Semi-supervised de-noising and de-aliasing for magnetic resonance imaging

Abstract

Systems and methods are provided for training a de-noising and de-aliasing machine learning model to generate de-noised and de-aliasing image data. The present disclosure provides techniques for training a de-noising and de-aliasing machine learning (ML) model to generate de-noised and de-aliasing imaging data. The method comprises: (1) training a first ML model using a first training data set comprising first image data to obtain a second ML model; and (2) training (a) the second ML model or (b) the third ML model using the second training data set to obtain a fourth ML model. The second training data set includes (i) the first image data and (ii) training image data obtained by applying at least one of the second ML model and the third ML model to the second image data. The de-noising and de-aliasing ML model may be a fourth ML model or derived from the fourth ML model.

Description

[0001] (This application is a divisional application of the application filed on May 26, 2023, with application number 202380097845.4 and invention title "Semi-supervised denoising and dealiasing for magnetic resonance imaging".) Technical Field

[0002] This application generally relates to reducing noise in medical imaging through machine learning, such as through two-step semi-supervised error correction and / or artifact correction (e.g., denoising and / or dealiasing) methods for reducing noise and / or artifacts in low-field diffuse magnetic resonance (MR) images. Background Technology

[0003] Magnetic resonance imaging (MRI) systems can be used to generate images of the human body's interior. MRI systems are used to detect magnetic resonance (MR) signals in response to an applied electromagnetic field. The MR signals generated by the MRI system can be processed to produce images, allowing observation of internal anatomical structures for diagnostic or research purposes. Accurately reconstructing the MR signals captured by the MRI system to a level sufficient for observing anatomical structures while removing sufficient noise is arguably challenging. Summary of the Invention

[0004] At least one aspect of this disclosure relates to a method for training a denoising and dealiasing machine learning (ML) model to generate denoised (or “de-denoised”) and / or de-aliased (or “de-aliased”) image data. The method includes: (1) training a first ML model using a first training dataset comprising first image data to obtain a second ML model; and (2) training (a) the second ML model or (b) a third ML model using a second training dataset to obtain a fourth ML model. The second training dataset comprises: (i) the first image data and (ii) training image data obtained by applying at least one of the second ML model and the third ML model to the second image data. The denoising and dealiasing ML model may be the fourth ML model or derived from the fourth ML model.

[0005] In some implementations, at least one of steps (1) and (2) includes supervised training processing. In some implementations, in step (2), a fourth ML model is obtained by training a third ML model using a second training dataset, and the third ML model has an architecture different from that of the second ML model. In some implementations, the second image data includes non-independent and non-identically distributed noise. In some implementations, the method may include applying a denoising and dealiasing ML model to the patient image to obtain a denoised patient image.

[0006] In some implementations, patient images are acquired using at least one of a low-field MR imaging system and a point-of-care (POC) MR imaging system. In some implementations, the first and second image data belong to separate domains. In some implementations, the method may include augmenting the training image data based on augmentation processing prior to step (2). In some implementations, the trained second ML model includes multiple convolutional neural network (CNN) layers.

[0007] In some implementations, the method may include generating a first training dataset by applying the raw imaging data to an image reconstruction pipeline. In some implementations, the method may include adding simulated image corruption to the raw imaging data. In some implementations, the third ML model is derived from a second ML model.

[0008] At least one other aspect of this disclosure relates to a method comprising: acquiring patient images using an imaging system; and applying a denoising and dealiasing machine learning (ML) model to the patient images to obtain denoised patient images. The denoising and dealiasing ML model may be obtained by: (1) training a first ML model using a first training dataset comprising first image data to obtain a second ML model; and (2) training (a) the second ML model or (b) a third ML model using a second training dataset to obtain a fourth ML model. The second training dataset comprises: (i) the first image data and (ii) training image data obtained by applying at least one of the second ML model and the third ML model to the second image data. The denoising and dealiasing ML model may be the fourth ML model or derived from the fourth ML model.

[0009] In some implementations, at least one of a low-field MR imaging system and a POC MR imaging system can be used to acquire patient images.

[0010] Another aspect of this disclosure relates to a system comprising: an imaging system configured to generate imaging data; and one or more processors configured to cause the imaging system to generate patient images and apply a denoising and dealiasing ML model to the patient images to generate denoised patient images. The denoising and dealiasing ML model can be obtained by: obtaining a first ML model using a first training dataset including first image data; and obtaining a second ML model using a second training dataset. The second training dataset includes: (i) the first image data and (ii) training image data obtained by applying at least one of the first ML model and a third ML model to the second image data. The denoising and dealiasing ML model can be the second ML model or derived from the second ML model.

[0011] These and other aspects and implementations are discussed in detail below. The foregoing information and the following detailed description include illustrative examples of the aspects and implementations and provide an overview or framework for understanding the nature and characteristics of the claimed aspects and implementations. The accompanying drawings provide illustrations and further understanding of the aspects and implementations and are incorporated in and constitute a part of this specification. The aspects may be combined, and it will be readily understood that features described in the context of one aspect of this disclosure may be combined with other aspects. The aspects may be implemented in any convenient form. In a non-limiting example, they may be carried on a suitable carrier medium (computer-readable medium) by a suitable computer program, which may be a tangible carrier medium (e.g., a disk) or an intangible carrier medium (e.g., a communication signal). The aspects may also be implemented using suitable devices, which may take the form of a programmable computer running a computer program arranged to implement the aspect. As used in the specification and claims, the singular forms “a,” “an,” and “the” include plural indicators unless the context clearly indicates otherwise. Attached Figure Description

[0012] The accompanying drawings are not intended to be drawn to scale. The same reference numerals and names in the various drawings indicate the same elements. For clarity, not every component may be labeled in every drawing. In the drawings:

[0013] Figure 1A Example components of a magnetic resonance imaging system according to one or more implementations are illustrated;

[0014] Figure 1B An example system is illustrated, which generates denoised image data based on one or more implementations of a machine learning model for training denoising and dealiasing.

[0015] Figure 2 A schematic diagram of an example data flow is depicted for a pipeline that reconstructs an example magnetic resonance image from one or more implementations.

[0016] Figure 3 A schematic diagram of an example data flow for the first stage of training processing based on one or more implementations of a machine learning model for training denoising and dealiasing to generate denoised image data is depicted.

[0017] Figure 4 A schematic diagram of an example data flow for the second stage of training processing based on one or more implementations of a machine learning model for training denoising and dealiasing to generate denoised image data is provided.

[0018] Figure 5 A flowchart is provided illustrating an example method for generating denoised image data by training a denoising and dealiasing machine learning model based on one or more implementations.

[0019] Figure 6A and Figure 6B This paper presents a comparison of example experimental data from one or more implementations for denoising magnetic resonance images; and

[0020] Figure 7 It is a block diagram of an example computing system applicable to the various arrangements described herein, implemented based on one or more examples. Detailed Implementation

[0021] The following is a detailed description of various concepts and implementations related to techniques, approaches, methods, devices, and systems used for training denoising and dealiasing machine learning models to generate denoised (e.g., noise removal or reduction) and / or dealiasing (e.g., removal or reduction of aliasing artifacts) or otherwise error correction and / or artifact correction image data. The various concepts introduced above and discussed in detail below can be implemented in any of a variety of ways, as the described concepts are not limited to any particular implementation. Examples of specific implementations and applications are provided primarily for illustrative purposes.

[0022] Magnetic resonance imaging (MRI) systems generate images for health assessment. MRI images are generated by “scanning” a patient while the MRI system applies a magnetic field to the patient and captures specific data. MRI scans produce raw scan data that can be transformed or otherwise processed into images, which can then be analyzed or examined to better assess the patient's health. MRI scans, which typically take longer, capture more raw data that can be used to generate images, while faster MRI scans, which require significantly less time for the patient to be on the MRI system, can generate images from less raw scan data. To enable faster scans with high image quality, MRI data is processed differently.

[0023] Not all data from an imaging system is available, and this unavailable data is called "noise," which can be misleading when interpreting images if not considered. An important consideration when processing medical images is increasing the ratio of available data (the "signal" being sought) to noise in the data. This is called the "signal-to-noise ratio," often simply called SNR. Since faster scans produce less raw data, ensuring a high "signal" score relative to the "noise" score allows faster scans to reveal more.

[0024] Faster scans can utilize weaker magnetic field strengths. In clinical low-field MRI, magnetic resonance (MR) sequences are designed to achieve a reasonable scan rate (SNR) within an acceptable scan time. While advantageously requiring significantly less patient time on the MRI system, faster MRI scans can generate images from less raw scan data compared to longer high-field MRI scans, but with a relatively lower SNR. Image denoising and dealiasing techniques can be used to further improve the SNR of scans captured using fast low-field MRI systems. Increased SNR improves the accuracy of various downstream processing tasks and can further reduce the scan time used by low-field MR systems.

[0025] Machine learning can be used to teach computers to perform tasks such as transforming raw scan data into images and reducing noise (“denoising” or “denoising”) without requiring them to be specifically programmed. This is particularly useful, for example, when constructing images from rapid raw scan data that can vary greatly from one patient to the next. This provides a machine learning model that has already been learned to perform a specific task, but the effectiveness of that model in different situations can vary greatly depending on how the model is trained (or “taught”) for the task. One type of machine learning approach is called “deep learning” and is based on multiple layers or levels of artificial neural networks.

