Image Reconstruction Method and Device

Abstract

The present disclosure relates to systems, methods, and apparatuses associated with reconstructing MR images based on undersampled magnetic resonance (MR) data. The MR data includes 2D or 3D information and may include multiple contrasts and multiple coils. The MR images are reconstructed using deep learning (DL) methods, which can accelerate the scanning and / or image generation process. By reconstructing the MR images based on respective subsets of contrasts, coils, and / or readout segments and then combining the reconstructed MR images to obtain a desired multi-contrast result, the challenges imposed by large amounts of MR data and hardware limitations are overcome.

Description

Technical Field

[0001] This application relates to magnetic resonance image reconstruction. Background Art

[0002] Multi-contrast MRI images of anatomical structures (such as the human brain or heart) can provide useful information about the characteristics of the anatomical structures and are thus commonly used in clinical practice. However, the acquisition of high-resolution MRI images requires collecting and encoding a large amount of MRI data (e.g., k-space data), which often results in long scan times that may affect image quality and increased sensitivity to motion artifacts. To address these issues, various acceleration techniques can be employed during image acquisition, e.g., undersampling k-space data and reconstructing MRI images based on the undersampled data. However, conventional acceleration techniques (such as compression sensing (CS)-based methods) are typically iterative and time-consuming, making these techniques unsuitable for processing large amounts of data generated in a multi-dimensional setting (e.g., multiple contrasts, multiple coils, multiple slice images, etc.) within a clinically acceptable time frame.

[0003] Therefore, there is a great need for systems, methods, and apparatuses for reconstructing high-resolution MRI images based on multi-dimensional MRI data and doing so within the requirements of clinical practice and the limitations of current available hardware (e.g., GPU memory, processing speed, etc.). Summary of the Invention

[0004] Systems, methods, and apparatuses associated with reconstructing magnetic resonance (MR) images based on a collection of undersampled MR data (e.g., k-space data) are described herein. The collection of undersampled MR data can be associated with an anatomical structure (such as the human heart or brain) and can include data associated with multiple contrast settings, multiple coils, and readout directions. The reconstruction of the MR image can be performed using deep learning-based methods and / or by partitioning the MR data set into smaller parts or subsets. For example, one or more neural networks can be used to reconstruct a first MR image of the anatomical structure based on a first portion of the undersampled MR data, where the first portion of the undersampled MR data corresponds to a first subset of the multiple contrast settings, a first subset of the multiple coils, or a first segment in the readout direction. A second MR image of the anatomical structure can be reconstructed based on a second portion of the undersampled MR data using one or more neural networks, where the second portion of the undersampled MR data corresponds to a second subset of the multiple contrast settings, a second subset of the multiple coils, or a second segment in the readout direction. Then, the first MR image and the second MR image can be combined to obtain the desired MR image with multi-contrast characteristics or features.

[0005] In an example, the first MR image can be reconstructed independently of the second MR image (e.g., the two MR images can be reconstructed in parallel). In an example, the second MR image is reconstructed based on the first MR image (e.g., in a sequential manner) to utilize the information contained in the first MR image. The first and second portions of the undersampled MR data can be selected based on different criteria. For example, the first and second portions of the MR data can be associated with different contrast settings, or the same contrast setting but different coils, or the same contrast setting and coils but different segments in the readout direction.

[0006] In an example, one or more neural networks can include a cascaded convolutional neural network (CNN) that includes one or more data consistency layers. In an example, one or more neural networks can include multiple depthwise separable convolutional layers. In an example, one or more neural networks can each have a structure determined via neural architecture search. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] A more detailed understanding of the examples disclosed herein can be obtained from the following description given by way of example in conjunction with the accompanying drawings.

[0008] Figure 1 is a block diagram illustrating an example framework for reconstructing magnetic resonance (MR) images using a deep learning-based method according to one or more embodiments described herein.

[0009] Figure 2 is a block diagram illustrating an example of combining two separately reconstructed MR images to obtain a target MR image with desired characteristics according to one or more embodiments described herein.

