional neural networks receives the first output and provides the color Doppler image as a second output.

8. An ultrasound imaging method comprising: receiving ultrasound signals corresponding to a first radio frequency signal group, the first radio frequency signal group comprising a first length, a first pulse repetition frequency, and a first sensitivity; and inputting the first radio frequency signal group into one or more artificial intelligence algorithms configured to output a second radio frequency signal group, a color Doppler image, or a power Doppler image, wherein the one or more artificial intelligence algorithms are trained using: reference ultrasound signals corresponding to at least one reference radio frequency signal group comprising at least one of a second higher pulse repetition rate, a second longer length, or a second higher sensitivity than the first radio frequency signal group.

9. The method of claim 8, wherein, the one or more artificial intelligence algorithms comprise one or more machine learning algorithms.

10. The method of claim 9, wherein, the one or more machine learning algorithms comprise one or more neural networks.

11. The method of claim 8, further comprising wall filtering the ultrasound signals corresponding to the first radio frequency signal group.

12. The method of claim 8, wherein, the first radio frequency signal group is input into a neural network comprising a series of convolutional networks.

13. The method of claim 12, further comprising training the neural network, wherein, the training comprises: providing at least one third RF signal group having at least one of a same third length, a third PRF, or a third sensitivity as the first RF signal group; and providing a corresponding color Doppler image generated from the reference RF signal group as a desired output.

14. The method of claim 13, further comprising generating a plurality of reference RF signal groups from a subset of the at least one reference RF signal group.

15. The method of claim 13, wherein, the at least one third RF signal group and the desired output are acquired from at least one of a flow phantom, a simulation, or in vivo.

16. The method of claim 13, wherein, the training further comprises applying a masked mean squared error (MSE) loss.

17. The method of claim 16, wherein, the masked mean squared error (MSE) loss is based at least in part on power Doppler data of the reference RF signal group.

18. The method of claim 13, wherein, the training further comprises applying an adversarial loss.

19. The method of claim 18, wherein, the adversarial loss is based on a binary cross-entropy loss generated by training a discriminator network to distinguish between color Doppler images generated from the first RF signal group and color Doppler images generated from the reference RF signal group.

20. The method of claim 19, wherein, the discriminator network has a conditional generative adversarial network architecture.

21. The method of claim 13, wherein, the first RF signal group and the reference RF signal group are wall filtered.

22. The method of claim 8, wherein, the first RF signal group comprises an undersampled signal group.

23. The method of claim 12, further comprising processing the first RF signal group with the neural network to generate a first power Doppler image based at least in part on the at least one reference RF signal group.

24. A non-transitory computer-readable medium containing instructions that, when executed, cause an imaging system to: receive ultrasound signals corresponding to a first radio frequency signal group, the first radio frequency signal group comprising a first length, a first pulse repetition frequency, and a first sensitivity; and input the first RF signal group into one or more artificial intelligence algorithms configured to output a second RF signal group, a color Doppler image, or a power Doppler image, wherein the one or more artificial intelligence algorithms are trained using: reference ultrasound signals corresponding to at least one reference RF signal group comprising at least one of a higher second pulse repetition rate, a longer second length, or a higher second sensitivity than the first RF signal group.

25. The non-transitory computer-readable medium of claim 24, wherein, the one or more artificial intelligence algorithms comprise one or more machine learning algorithms.

26. The non-transitory computer-readable medium of claim 25, wherein, the one or more machine learning algorithms comprise one or more neural networks.

Images