[0026] While deep learning-based image denoising and dealiasing methods are successful for various image denoising and dealiasing tasks, they typically require a sufficiently large dataset of "clean" images (i.e., low-noise or denoised images) for training. Obtaining such "clean" images for MRI in clinical settings is difficult or otherwise infeasible. In such cases, unsupervised denoising and dealiasing can be used to train deep learning-based denoising and dealiasing models without requiring clean data. Such methods typically require noise to be independent and identically distributed (iid) in the image. In contrast, noise distributions in images obtained through complex MR reconstruction pipelines may be non-iid and therefore incompatible with such methods.

[0027] To address these shortcomings, the technical solution disclosed herein employs a two-stage processing approach to train denoising and dealiasing machine learning models to generate denoised image data. The technique described herein can effectively remove relevant MR noise without requiring clean images from the target domain (e.g., the domain of clinical MR images captured from a low-field MR system). In the first training step, supervised training processing can be performed to train denoising and dealiasing machine learning models (e.g., a denoising and dealiasing convolutional neural network (DNCNN) or similar) using a training set from the source domain (e.g., the domain of MR images captured using high-field MRI and having a clean reference image).

[0028] Once the denoising and dealiasing machine learning model is trained, it can be applied to a variety of images captured from the target domain (e.g., low-field MR images not necessarily associated with a clean reference image). The output of the denoising and dealiasing machine learning model, when performed on images from the target domain, undergoes data augmentation (including, for example, image sharpening, affine transformation, elastic deformation, insertion or addition of different geometric objects, intensity enhancement, etc., or any combination thereof) to increase the size of the dataset, which is then included as part of a second training set to retrain the denoising and dealiasing machine learning model. The denoising and dealiasing machine learning model can then be retrained using supervised learning methods leveraging the second training set.

[0029] The techniques described herein can be extended to accommodate challenging clinical MRI reconstructions on portable low-field MRI systems (e.g., less than about 0.5T, less than about 0.2T, between about 100mT and about 400mT, between about 200mT and about 300mT, between about 1mT and 100mT, between about 50mT and about 100mT, between about 40mT and about 80mT, and about 64mT, etc.). Advantages of the techniques described herein include the ability to train denoising and dealiasing machine learning models using data that includes relevant noise but not a clean reference image. The techniques described herein offer competitive performance compared to unsupervised methods and are robust across varying noise levels. Therefore, the systems and methods described herein provide technical improvements over conventional MRI image denoising and dealiasing methods.

[0030] Figure 1A An example MRI system is illustrated that can be used with a denoising and dealiasing model trained using the techniques described herein. Figure 1A In this MRI system 100, a computing device 104, a controller 106, a pulse sequence storage library 108, a power management system 110, and a magnetic component 120 may be included. The MRI system 100 is exemplary, and in addition to... Figure 1A Other than or in place of the components shown Figure 1A The MRI system may also have one or more other components of any suitable type, as shown in the components illustrated. Additionally, the implementation of components for a particular MRI system may differ from that described herein. Examples of low-field MRI systems may include portable MRI systems that may have field strengths of less than or equal to 0.5 T, less than or equal to 0.2 T, in the range of 1 mT to 100 mT, in the range of 50 mT to 0.1 T, in the range of 40 mT to 80 mT, approximately 64 mT, etc., in non-limiting examples.

[0031] Magnetic component 120 may include a B0 magnet 122, a spacer 124, radio frequency (RF) transmitting and receiving coils 126, and a gradient coil 128. The B0 magnet 122 can be used to generate a main magnetic field B0. The B0 magnet 122 can be any suitable type or combination of magnetic components capable of generating a useful main magnetic field B0. In some embodiments, the B0 magnet 122 may be one or more permanent magnets, one or more electromagnets, one or more superconducting magnets, or a hybrid magnet comprising one or more permanent magnets and one or more electromagnets or one or more superconducting magnets. In some embodiments, the B0 magnet 122 may be configured to generate a B0 magnetic field having a field strength less than or equal to 0.2 T or in the range of 50 mT to 0.1 T.

[0032] In some implementations, the B0 magnet 122 may include a first B0 magnet and a second B0 magnet, each of which may include a permanent magnet block arranged in a concentric ring around a common center. The first and second B0 magnets may be arranged in a biplane configuration such that the imaging region can be located between the first and second B0 magnets. In some embodiments, the first and second B0 magnets may each be coupled to and supported by a ferromagnetic yoke configured to capture and guide magnetic flux from the first and second B0 magnets.

[0033] Gradient coil 128 can be arranged to provide a gradient field, and in a non-limiting example, it can be arranged to generate gradients in three substantially orthogonal directions (X, Y, and Z) within the B0 field. Gradient coil 128 can be configured to encode the transmitted MR signal by systematically varying the B0 field (generated by B0 magnet 122 or shim 124) to encode the spatial location of the received MR signal as a function of frequency or phase. In a non-limiting example, gradient coil 128 can be configured to vary the frequency or phase as a linear function of spatial location along a particular direction, although a more complex spatial encoding distribution can also be provided by using a non-linear gradient coil. In some embodiments, in a non-limiting example, gradient coil 128 can be implemented using a laminate (e.g., a printed circuit board).

[0034] MRI scans are performed by exciting and detecting the emitted MR signals using transmit and receive coils (referred to herein as radio frequency (RF) coils). The transmit and receive coils may include separate coils for transmitting and receiving, multiple coils for transmitting or receiving, or the same coil for both. Thus, the transmit / receive assembly may include one or more coils for transmitting, one or more coils for receiving, or one or more coils for both transmitting and receiving. The transmit / receive coil may be referred to as a Tx / Rx or Tx / Rx coil to generally refer to various configurations of the transmit and receive magnetic assemblies of an MRI system. These terms are used interchangeably herein. Figure 1A In this configuration, the RF transmitting and receiving coil 126 may include one or more transmitting coils that can be used to generate RF pulses to induce an oscillating magnetic field B1. The transmitting coils (one or more) may be configured to generate any suitable type of RF pulse.

[0035] The power management system 110 includes electronic devices for providing operating power to one or more components of the MRI system 100. In a non-limiting example, the power management system 110 may include one or more power sources, energy storage devices, gradient power components, transmit coil assemblies, or any other suitable power electronic devices necessary to provide suitable operating power to energize and operate the components of the MRI system 100. Figure 1A As shown, the power management system 110 may include a power supply system 112, an amplifier (one or more) 114, a transmitting / receiving circuit system 116, and may optionally include a thermal management component 118 (e.g., a cryogenic cooling device for a superconducting magnet, a water cooling device for an electromagnet).

[0036] Power system 112 may include electronics for providing operating power to the magnetic components 120 of MRI system 100. In a non-limiting example, the electronics of power system 112 may provide operating power to one or more gradient coils (e.g., gradient coil 128) to generate one or more gradient magnetic fields, thereby providing spatial encoding of the MR signal. Additionally, the electronics of power system 112 may provide operating power to one or more RF coils (e.g., RF transmit and receive coil 126) to generate or receive one or more RF signals from a subject. In a non-limiting example, power system 112 may include a power source configured to provide power from mains power to the MRI system or an energy storage device. In some embodiments, the power source may be an AC-to-DC power source for converting AC power from mains power to DC power for use by the MRI system. In some embodiments, the energy storage device may be any of a battery, capacitor, supercapacitor, flywheel, or any other suitable energy storage device capable of bidirectionally receiving (e.g., storing) power from mains power and supplying power to the MRI system. Additionally, the power system 112 may include additional power electronic devices, including but not limited to power converters, switches, buses, drivers, and any other suitable electronic devices for supplying power to the MRI system.

[0037] Amplifier 114 (one or more) may include one or more RF receive (Rx) preamplifiers for amplifying MR signals detected by one or more RF receive coils (e.g., coil 126), one or more RF transmit (Tx) power components configured to power one or more RF transmit coils (e.g., coil 126), one or more gradient power components configured to power one or more gradient coils (e.g., gradient coil 128), and may power one or more shim power components configured to power one or more shims (e.g., shim 124). In some implementations, shim 124 may be implemented using a permanent magnet, an electromagnet (e.g., a coil), or a combination thereof. Transmit / receive circuitry 116 may be used to select whether the RF transmit coil or the RF receive coil is being operated.

[0038] like Figure 1AAs shown, the MRI system 100 may include a controller 106 (also referred to as a console) that may include control electronics for sending instructions to and receiving information from a power management system 110. The controller 106 may be configured to implement one or more pulse sequences that determine instructions sent to the power management system 110 to operate the magnetic component 120 according to a desired sequence (e.g., parameters for operating the RF transmit and receive coils 126, parameters for operating the gradient coil 128, etc.). The pulse sequence may generally describe the order and timing of the operation of the RF transmit and receive coils 126 and the gradient coil 128 to acquire resulting MR data. In a non-limiting example, the pulse sequence may indicate the order and duration of transmit pulses, gradient pulses, and the acquisition time of the receive coils to acquire MR data.