[0010] Figure 3A 、 Figure 3B and Figure 3C is a block diagram illustrating an example manner of selecting a subset of MR data from a multi-dimensional MR dataset for the purpose of reconstructing an MR image according to one or more embodiments described herein.

[0011] Figure 4 is a block diagram illustrating an example of MR image reconstruction using an artificial neural network according to one or more embodiments described herein.

[0012] Figure 5 is a block diagram illustrating an example neural network that can be used to reconstruct an MR image based on undersampled MR data according to one or more embodiments described herein.

[0013] Figure 6 is a flowchart illustrating an example process for training a neural network to perform the image reconstruction task described herein.

[0014] Figure 7 is a block diagram illustrating example components of a device that can be used to perform the image reconstruction tasks described herein. Detailed Description

[0015] In the figures of the accompanying drawings, the present disclosure is illustrated by way of example and not limitation.

[0016] Figure 1 is a block diagram illustrating an example framework 100 for reconstructing magnetic resonance (MR) images (e.g., 102a-n shown in the figures) of an anatomical structure (e.g., a human heart, brain, etc.) based on a multi-dimensional MR data set 104 collected via a magnetic resonance imaging (MRI) device. As such, the multi-dimensional MR data set 104 can include k-space data that represents spatial frequency and / or temporal phase information of the scanned anatomical structure. It should be noted that when used herein, the term multi-dimensional can refer not only to positions and / or orientations in time and space (e.g., two-dimensional (2D), three-dimensional (3D), 2D / 3D plus time, etc.), but also to settings, elements, components, parameters, features, factors, etc. that make up a complete entity. For example, a multi-dimensional MR data set (such as Figure 1 the multi-dimensional MR data set 104 shown) can mean not only that the data set can include 2D or 3D MR data, but also that the data can contain multiple contrast settings, multiple coils, multiple segments along a readout direction or a phase-encoding direction, etc.

[0017] The multi-dimensional MR data set 104 can be collected from a single scan or multiple scans and can include multiple slice images obtained along a particular axis (e.g., a longitudinal axis or a transverse axis). Since the data set 104 can contain multiple dimensions (e.g., multiple contrasts, multiple coils, etc.), the amount of data can become too large to process within a reasonable time if fully sampled. Thus, in one or more embodiments described herein, the multi-dimensional MR data set 104 can correspond to undersampled MR data (e.g., undersampled k-space data) obtained using various subsampling techniques (e.g., for the purpose of accelerating scan and / or image reconstruction operations), and the MR images 102a-n can be reconstructed based on the undersampled MR data using a deep learning (DL) model learned by a neural network 106. Examples of subsampling techniques can include Cartesian sampling, radial sampling, spiral sampling, Poisson disk sampling, etc.

[0018] For some clinical applications, even undersampled MR data may be impractical to process given the limitations of current available hardware (e.g., processor speed, graphics processing unit (GPU) memory, etc.) and / or the requirements of the clinical application (e.g., shorter scan times, higher quality images, etc.). Thus, in one or more embodiments described herein, MR images 102a-n may be reconstructed based on corresponding portions or subsets of the multi-dimensional MR dataset 104. For example, MR image 102a may be reconstructed based on a first portion 108a of the multi-dimensional MR dataset 104 corresponding to a first subset of contrast settings, a first subset of coils, and / or a first segment in the readout direction, MR image 102b may be reconstructed based on a second portion 108b of the multi-dimensional MR dataset 104 corresponding to a second subset of contrast settings, a second subset of coils, and / or a second segment in the readout direction, and MR image 102n may be reconstructed based on an nth portion 108n of the multi-dimensional MR dataset 104 corresponding to an nth subset of contrast settings, an nth subset of coils, and / or an nth segment in the readout direction.