[0039] Pulse sequences can be organized into a series of time periods. In a non-limiting example, a pulse sequence may include a pre-programmed number of pulse repetition periods, and applying a pulse sequence may include operating an MRI system according to parameters of the pulse sequence during the pre-programmed number of pulse repetition periods. Within each time period, the pulse sequence may include parameters for generating RF pulses (e.g., parameters identifying transmission duration, waveform, amplitude, phase, etc.), parameters for generating gradient fields (e.g., parameters identifying transmission duration, waveform, amplitude, phase, etc.), timing parameters controlling when to generate RF or gradient pulses or when to configure (one or more) receiving coils to detect MR signals generated by the subject, and other functionalities. As described herein, in some embodiments, the pulse sequence may include parameters for specifying one or more navigation RF pulses.

[0040] Examples of pulse sequences include zero-echo-time (ZTE) pulse sequences, equilibrium steady-state free precession (bSSFP) pulse sequences, gradient echo pulse sequences, inversion recovery pulse sequences, diffusion-weighted imaging (DWI) pulse sequences, spin echo pulse sequences (including conventional spin echo (CSE) pulse sequences, fast spin echo (FSE) pulse sequences, turbine spin echo (TSE) pulse sequences, or any multi-spin echo pulse sequence such as diffusion-weighted spin echo pulse sequences, inversion recovery spin echo pulse sequences, arterial spin labeling pulse sequences, etc.), and Overhauser imaging pulse sequences, etc.

[0041] like Figure 1AAs shown, controller 106 can communicate with computing device 104, which can be programmed to process received MR data. In a non-limiting example, computing device 104 can process the received MR data using any suitable image reconstruction processing (including the execution of a denoising and dealiasing machine learning model trained using the techniques described herein) to generate one or more MR images. Additionally or alternatively, controller 106 can process the received MR data using any suitable image denoising and / or dealiasing processing to generate one or more denoised and / or dealiased MR images. Controller 106 can provide computing device 104 with information relating to one or more pulse sequences for the computing device to process the data. In a non-limiting example, controller 106 can provide computing device 104 with information relating to one or more pulse sequences, and computing device 104 can perform image denoising and / or dealiasing processing at least in part based on the provided information.

[0042] The computing device 104 may be any electronic device configured to process acquired MR data and generate one or more images of the subject being imaged. The computing device 104 may include at least one processor and a memory (e.g., processing circuitry). The memory may store processor-executable instructions that, when executed by the processor, cause the processor to perform one or more of the operations described herein. The processor may include a microprocessor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a graphics processing unit (GPU), a tensor processing unit (TPU), and combinations thereof. The memory may include, but is not limited to, electronic, optical, magnetic storage or transmission devices or any other storage or transmission device capable of providing program instructions to the processor. The memory may also include floppy disks, CD-ROMs, DVDs, magnetic disks, memory chips, ASICs, FPGAs, read-only memory (ROM), random access memory (RAM), electrically erasable programmable ROM (EEPROM), erasable programmable ROM (EPROM), flash memory, optical media, or any other suitable memory from which the processor can read instructions. The instructions may include code generated from any suitable computer programming language. The computing device 104 may include a combination of Figure 7 The computer system 700 is described in terms of any or all components, and performs any or all functions of the computer system 700. In some implementations, the computing device 104 may be located in the same room as the MRI system 100, or may be coupled to the MRI system 100 via a wired or wireless connection.

[0043] In some implementations, computing device 104 may be a fixed electronic device, such as a desktop computer, server, rack-mount computer, or any other suitable fixed electronic device that can be configured to process MR data and generate one or more images of the subject being imaged. Alternatively, computing device 104 may be a portable device, such as a smartphone, personal digital assistant, laptop computer, tablet computer, or any other portable device that can be configured to process MR data and generate one or more images of the subject being imaged. In some implementations, computing device 104 may include multiple computing devices of any suitable type, as the aspects of this disclosure provided herein are not limited in this respect. In some implementations, operations described as being performed by computing device 104 may alternatively be performed by controller 106, or vice versa. In some implementations, certain operations may be performed by both controller 106 and computing device 104 via communication between the devices.

[0044] The MRI system 100 may include one or more external sensors 178. These external sensors may assist in detecting one or more error sources (e.g., motion, noise) that degrade image quality. A controller 106 may be configured to receive information from one or more external sensors 178. In some embodiments, the controller 106 of the MRI system 100 may be configured to control the operation of one or more external sensors 178 and to collect information from one or more external sensors 178. Data collected from one or more external sensors 178 may be stored in a suitable computer memory and may be used to assist various processing operations of the MRI system 100.

[0045] As described above, the techniques described herein can be used to train denoising and dealiasing machine learning models for images in target domains where clean references may not be available. This enables the training of machine learning models that outperform supervised and unsupervised techniques in the target domain in terms of accuracy, and that are capable of denoising and dealiasing images where noise is non-iID.

[0046] The training processing described in this paper can be used to train accurate models based on undersampled and non-Cartesian MR data.

[0047] Figure 1B An example system 150 is illustrated, based on one or more implementations of a machine learning model for training denoising and dealiasing to generate denoised and / or dealiased image data. In a non-limiting example, system 150 can be used for combination. Figure 5 The example method 500 described herein, in whole or in part, and any other operations described herein. In some implementations, system 150 forms, for example, a combination... Figure 1AThe described MRI system 100 is part of an MRI system. In some implementations, system 150 may be external to the MRI system, but communicate with the MRI system (or its components) to perform as described herein. Figure 5 Example method 500.

[0048] like Figure 1B As shown, an embodiment of example system 150 may include controller 106, training platform 160, and user interface 176. User interface 176 may present or enable the examination of any reconstructed MR images generated using the techniques described herein. In a non-limiting example, user interface 176 may provide input related to performing such techniques by receiving input or configuration data related to training processing, MR scanning, or MR image reconstruction. User interface 176 may enable a user to select the type of imaging to be performed by the MRI system (e.g., diffusion-weighted imaging, etc.), select the sampling density for the MR scan, or define any other type of parameters related to MR imaging or model training as described herein. In some implementations, user interface 176 may display reconstructed and denoised and / or dealiased images generated from MR data acquired from the MRI system via a display communicating with user interface 176. User interface 176 may enable a user to initiate imaging with the MRI system or to perform or coordinate any machine learning techniques described herein.

[0049] In a non-limiting example, controller 106 can control various aspects of example system 150 for integration. Figure 5 The example method 500 described herein includes at least a portion of it, as well as any other operations described herein. In some implementations, controller 106 can control operations such as... Figure 1A The described MRI system 100 and other MRI systems include one or more operations. Additionally or alternatively, Figure 1A The computing device 104 can perform some or all of the functionality of the controller 106. In such an implementation, the computing device 104 can communicate with the controller 106 to exchange information as needed to achieve useful results.

[0050] Controller 106 may be implemented using software, hardware, or a combination thereof. Controller 106 may include at least one processor and memory (e.g., processing circuitry). The memory may store processor-executable instructions that, when executed by the processor, cause the processor to perform one or more of the operations described herein. The processor may include a microprocessor, ASIC, FPGA, GPU, TPU, etc., or a combination thereof. The memory may include, but is not limited to, electronic, optical, magnetic storage or transmission devices or any other storage or transmission device capable of providing program instructions to the processor. The memory may also include floppy disks, CD-ROMs, DVDs, magnetic disks, memory chips, ASICs, FPGAs, ROMs, RAMs, EEPROMs, EPROMs, flash memory, optical media, or any other suitable memory from which the processor can read instructions. Instructions may include code generated from any suitable computer programming language. Controller 106 may include combinations of... Figure 7 Any or all components of the computer system 700 described, and performing any or all functions of the computer system 700.

[0051] Controller 106 can be configured to perform one or more functions described herein. Controller 106 can store or capture MR spatial frequency data 170. MR systems (such as those combined with...) can be used. Figure 1A The described MRI system 100, etc., is used to acquire MR spatial frequency data 170. In some implementations, the MR spatial frequency data 170 may be acquired externally and provided to the controller 106 via one or more communication interfaces. The MR spatial frequency data 170 may be undersampled relative to the Nyquist sampling criterion. In some embodiments, the spatial frequency domain data may include a number of data samples less than 90% (or less than 80%, or less than 75%, or less than 70%, or less than 65%, or less than 60%, or less than 55%, or less than 50%, or less than 40%, or less than 35%, or any percentage between 25% and 100%) of the number of data samples required by the Nyquist criterion. Similarly, the MR spatial frequency data 170 may be non-Cartesian data. As described herein, the MR spatial frequency data 170 may be represented in the k-space domain. The MR spatial frequency data 170 may be generated by an MR scanner that can utilize suitable pulse sequences and sampling techniques. In some implementations, a Cartesian sampling scheme may be used to collect the MR spatial frequency data 170. Alternatively, non-Cartesian sampling schemes such as radial, spiral, rose, or Lissajou sampling schemes can be used to generate MR spatial frequency data170.

[0052] Controller 106 may include machine learning model executor 172. Machine learning model executor 172 may execute an image reconstruction pipeline to generate a reconstructed image from MR spatial frequency data 170. Machine learning model executor 172 may use the reconstructed image as input to execute a denoising and dealiasing machine learning model (such as machine learning model 168, which may be stored in the memory of controller 106 or computing device 104 in some implementations) to generate a denoised and / or dealiased image 174. Machine learning model 168 may be similar to or may include any denoising and dealiasing model described herein. As described herein, machine learning model 168 may be or may include a variational reconstruction network. Machine learning model executor 172 may use machine learning model 168 as input to execute machine learning model 168 to generate a denoised and / or dealiased image 174 (e.g., as part of an image reconstruction pipeline, etc.).

[0053] In a non-limiting example, the machine learning model 168 can be implemented by the training platform 160. Figure 5 Example method 500 is used for training. As described further in detail herein, machine learning model 168 can generate a denoised and / or dealiased image 174 from a reconstructed image generated based on MR spatial frequency data 170. In a non-limiting example, the denoised and / or dealiased image 174 generated by machine learning model 168 can be presented for inspection by a user at user interface 176. The denoised and / or dealiased image 174 can be stored in one or more data structures in the memory of controller 106 at the time of generation.

[0054] Training platform 160 may be or may include Figure 1A The computing device 104. Alternatively, the training platform 160 (or any component thereof) may be implemented as part of the controller 106. The training platform 160 may include at least one processor and memory (e.g., processing circuitry). The memory may store processor-executable instructions that, when executed by the processor, cause the processor to perform one or more of the operations described herein. The processor may include a microprocessor, ASIC, FPGA, GPU, TPU, etc., or a combination thereof. The memory may include, but is not limited to, electronic, optical, magnetic storage or transmission devices or any other storage or transmission device capable of providing program instructions to the processor. The memory may also include floppy disks, CD-ROMs, DVDs, magnetic disks, memory chips, ASICs, FPGAs, ROMs, RAMs, EEPROMs, EPROMs, flash memory, optical media, or any other suitable memory from which the processor can read instructions. Instructions may include code generated from any suitable computer programming language. The training platform 160 may include combinations of Figure 7The computer system 700 described herein includes any or all components and performs any or all functions of the computer system 700. In some implementations, the training platform 160 may be a desktop computer, server, rack computer, distributed computing environment, or any other computing system that can be configured to train the machine learning model 168 using the detraining techniques described herein. The training platform 160 may include any number of any suitable type of computing devices.

[0055] Training platform 160 may include a first set 162 of MR training data, a second set 166 of MR training data, a model training component 164, and a machine learning model 168 (e.g., which may be trained and retrained by model training component 164 as described herein). Model training component 164 may be implemented using any suitable combination of software or hardware. Additionally or alternatively, model training component 164 may be implemented by one or more servers or distributed computing systems, including cloud computing systems. In some implementations, model training component 164 may be implemented using one or more virtual servers or computing systems. Model training component 164 may be combined with... Figure 5 The example method 500 described is used to train the machine learning model 168, as well as any other operations related to the training of the denoising and dealiasing model as described herein. These training processes can be similar to combining... Figure 3 and Figure 4 The training processes described are at various levels, which can be implemented by the model training component 164 to train the machine learning model 168.

[0056] As described herein, model training component 164 can train machine learning model 168 using a first set 162 of MR training data and a second set 166 of MR training data. The first set 162 of MR training data can store multiple batches of MR spatial frequency data associated with corresponding clean reference images (e.g., images excluding noise or with noise removed). The first set 162 of MR training data can include raw data or reconstructed images that include noise captured using a high-field MRI system. In a non-limiting example, the first set 162 of MR training data can include images from a source domain (e.g., images captured using different types of MR systems, images captured from a specific patient population, etc.).

[0057] The first set 162 of MR training data may be previously generated by an MR scanner (e.g., including multiple historical MRI scans). The first set 162 of MR training data may include images reconstructed from MR spatial frequency data (e.g., k-space domain data, non-Cartesian data, etc.). In a non-limiting example, the reconstructed images in the MR training data repository 162 may be enhanced by applying affine transformations to create images with different orientations and sizes, by adding noise to create images with different SNRs, by introducing motion artifacts, by incorporating phase or signal modulation for more complex sequences (such as echo trains), or by modeling the dephasing of the data to adapt the model to diffusion-weighted imaging of sequence-like patterns.

[0058] In a non-limiting example, model training component 164 can perform any of the functionalities described herein to train machine learning model 168, including performing the two-stage training processes described herein. In the first training step, supervised training processes can be performed to train machine learning model 168 using a first set 162 of MR training data (e.g., based on clean reference images included therein), which machine learning model 168 may include DNCNN or another suitable denoising and dealiasing model as described herein. Once machine learning model 168 has been trained, model training component 164 can apply machine learning model 168 to images in the target domain (e.g., captured from a patient population corresponding to the target domain using a low-field MRI system). Images in the target domain may not necessarily be associated with clean reference images. Model training component 164 can utilize the output generated when performing machine learning model 168 on images in the target domain as clean reference images for retraining.

[0059] Images from the target domain (including the output of machine learning model 168) can be stored as part of a second set 166 of MR training data. In some embodiments, the second set 166 of MR training data may combine images from the source domain (and their corresponding clean reference images) with images from the target domain (e.g., noisy and corresponding clean images generated using machine learning model 168). In some implementations, the second set 166 of MR training data may include only images corresponding to the target domain (e.g., noisy and clean). In some embodiments, model training component 164 may perform data augmentation (including, for example, image sharpening, affine transformation, elastic deformation, insertion or addition of different geometric objects and / or intensity enhancement) to increase the size of the second set of MR training data. Model training component 164 may then retrain machine learning model 168 based on the second set 166 of MR training data using the techniques described herein. As described in the text, once the machine learning model 168 has been retrained (e.g., the training process has been terminated), the training platform 160 can provide the trained machine learning model 170 to the controller 106, so that the machine learning model executor 172 can use the machine learning model 168 to generate a denoised and / or dealiased image 174.

[0060] Figure 2 A schematic diagram 200 illustrates an example data flow of an example magnetic resonance image reconstruction pipeline based on one or more implementations. The image reconstruction pipeline shown in schematic diagram 200 can use mathematical operations to transform raw scan data (e.g., k-space data) into image data. The raw input data can be non-Cartesian scan data. The operation of the image reconstruction pipeline can be represented by Equation 1 below.

[0061] (1)

[0062] In Equation 1 above, Corresponding to the coil decorrelation operation, Corresponding to the gridding operation, This corresponds to the coil combination operation, and abs(.) corresponds to the amplitude operation. Due to sampling artifacts, coil correlation, and Rician bias, the image reconstruction process shown in schematic 200 may result in spatially correlated non-uniform noise in the reconstructed image. The reconstruction pipeline used in further operations described herein is... express.

[0063] In step 205, the original input data is... Applied to coil decorrelation operation Original input data This includes data from MRI scans that can be converted into visible images. Raw input data. Includes noise (data represented by wavy lines in this document indicates that the data includes noise). Operation This can be a transformation operation, such as a pre-whitening matrix. Hermitian adjoint or conjugate transpose, etc. Transformation operations. The output can be used as input to the next stage of the image reconstruction pipeline.

[0064] In step 210, the transformation operation The output (e.g., dereferenced medical image data) can be provided as input to the gridding operation. Gridding operation This can include operations for transforming decorrelational medical image data from the spatial frequency domain (e.g., k-space data) to the image domain. Meshization operations. It can compensate for sampling density in non-Cartesian spatial frequency data. (Gridding operation) The output may include one or more medical images, each corresponding to a set of MR signals captured by the RF receiving coil.

[0065] In step 215, the meshing operation can be performed. The generated medical images are used in coil assembly operations. Coil combination operation Medical images (each corresponding to the MR signal response of multiple corresponding RF receiving coils) can be combined into a specified array. A single noisy medical image. In some embodiments, amplitude manipulation can be applied to coil combination manipulation. The output is used to produce noisy medical images 220 ( ).

[0066] Once the noisy medical image 220 has been generated, it can be fed as input to a machine learning model (e.g., machine learning model 168) to generate a denoised and / or dealiased image (e.g., which can be specified as...). The two-level training technique described in this paper can be used to train images for generating denoised and / or dealiased images. The machine learning model. In the unrestricted example, the original input data... This can correspond to the target domain used to train machine learning models using the techniques described in this paper.

[0067] Figure 3A sample data flow diagram 300 depicts the first stage of a training process based on one or more implementations of a machine learning model (e.g., Machine Learning Model 168, etc.) for training denoising and / or dealiasing image data. The two-step training process described herein can be utilized when a clean reference image is available in the source domain (e.g., an image captured from a patient population or MR system) but the corresponding clean reference image is not available in the target domain (e.g., from a clinical population using scans from the low-field MR system described herein). In the first stage of the two-stage training process, an initial model can be trained using training data available from the source domain. The trained model can then be performed on noisy images from the target domain to generate denoised and / or dealiased images, which can be used as part of a second set of training data for subsequent training steps. Data augmentation (such as image sharpening, etc.) can be performed to expand the second set of training data. Non-limiting examples of data augmentation techniques include affine transformations, elastic deformation, insertion or addition of different geometric objects and / or intensity enhancement, etc. The initial model can then be preserved based on denoised and / or dealiased images generated from noisy target domain data, rather than using simulated image corruption (e.g., noisy data, MR frequency data acquired at sub-Nyquist rates to simulate aliasing artifacts, etc.). This improves the accuracy of denoising and / or dealiasing processing because the second step of training uses data from the target domain used in inference and testing, rather than using simulated data from training and target domain data from inference and testing.