[0019] The portions (or subsets) of the multi-dimensional MR dataset 104 used to reconstruct MR images 102a-n may be selected based on combinations of dimensions described herein. For example, Figure 1 Each of the illustrated MR data portions 108a-108n may correspond to a respective contrast setting (e.g., a subset of one or more contrast settings), a respective coil (e.g., a subset of one or more coils), a respective segment along the readout direction (e.g., a subset of one or more segments), etc. As another example, MR data portions 108a-n may correspond to the same contrast setting (e.g., the same subset of contrast settings), but different coils (e.g., different subsets of coils) and / or different segments along the readout direction. As yet another example, MR data portions 108a-n may correspond to the same contrast setting (e.g., the same subset of contrast settings) and the same coil (e.g., the same subset of coils), but different segments along the readout direction. As yet another example, MR data portions 108a-n may correspond to the same contrast setting (e.g., the same subset of contrast settings) and the same segment in the readout direction (e.g., the same subset of segments), but different coils (e.g., different subsets of coils). Other dimension combinations are also contemplated herein. However, for ease of description, these combinations are not all listed individually herein.

[0020] Portions (or subsets) 108a-n of the multi-dimensional MR dataset 104 can be selected (e.g., identified) from the multi-dimensional MR dataset 104 based on the dimensions and / or segments of the multi-dimensional MR dataset 104 and / or various indicators (e.g., markers, tags, and / or other types of identifiers) that may be included in the multi-dimensional MR dataset 104. For example, portions of the MR data corresponding to a particular contrast setting may be marked in the multi-dimensional MR dataset 104 by a unique identifier corresponding to the contrast setting such that portions of the MR data can be selected from the multi-dimensional MR dataset 104 based on the unique identifier.

[0021] By reconstructing the MR images 102a-n based on corresponding portions (or subsets) of the multi-dimensional MR dataset 104 or by using portions (or subsets) of the large MR dataset to train a neural network, the framework described herein can allow for the processing of large amounts of data generated by multi-contrast, high-resolution MRI procedures while alleviating the constraints imposed by currently available hardware. For example, the GPU used to implement the neural network 106 may have a limited amount of memory and thus may not be able to accommodate the entire multi-dimensional MR dataset during training or testing / inference. By dividing the multi-dimensional MR dataset 104 into smaller portions that are subsets corresponding to contrast settings, coils, and / or readout segments, the MR dataset 104 can be processed using deep learning-based techniques regardless of the aforementioned hardware limitations. Since the amount of data to be processed is smaller, the reconstruction speed of individual MR images can also become faster. Individual MR images can be reconstructed in parallel (e.g., independently of each other) and in a sequential manner. In the latter case, the reconstruction of a second MR image (e.g., based on a second subset of the MR data) can utilize the features and / or characteristics of a first MR image reconstructed previously (e.g., using a first subset of the MR data).

[0022] The MR images 102a-n generated using the techniques described herein can be combined to derive an image (e.g., a 2D image, a 3D image, a 2D or 3D plus time image, etc.) having certain desired features, tissue contrast, intensity, etc. Figure 2 An example is shown where two images 202a and 202b reconstructed based on data collected from two coils respectively are combined into a target image 204 that contains information from both images 202a and 202b. Various techniques can be employed to combine individually reconstructed MR images and / or to determine diagnostic metrics based on individually reconstructed MR images and / or combined MR images. For example, by Figure 1Averaging all or a subset of the MR images 102a-n shown to generate a composite proton density weighted (PDW) image and / or a T1 weighted (T1W) image. As another example, a T2* map can be calculated by applying multi-dimensional integration (MDI) to the MR images 102a-n. As yet another example, one or more field maps can be extracted from the MR images 102a-n and used to generate quantitative susceptibility mapping (QSM) (e.g., using an L2 norm optimization method with dynamic linear artifact regularization).