[0068] As shown in the figure, by adding simulated structured noise 305 to clean reference data 315 from the source domain ( To generate noisy images from the source domain. The first training set 310. In a non-restrictive example, structured noise 305 can be simulated by generating Gaussian noise and adding it to clean reference data 315. Other noises can also be generated, such as noise from a Poisson distribution and other types of random structured noise. Clean reference data 315 can be non-Cartesian frequency domain data (e.g., k-space data) from a previous MR scan. This is achieved via the image reconstruction pipeline. (for example, combining) Figure 2 The description of reconstructed pipelines, etc.) propagates the first training set to generate noisy images 320 ( Similar techniques can be used (e.g., by rebuilding pipelines). Propagate clean reference data 315 to generate the corresponding clean reference image 325. ).

[0069] Using the generated noisy image 320 and the clean reference image 325 from the source domain, an initial machine learning model can be trained using appropriate training techniques. In the unrestricted example, the initial machine learning model It can be a combination Figure 1B The machine learning model 168 is described. In one embodiment, the initial machine learning model... This can be a DNCNN with multiple convolutional layers (e.g., twenty convolutional layers). Each convolutional layer can have a predetermined kernel size (e.g., 3×3) and a predetermined stride or bias term. In some implementations, each convolutional layer can apply 64 filters to the data produced by the previous layer. In some implementations, the machine learning model... This can include one or more activation layers (e.g., rectified linear units (ReLU), etc.) positioned between convolutional layers. Initial machine learning model It may include an input layer that receives a noisy image as input and produces a denoised and / or dealiased image at the output layer. Initial machine learning model. It can include any suitable number of convolutional layers, activation layers, or other types of neural network layers (e.g., pooling layers, fully connected layers, etc.).

[0070] Training the initial machine learning model This can include supervised learning processes (e.g., stochastic gradient descent and backpropagation, Adam optimizer, etc.) to iteratively fine-tune the initial machine learning model. The trainable parameters. Any suitable loss function can be used to train the initial machine learning model. Such loss functions include L1 loss, L2 loss, mean squared error (MSE) loss, binary cross-entropy (BCE), classification cross-entropy (CC), or sparse classification cross-entropy (SCC) loss functions. These can be based on a machine learning model given a noisy image 320 as input. The output is used to calculate the loss based on the corresponding clean reference image 325. Supervised training processing is illustrated by the dashed arrows in diagram 300, indicating the machine learning model. The machine learning model is trained to generate a corresponding clean reference image 325 from a noisy input image 320 (generated by adding structured noise, as described in this paper). Having already been trained, it can combine Figure 4 The second stage of the training process described utilizes machine learning models. .

[0071] Figure 4 The description outlines a method for retraining denoising and dealiasing machine learning models based on one or more implementations (e.g., machine learning model 168, combined with...). Figure 3 Described machine learning model Example data stream diagram 400 of the second stage of training processing to generate denoised and / or dealiased image data. Figure 4 The second stage of the two-stage training process is shown, in which the machine learning model... This is used to generate clean images in the target domain from noisy images in the target domain (e.g., clinical data captured using a low-field MRI system). The clean images are then used with corresponding noisy images in the clinical domain to retrain the machine learning model. Thus, retrained machine learning is obtained. .

[0072] As shown in schematic diagram 400, a noisy image 430 (specified as) from the target domain (e.g., captured using a low-field MR system) ) is provided as a machine learning model The machine learning model is executed to produce a clean image 435 (specified as) from the target domain as input. ). Execute machine learning models This can include using machine learning models To propagate the input data (e.g., various noisy images 430 from the target domain). At this stage of the training process, such as combining... Figure 3 The machine learning model described The model has already been trained on images from the source domain. Individual noisy images 430 from the target domain can be fed into the machine learning model. , to generate a set 435 of corresponding clean images from the target domain.

[0073] In order to generate a second set of training data (e.g., Figure 1B The second set of MR training data (166 etc.) was used to retrain the machine learning model. Data augmentation can be performed to increase the size of the clean image (435) generated for the target domain. This can be applied to machine learning models. A non-exhaustive list of examples of data augmentation processing for the generated image includes image sharpening (e.g., using a random Gaussian kernel), various transformations (e.g., rotation, cropping, horizontal or vertical flipping), or other data augmentation techniques (e.g., affine transformation, elastic deformation, insertion or addition of different geometric objects, intensity enhancement, etc.). Where appropriate, one or more data augmentation techniques may also be applied to the noisy input image from the target domain. Data augmentation can be used to remove some residual noise (e.g., image sharpening) or to increase the size of the training dataset (e.g., copying and horizontal / vertical flipping). Applying data augmentation techniques to the clean image 435 of the target domain produces an enhanced clean image 440 of the target domain (specified as...). ).

[0074] Once the enhanced clean image 440 of the target domain has been generated using data augmentation techniques, the various enhanced clean images 440 of the target domain can be transformed. To generate corresponding enhanced clean frequency data 445 (represented as) for the target domain. As described in this paper, the enhanced clean frequency data 445 from the target domain, together with the noisy image 430 from the target domain, can be used to generate a second training set to retrain the machine learning model. .

[0075] The corresponding set of clean images can be used as the second training set 415 (denoted as...). As part of retraining machine learning models The second training set 415 may include spatial frequency data from the target domain (e.g., enhanced clean frequency data 445 of the target domain) and spatial frequency data corresponding to a clean reference image from the source domain (e.g., clean reference data 315). Thus, the second training set 415 may include spatial frequency data from the source domain and clean spatial frequency data generated for the target domain. In some embodiments, the second training set 415 includes only the clean spatial frequency data corresponding to the target domain (e.g., enhanced clean frequency data 445 of the target domain). In some embodiments, simulated noise 405 may be added to the second training set 415 using techniques similar to those described herein. In one embodiment, simulated noise 405 is added to the second training set 415 to generate a second set 410 of noisy spatial frequency data (designated as...). In some other embodiments, noise is not added to the second training set 415, and the transformation of the noisy image 430 in the target domain (e.g., from the image domain to the frequency domain) is used as the second set 410 of noisy spatial frequency data. The simulated noise 405 can be any type of simulated image corruption, including simulated noise or MR frequency data acquired at an sub-Nyquist rate to simulate aliasing artifacts.

[0076] Corresponding reconstruction transformation The second set 410 and the second training set 415, which can be applied to noisy spatial frequency data, can respectively generate a second set 420 of noisy images (specified as...). The second set 420 of noisy images and the second set 425 of clean images are used. In the embodiment that utilizes the transformation of the noisy image 430 of the target domain (e.g., from image to frequency domain) as the second set 410 of noisy spatial frequency data, the noisy image 430 from the target domain and the enhanced clean image 440 of the target domain can be used as the second set 420 of noisy images and the second set 425 of clean images, respectively, to retrain the machine learning model. .

[0077] Once the second set 420 of noisy images and the second set 425 of clean images have been generated, the machine learning model can be retrained. To obtain a retrained machine learning model In one embodiment, the machine learning model is retrained. This can include starting from scratch (e.g., from a machine learning model). Retrain the machine learning model starting from the default values ​​of the trainable parameters. In some embodiments, the machine learning model is retrained. It can be included in Figure 3 The first stage of training described in the text is followed by the use of a machine learning model. Trainable parameters for training machine learning models .

[0078] Retraining the machine learning model This can include supervised learning processes (e.g., stochastic gradient descent and backpropagation, Adam optimizer, etc.) to iteratively fine-tune the machine learning model. The trainable parameters are used to obtain a retrained machine learning model. As described in this article, any suitable loss function can be used to retrain a machine learning model. Such as L1 loss, L2 loss, MSE loss, BCE loss, CC loss, or SCC loss. A machine learning model can be based on the input of noisy images from a second set 420 of noisy images. The output is used to calculate the loss based on the corresponding clean reference images from the second set 425 of clean images. Supervised training processing is illustrated by the dashed arrows in diagram 400, indicating the retrained machine learning model. A second set 425 is trained to generate corresponding clean images from a second set 420 of noisy images.

[0079] Figure 5 Examples are given of training machine learning models using two-level supervised learning processes based on one or more implementations (e.g., Figure 1B A flowchart of an example method 500 for generating denoised and / or dealiased MR images (e.g., machine learning model 168, etc.). Method 500 can use any suitable computing system (e.g., training platform 160, controller 106, or computing device 104 in Figure 1). Figure 7The computational system 700, etc., can be used to execute this method. It should be understood that some steps of method 500 can be executed in parallel (e.g., simultaneously) or sequentially, while still achieving useful results. As described herein, method 500 can be executed iteratively to update or otherwise train the denoising and dealiasing machine learning model.

[0080] Method 500 may include action 505, wherein a first machine learning model (e.g., machine learning model 168) is initially trained using a first training set corresponding to the source domain (e.g., a first set 162 of MR training data) to obtain a second machine learning model. As described herein, the source domain may include images captured from a patient population that differ from images corresponding to the target domain. The source domain may include images captured using an MR system of a different type than the target domain or images captured from a specific patient population. In a non-limiting example, raw MR scan data (e.g., spatial frequency data) can be applied to an image reconstruction pipeline (such as combining...) Figure 2 The image reconstruction pipeline (described above) is used to generate an image corresponding to the source domain. Simulated noise data can be added to the original scan data before reconstruction to simulate a noisy image, which can be paired with a corresponding clean image (without simulated noise) for use as a reference for supervised learning. The simulated noise data can be any type of simulated image corruption. In embodiments, in non-limiting examples, images in the first training set can be enhanced by applying affine transformations to create images with different orientations and sizes, by adding noise to create images with different SNRs, by introducing motion artifacts, by incorporating phase or signal modulation for more complex sequences (such as echo trains), or by modeling dephasing of the data to adapt the model to diffusion-weighted imaging of sequence-like data.

[0081] The first and second machine learning models can be DNCNN models with multiple convolutional layers (e.g., twenty convolutional layers). Each convolutional layer can have a predetermined kernel size (e.g., 3×3) and a predetermined stride or bias term. In some implementations, each convolutional layer can apply 64 filters to the data generated by the previous layer. In a non-limiting example, training the first machine learning model can include performing supervised learning processes (e.g., stochastic gradient descent and backpropagation, Adam optimizer, etc.) to iteratively adjust the trainable parameters of the first machine learning model until predetermined training termination conditions have been met (e.g., predetermined model accuracy has been achieved, a predetermined amount of training data has been used to train the model, etc.). The first machine learning model can be trained using any suitable loss function, such as L1 loss, L2 loss, MSE loss, BCE loss, CC loss, or SCC loss function, etc. The loss can be calculated based on the output of the first machine learning model when a noisy image from the first training set is provided as input compared to a corresponding clean reference image in the first training set. Once trained using the first training set, the first machine learning model is referred to as the second machine learning model.

[0082] Method 500 may include action 510, wherein a second machine learning model obtained in action 505 is used to generate a denoised and / or dealiased training image (e.g., a clean image 435 of the target domain) corresponding to the target domain. The denoised and / or dealiased training image may be generated based on noisy images captured using a low-field MR system or a point-of-care (POC) MR imaging system. The noisy image corresponding to the target domain may include non-independent and non-identically distributed noise. To generate a clean image of the target domain, the second machine learning model obtained in action 505 may be executed using the noisy image corresponding to the target domain. Executing the second machine learning model may include propagating the various noisy images from the target domain through the trained second machine learning model until a corresponding clean output image is generated.

[0083] Method 500 may include action 515, in which a second training dataset is generated. The second training dataset may include denoised and / or dealiased training images generated in action 515 or images derived therefrom. In a non-limiting example, the second training dataset may be generated by performing data augmentation on images from the target domain (e.g., noisy images and / or clean images, as appropriate). Data augmentation may be used to remove some residual noise (e.g., image sharpening) or to increase the size of the training dataset (e.g., copying and horizontal / vertical flipping). A non-exhaustive list of examples of data augmentation processes that may be applied to images corresponding to the target domain includes image sharpening (e.g., using a random Gaussian kernel), various transformations (e.g., rotation, cropping, horizontal or vertical flipping), or other data augmentation techniques. Clean images corresponding to the target domain, along with their augmented deformations, may be included as part of the second training dataset, which may be used in action 525 to retrain a second machine learning model or in action 525 to train a third machine learning model.

[0084] Method 500 may include action 520, in which a second machine learning model trained in action 505 is retrained based on a second training set. In some implementations, simulated noise data may be added to a clean image corresponding to the target domain to train the second machine learning model. The simulated noise data can be any type of simulated image corruption. In a non-limiting example, a clean image with simulated noise may be propagated through the second machine learning model, which is then trained based on a loss computed using the corresponding clean image of the target domain. In another embodiment, as described herein, a noisy image corresponding to the target domain from which the clean image is generated may be used as input data, which is propagated through the second machine learning model, which is then retrained using the computed loss function.

[0085] The second machine learning model can be retrained from scratch, or it can be retrained based on its state after being trained on the source domain. In some implementations, overfitting treatment can be used to train the second machine learning model based on a second training set. Retraining the second machine learning model may include performing supervised learning processes (e.g., stochastic gradient descent and backpropagation, Adam optimizer, etc.) to iteratively readjust the trainable parameters of the second machine learning model, ultimately obtaining a retrained machine learning model. The retrained machine learning model can then be deployed and executed using patient images captured using a low-field MR imaging system or a POC MR imaging system to obtain denoised and / or dealiased patient images.

[0086] Method 500 may include action 525, in which a third machine learning model trained in action 505 is retrained based on a second training set. In some implementations, the third machine learning model may have the same architecture as the second machine learning model. In alternative embodiments, the third machine learning model may have a different architecture than the second machine learning model. The third machine learning model may be trained based on the second training set. In some implementations, simulated noise data may be added to a clean image corresponding to the target domain to train the third machine learning model. The simulated noise data may be any type of simulated image corruption. In a non-limiting example, a clean image with simulated noise may be propagated through the third machine learning model, which is then trained based on a loss computed using the corresponding clean image of the target domain. In another embodiment, as described herein, a noisy image corresponding to the target domain from which the clean image is generated may be used as input data, which is propagated through the third machine learning model, which is then trained using the computed loss function. Training the third machine learning model may include performing supervised learning processes (e.g., stochastic gradient descent and backpropagation, Adam optimizer, etc.) to iteratively readjust the trainable parameters of the machine learning model, ultimately obtaining a trained fourth machine learning model. A third machine learning model can be deployed and executed using patient images captured by a low-field MR imaging system or a POC MR imaging system to obtain denoised and / or dealiased patient images.

[0087] Figure 6A and Figure 6B Example experimental data comparison methods for denoising and / or dealiasing magnetic resonance images are shown based on one or more implementations. Figure 6A Example comparisons of outputs from different denoising and / or dealiasing implementations are shown. Figure 6A and Figure 6B The image depicts corresponding noisy MR images generated from MR data acquired using DWI pulse sequences, and four corresponding denoised and / or dealiased MR images generated from the noisy MR images using various denoising and / or dealiasing methods. Each MR image shows a magnified portion of the corresponding MR image (each larger bounding box is a magnification of its corresponding smaller bounding box). Dedenoising and / or dealiasing methods include block matching and 3D filtering (BM3D), Nr2N, and supervised learning (Sup) using images from the source domain (e.g., in...). Figure 3 The first-level training process (the machine learning model trained after the first level) and the Sequential Semi-Supervised Learning (SeqSSL), which is a retrained machine learning model generated using the two-level training process described in this paper, are used. For BM3D, hyperparameters are manually tuned for each slice. To train the Nr2N model, the model is trained to predict the image at σ = 0.05 from an input with a noise level of σ = 0.1.

[0088] The SeqSSL model was trained using 400 T1-weighted and T2-weighted images from the Human Connectome Project (HCP), along with 400 T1-weighted, T2-weighted, and fluid-attenuated inversion recovery (FLAIR) images acquired at 64 mT as image data from the source domain. Image data from the target domain consisted of DWI images captured using a low-field (e.g., 64 mT) MR system. Specifically, the target domain included b = 890 s / mm. 2 and b=0 s / mm 2 289 and 308 DWI images were obtained at the location, with TE / TR = 24 ms / 34 ms and a resolution of 2×2×6 mm. 3 The scan time was 8 minutes (for b=890 s / mm). 2 ) and 1.5 minutes (for b=0 s / mm) 2 The architecture of the machine learning model used is an unbiased DNCNN with 20 convolutional layers. The Adam optimizer is used, and patch-based training is combined with L1 and structural similarity index (SSIM).

[0089] like Figure 6A and Figure 6B As shown, the results from BM3D are competitive with this technique; however, the BM3D method requires careful, manual parameter tuning. Nr2N results in inconsistent levels of denoising and / or dealiasing. This is likely due to the fact that in practice, the noise variance present in images is variable, while Nr2N requires training with a fixed noise level. Some degree of oversmoothing can be observed in Sup. SeqSSL mitigates the oversmoothing problem and performs more consistently across all images.

[0090] The proposed error correction and / or artifact correction framework was qualitatively evaluated by four expert raters with backgrounds in MR physics, clinical science, and / or radiology. Images before and after denoising and / or dealiasing were presented to the raters as a side-by-side comparison. Users were asked to rate the denoised and / or dealiased images as “far better,” “significantly better,” “same,” “significantly worse,” or “far worse” in terms of noise, sharpness, and overall quality. Raters were also asked as input whether the denoised and / or dealiased images exhibited consistent image characteristics in terms of contrast, geometric fidelity, and the presence or absence of artifacts. The results of this qualitative evaluation are presented in Tables 1 and 2 below.

[0091] Table 1

[0092] Table 2

[0093] As the results above show, for all categories, all raters indicated "same," "significantly better," and "far better." At least 88.8%, 42.5%, and 82.5% voted "significantly better" and "far better," respectively, for reduced noise, sharpness, and overall quality. Raters also rated "yes" for all consistency issues.

[0094] Figure 7 This is a component diagram of an example computing system applicable to various implementations described herein, based on examples. In a non-limiting example, computing system 700 can be implemented... Figure 1A and Figure 1B The computing device 104, controller 106, or training platform 160, or various other example systems and devices described in this disclosure.

[0095] The computing system 700 includes a bus 702 or other communication components for communicating information and a processor 704 coupled to the bus 702 for processing information. The computing system 700 also includes a main memory 706, such as RAM or other dynamic storage devices, coupled to the bus 702 for storing information and instructions to be executed by the processor 704. The main memory 706 can also be used to store location information, temporary variables, or other intermediate information during instruction execution by the processor 704. The computing system 700 may also include a ROM 708 or other static storage device coupled to the bus 702 for storing static information and instructions used by the processor 704. A storage device 710, such as a solid-state device, a magnetic disk, or an optical disk, is coupled to the bus 702 for persistently storing information and instructions.

[0096] The computing system 700 can be coupled via a bus 702 to a display 714, such as a liquid crystal display or an active matrix display, for displaying information to a user. An input device 712, such as a keyboard including alphanumeric keys and other keys, can be coupled to the bus 702 for communicating information and command selection to the processor 704. In another implementation, the input device 712 has a touchscreen display. The input device 712 can include any type of biosensor or cursor control, such as a mouse, trackball, or cursor arrow keys, for communicating directional information and command selection to the processor 704 and for controlling cursor movement on the display 714.

[0097] In some implementations, computing system 700 may include a communication adapter 716, such as a network adapter. Communication adapter 716 may be coupled to bus 702 and may be configured to communicate with computing or communication networks or other computing systems. In various exemplary implementations, communication adapter 716 can be used to implement any type of network configuration, such as wired (e.g., via Ethernet), wireless (e.g., via Wi-Fi, Bluetooth), pre-configured satellite (e.g., via GPS), ad-hoc, LAN, and WAN, etc.

[0098] According to various implementations, the processing of the exemplary implementation described herein can be implemented by a computing system 700 in response to a processor 704 executing an implementation of instructions contained in main memory 706. Such instructions can be read into main memory 706 from another computer-readable medium, such as storage device 710. Execution of the implementation of the instructions contained in main memory 706 causes the computing system 700 to perform the exemplary processing described herein. Alternatively, one or more processors in a multiprocessing implementation can be used to execute the instructions contained in main memory 706. In alternative implementations, hardwired circuitry can be used to replace or combine with software instructions to implement the exemplary implementation. Therefore, the implementation is not limited to any particular combination of hardware circuitry and software.

[0099] Potential embodiments include, but are not limited to:

[0100] Example AA: A method comprising training a denoising and dealiasing machine learning model, i.e., a denoising and dealiasing ML model, to generate denoised and / or dealiased imaging data, wherein training the denoising and dealiasing ML model comprises: (1) training a first ML model using a first training dataset comprising first image data to obtain a second ML model; and (2) training (a) the second ML model or (b) a third ML model using a second training dataset to obtain a fourth ML model, wherein the second training dataset comprises (i) the first image data and (ii) training image data obtained by applying at least one of the second ML model and the third ML model to the second image data, and wherein the denoising and dealiasing ML model is the fourth ML model or derived from the fourth ML model.

[0101] Example AB: The method according to Example AA, wherein at least one of steps (1) and (2) includes supervised training processing.

[0102] Example AC: The method according to Example AA or AB, wherein, in step (2), the third ML model is trained by using the second training dataset to obtain the fourth ML model, and wherein the third ML model has an architecture different from that of the second ML model.

[0103] Example AD: The method according to any one of Examples AA to AC, wherein the second image data includes noise.

[0104] Example AE: The method according to Example AD, wherein the noise includes non-independent and non-uniformly distributed noise.

[0105] Example AF: The method according to any one of Examples AA to AE further includes: applying the denoising and dealiasing ML model to the patient image to obtain a denoised and / or dealiased patient image.

[0106] Example AG: The method according to Example AF, wherein the patient image is acquired using at least one of a low-field magnetic resonance imaging system, i.e., a low-field MR imaging system, and a real-time MR imaging system, i.e., a POC MR imaging system.

[0107] Example AH: The method according to any one of Examples AA to AG, wherein the first image data and the second image data belong to separate domains.

[0108] Example AI: The method according to any one of Examples AA to AH further includes: enhancing the training image data based on enhancement processing before step (2).

[0109] Example AJ: The method according to any one of Examples AA to AI, wherein the enhancement process includes image sharpening, affine transformation, elastic deformation, insertion or addition of different geometric objects, or any combination of intensity enhancement.

[0110] Example AK: The method according to any one of Examples AA to AJ, wherein the trained second ML model includes multiple convolutional neural network layers, i.e., multiple CNN layers.

[0111] Example AL: The method according to any one of Examples AA to AK further includes: generating the first training dataset by applying the raw imaging data to an image reconstruction pipeline.

[0112] Example AM: The method according to Example AL further includes adding simulated image damage to the original imaging data.

[0113] Example AN: The method according to Example AM, wherein the image corruption includes at least one of noise and / or aliasing artifacts.

[0114] Example AO: The method according to Example AM ​​or AN, wherein adding simulated image corruption includes adding simulated noise data.

[0115] Example AP: The method according to any one of Examples AM to AO, wherein adding simulated image corruption includes acquiring MR frequency data at a sub-Nyquist rate to simulate aliasing artifacts.

[0116] Example AQ: The method according to any one of Examples AA to AP, wherein the third ML model is derived from the second ML model.

[0117] Example AR: An apparatus or system capable of performing any of the methods described according to Examples AA to AQ.

[0118] Example BA: A method comprising: acquiring a patient image using an imaging system, and applying a denoising and dealiasing machine learning model, i.e., a denoising and dealiasing ML model, to the patient image to obtain a denoised and / or dealiased patient image, the denoising and dealiasing ML model being obtained by: (1) training a first ML model using a first training dataset comprising first image data to obtain a second ML model; and (2) training (a) the second ML model or (b) a third ML model using a second training dataset to obtain a fourth ML model, wherein the second training dataset comprises (i) the first image data and (ii) training image data obtained by applying at least one of the second ML model and the third ML model to the second image data, wherein the denoising and dealiasing ML model is the fourth ML model or derived from the fourth ML model.

[0119] Example BB: The method according to Example BA, wherein the patient image is acquired using at least one of a low-field MR imaging system and a POC MR imaging system.

[0120] Example BC: An apparatus or system capable of performing Example BA or BB.

[0121] Example CA: A system comprising: an imaging system configured to generate imaging data; and one or more processors configured to cause the imaging system to generate patient images and apply a denoising and dealiasing ML model to the patient images to generate denoised and / or dealiased patient images, the denoising and dealiasing ML model being obtained by: obtaining a first ML model using a first training dataset comprising first image data; and obtaining a second ML model using a second training dataset, wherein the second training dataset comprises (i) the first image data and (ii) training image data obtained by applying at least one of the first ML model and a third ML model to the second image data, and wherein the denoising and dealiasing ML model is the second ML model or derived from the second ML model.

[0122] Example CB: The system described in Example CA is used to perform any of the methods disclosed herein, such as any of the examples AA to BB.

[0123] Example DA: An apparatus or system for performing any of the methods disclosed herein (such as any of Examples AA to BB).

[0124] The implementations described herein are illustrated with reference to the accompanying drawings. The drawings illustrate certain details of specific implementations of the systems, methods, and programs described herein. The use of the drawings to describe the implementations should not be construed as imposing any limitations that may exist in the drawings on this disclosure.

[0125] It should be understood that unless the element is explicitly defined by the phrase “part for…”, the elements of the claims herein shall not be construed under the provisions of section 112(f) of title 35 of the United States Code.

[0126] As used herein, the term "circuit" can include hardware configured to perform the functions described herein. In some implementations, each respective "circuit" can include a machine-readable medium for configuring hardware to perform the functions described herein. A circuit can be embodied as one or more circuit system components, including but not limited to processing circuit systems, network interfaces, peripheral devices, input devices, output devices, sensors, etc. In some implementations, a circuit can take the form of one or more analog circuits, electronic circuits (e.g., integrated circuits (ICs), discrete circuits, system-on-a-chip (SoC) circuits), telecommunications circuits, hybrid circuits, and any other type of "circuit". In this respect, "circuit" can include any type of component for performing or facilitating the implementation of the operations described herein. In non-limiting examples, a circuit as described herein can include one or more transistors, logic gates (e.g., NAND, AND, NOR, OR, XOR, NOT, XNOR), resistors, multiplexers, registers, capacitors, inductors, diodes, wiring, etc.

[0127] The term "circuit" may also include one or more processors communicatively coupled to one or more memories or memory devices. In this respect, one or more processors may execute instructions stored in memory or instructions otherwise accessible to one or more processors. In some implementations, one or more processors may be embodied in various ways. One or more processors may be constructed in a manner sufficient to at least perform the operations described herein. In some implementations, one or more processors may be shared by multiple circuits (e.g., circuits A and B may include or otherwise share the same processor, which, in some example implementations, may execute instructions stored or otherwise accessed via different regions of memory). Alternatively or additionally, one or more processors may be configured to perform or otherwise execute certain operations independently of one or more coprocessors.

[0128] In other example embodiments, two or more processors may be coupled via a bus to enable independent, parallel, pipelined, or multithreaded instruction execution. Each processor may be implemented as one or more general-purpose processors, ASICs, FPGAs, GPUs, TPUs, digital signal processors (DSPs), or other suitable electronic data processing components configured to execute instructions provided by memory. One or more processors may take the form of a single-core processor, a multi-core processor (e.g., a dual-core, triple-core, or quad-core processor), a microprocessor, etc. In some implementations, one or more processors may be external to the device; in non-limiting examples, one or more processors may be remote processors (e.g., cloud-based processors). Alternatively or additionally, one or more processors may be internal to the device or local. In this regard, a given circuit or its components may be arranged locally (e.g., as part of a local server, local computing system) or remotely (e.g., as part of a remote server such as a cloud-based server). For this purpose, a “circuit” as described herein may include components distributed across one or more locations.

[0129] An exemplary system for implementing the entire system or a part thereof may include a general-purpose computing device in the form of a computer, comprising a processing unit, system memory, and a system bus for coupling various system components, including the system memory, to the processing unit. Various memory devices may include non-transitory volatile storage media, non-volatile memory media, non-transitory storage media (e.g., one or more volatile or non-volatile memories), etc. In some implementations, the non-volatile medium may take the form of ROM, flash memory (e.g., flash memory such as NAND, 3D NAND, NOR, 3D NOR, etc.), EEPROM, MRAM, magnetic storage, hard disk, optical disk, etc. In other implementations, the volatile storage medium may take the form of RAM, TRAM, ZRAM, etc. Combinations of the above are also included within the scope of machine-readable media. In this regard, in a non-limiting example, machine-executable instructions include instructions and data that cause a general-purpose computer, a special-purpose computer, or a special-purpose processor to perform certain functions or groups of functions. According to the example implementation described herein, each corresponding memory device may be operable to maintain or otherwise store information relating to operations performed by one or more associated circuits, including processor instructions and associated data (e.g., database components, object code components, script components).

[0130] It should also be noted that, as described herein, the term "input device" can include any type of input device, including but not limited to keyboards, keypads, mice, joysticks, or other input devices that perform similar functions. In contrast, as described herein, the term "output device" can include any type of output device, including but not limited to computer monitors, printers, fax machines, or other output devices that perform similar functions.

[0131] It should be noted that although the figures herein may illustrate a specific order and composition of method steps, it should be understood that the order of these steps may differ from that depicted. In non-limiting examples, two or more steps may be performed simultaneously or partially simultaneously. Furthermore, some method steps performed as discrete steps may be combined, steps performed as combined steps may be divided into discrete steps, the sequence of certain processes may be reversed or otherwise varied, and the nature or number of discrete processes may be altered or changed. According to alternative implementations, the order or sequence of any element or device may be varied or replaced. Therefore, all such variations are intended to be included within the scope of this disclosure as defined in the appended claims. Such variations will depend on the machine-readable medium and hardware system chosen, as well as the designer's choice. It should be understood that all such variations are within the scope of this disclosure. Similarly, the software and web implementations of this disclosure can be accomplished using standard programming techniques with rule-based logic and other logic to perform various database search steps, relevance steps, comparison steps, and decision steps.

[0132] While this specification contains numerous specific implementation details, these should not be construed as limiting the scope of any invention or claimable content, but rather as descriptions of features specific to the systems and methods described herein. Some features described in this specification in the context of individual implementations may also be implemented in combination within a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations. Furthermore, although features may be described above as functioning in certain combinations and even initially claimed in this way, in some cases, one or more features from the claimed combination may be removed from the combination, and the claimed combination may involve sub-combinations or variations of sub-combinations.

[0133] In some cases, multitasking and parallel processing can be advantageous. Furthermore, the separation of the various system components in the above implementation should not be interpreted as a requirement for such separation in all implementations, and it should be understood that the described program components and systems can often be integrated together in a single software product or encapsulated in multiple software products.

[0134] Some exemplary implementations and approaches have now been described. It is clear that the foregoing, presented by way of example, is illustrative and not restrictive. In particular, while many of the examples presented herein involve specific combinations of method actions or system elements, these actions and elements can be combined in other ways to accomplish the same goal. Actions, elements, and features discussed in conjunction with only one implementation are not intended to exclude similar roles in other implementations.

[0135] The wording and terminology used herein are for descriptive purposes and should not be considered limiting. The use of “including,” “containing,” “having,” “comprising,” “involving,” “characterized by,” “featured in,” and variations thereof is intended to cover items listed herein, their equivalents and additions, and alternative implementations that are exclusively comprised of the items listed thereafter. In one implementation, the system and method described herein consist of one, more than one, or all of the described elements, actions, or components.

[0136] Any reference to an implementation of a system or method, element, or action mentioned in the singular herein may include implementations that include multiple such elements, and any plural reference to any implementation, element, or action herein may include an implementation that includes only a single element. References in singular or plural form are not intended to limit the currently disclosed systems or methods, their components, actions, or elements to a single or plural configuration. A reference to any action or element based on any information, action, or element may include an action or element that is at least partially based on an implementation of that information, action, or element.

[0137] Any implementation disclosed herein may be combined with any other implementation, and references to “implementation,” “some implementations,” “alternative implementations,” “various implementations,” or “an implementation,” etc., are not necessarily mutually exclusive and are intended to indicate that a particular feature, structure, or characteristic described in connection with an implementation may be included in at least one implementation. Such terms as used herein do not necessarily refer to the same implementation. Any implementation may be combined inclusively or exclusively with any other implementation in any manner consistent with the aspects and implementations disclosed herein.

[0138] A reference to "or" can be interpreted as inclusive, such that any term described using "or" can refer to any one of the single, multiple, or all of the terms described.

[0139] Where a reference numeral follows a technical feature in an accompanying drawing, detailed description, or any claim, the reference numeral is included solely for the purpose of enhancing the comprehensibility of the drawing, detailed description, and claims. Therefore, the reference numeral, and its absence, have no limiting effect on the scope of any claim element.

[0140] The foregoing description of the implementations is presented for illustrative and descriptive purposes. The foregoing description is not intended to be exhaustive or to limit this disclosure to the precise form disclosed, and variations and modifications are possible, or can be derived from, this disclosure in light of the above teachings. The implementations have been chosen and described to explain the principles of this disclosure and its practical application, thereby enabling those skilled in the art to utilize various implementations and have various modifications suitable for the particular intended use. Other substitutions, modifications, alterations, and omissions may be made in the design, operating conditions, and implementation of the implementations without departing from the scope of this disclosure as set forth in the appended claims.

[0141] Cross-reference to related applications

[0142] This application claims priority to U.S. Provisional Patent Application No. 63 / 489,918, filed March 13, 2023, the contents of which are incorporated herein by reference in their entirety for all purposes.

Claims

1. A method comprising: Training a denoising and dealiasing machine learning model, i.e., a denoising and dealiasing ML model, to generate denoised and / or dealiased imaging data, wherein training the denoising and dealiasing ML model includes: (1) Train a first ML model using a first training dataset including the first image data to obtain a second ML model; and (2) Using a second training dataset to train (a) the second ML model or (b) the third ML model to obtain a fourth ML model, wherein the second training dataset includes (i) the first image data and (ii) training image data obtained by applying at least one of the second ML model and the third ML model to the second image data, and wherein the denoising and dealiasing ML model is the fourth ML model or derived from the fourth ML model.

2. The method according to claim 1, wherein, At least one of steps (1) and (2) includes supervised training processing.

3. The method according to claim 1, wherein, In step (2), the third ML model is trained using the second training dataset to obtain the fourth ML model, wherein the third ML model has an architecture different from that of the second ML model.

4. The method according to claim 1, wherein, The second image data includes noise that is neither independent nor identically distributed.

5. The method according to claim 1, further comprising: The denoising and dealiasing ML model is applied to the patient image to obtain a denoised and / or dealiased patient image.

6. The method according to claim 5, wherein, The patient images were acquired using at least one of a low-field magnetic resonance imaging system (LMR imaging system) and a point-of-care (POC) MR imaging system.

7. The method according to claim 1, wherein, The first image data and the second image data belong to separate domains.

8. The method according to claim 1, further comprising: The training image data is enhanced based on enhancement processing prior to step (2).

9. The method according to claim 1, wherein, The trained second ML model includes multiple convolutional neural network layers, i.e., multiple CNN layers.

10. The method according to claim 1, further comprising: The first training dataset is generated by applying the raw imaging data to the image reconstruction pipeline.

11. The method of claim 10, further comprising: Simulated image corruption is added to the original imaging data.

12. The method according to claim 1, wherein, The third ML model is derived from the second ML model.

13. A method comprising: Patient images are acquired using an imaging system, and denoising and dealiasing machine learning models, i.e., denoising and dealiasing ML models, are applied to the patient images to obtain denoised and / or dealiased patient images. These denoising and dealiasing ML models are obtained through the following operations: (1) Train a first ML model using a first training dataset including the first image data to obtain a second ML model; and (2) Using a second training dataset to train (a) the second ML model or (b) the third ML model to obtain a fourth ML model, wherein the second training dataset includes (i) the first image data and (ii) training image data obtained by applying at least one of the second ML model and the third ML model to the second image data, wherein the denoising and dealiasing ML model is the fourth ML model or derived from the fourth ML model.

14. The method according to claim 13, wherein, The patient images were acquired using at least one of a low-field MR imaging system and a POC MR imaging system.

15. A system comprising: An imaging system configured to generate imaging data; as well as One or more processors are configured to cause the imaging system to generate patient images and apply denoising and dealiasing ML models to the patient images to generate denoised and / or dealiased patient images, the denoising and dealiasing ML models being obtained through the following means: The first ML model is obtained using a first training dataset that includes the first image data; as well as A second training dataset is used to obtain a second ML model, wherein the second training dataset includes (i) the first image data and (ii) training image data obtained by applying at least one of the first ML model and the third ML model to the second image data, and wherein the denoising and dealiasing ML model is the second ML model or is derived from the second ML model.

Images