[0023] Figures 3A to 3C Illustrates an example manner for selecting a portion (or subset) of a large MR data set to reconstruct an MR image. Figure 3A The example in shows that the MR data set can include data (e.g., k-space data) containing different contrast settings 302a-302c, and an MR image can be reconstructed based on a portion of the MR data set corresponding to one or more of the contrast settings 302a-302c. Similarly, Figure 3B The example in shows that the MR data set can include data (e.g., k-space data) collected along the readout direction R, and an MR image can be reconstructed based on a portion of the MR data set corresponding to one or more segments (e.g., 304a, 304b, and 304c) along the readout direction R. Finally, Figure 3C The example in shows that the MR data set can include data (e.g., k-space data) collected using multiple coils (e.g., 306a and 306b), and an MR image (e.g., I1, I2, etc.) can be reconstructed based on a portion of the MR data set corresponding to one or more of the coils (e.g., 306a and / or 306b). It should be noted that even though Figures 3A to 3C shows selecting a portion of the MR data based on only one of the contrast settings, readout segments, or coils, the drawings should not be construed as limiting the ways in which a portion of the MR data can be selected. Instead, as described above, a multi-dimensional MR data set can be divided into smaller portions or subsets in order to reconstruct corresponding MR images based on a dimensional combination including a combination of contrast settings, readout segments, and / or coils.

[0024] Figure 4 Illustrates an example of using a deep learning (DL) method to reconstruct an MR image based on undersampled, multi-contrast, multi-coil MR data (e.g., 2D or 3D MR data). As shown, the reconstruction can be performed using an artificial neural network (ANN) 402. The input 404 to the ANN can be a multi-dimensional undersampled MR data set, such as Figure 1 the multi-dimensional MRI data set 104 shown. The input 404 to the ANN can also be a portion or subset of the multi-dimensional undersampled MR data set, such as Figure 1The subsets of the MR data 108a, 108b,..., or 108n shown. As explained herein, the multi-dimensional undersampled MR data sets can include data collected across multiple contrast settings, multiple coils, etc. Thus, when subsets of the MR data sets are used for reconstruction, subsets of the MR data can be selected from the multi-dimensional undersampled MR data sets based on one or a combination of contrast settings, coils, etc. In either case (e.g., whether the entire MR data set or a portion of the MR data set is provided as the input 404 to the ANN at inference time (e.g., once the ANN is trained and made online)), a portion or subset of the multi-dimensional MRI data set can be used to train the ANN 402 to learn a model for reconstructing MR images based on the multi-dimensional MR data. For example, during training, the multi-dimensional (e.g., multi-contrast, multi-coil) MRI training data set can be divided into smaller portions, each smaller portion corresponding to a specific contrast setting, coil, and / or segment along the readout direction (e.g., the segmentation can be performed randomly). Then, the smaller portions of the MR data can be provided to the ANN 402 to reconstruct MR images corresponding to the specific contrast settings, coils, and / or segments (e.g., to avoid loading the entire data set into the GPU memory). In this way, the constraints imposed by hardware limitations (such as limited GPU memory) can be overcome, and once the ANN 402 is appropriately trained (e.g., offline), only a forward pass can be employed to generate reconstructed MR images based on the undersampled MR data (e.g., the entire multi-dimensional data set or a portion thereof), resulting in a faster processing speed compared to conventional image reconstruction methods (e.g., CS-based methods).

[0025] It should be noted that even though the input 404 to the neural network 402 is shown in Figure 4 as including an undersampled MR data set (e.g., k-space data), the input 404 can include, for example, an MR image derived from the undersampled MR data set by applying an inverse Fourier transform to the MR data set.

[0026] The training of the ANN 402 can be formulated as learning a function f nn based on a large data set that maps undersampled (e.g., zero-filled) MR data (e.g., k-space measurements) to one or more fully sampled MR images by minimizing a loss function, as exemplified below:

[0027] f nn : x z → y min θ (L(f nn (x z (θ), y))

[0028] where y = f nn (x z|θ) can represent the MR image reconstructed by the ANN 402 using the parameters θ (e.g., the weights of the ANN 402) during forward propagation, y can represent the gold standard image, and x z can represent undersampled MR data (e.g., k-space data), and L can represent the loss function. The ANN 402 can utilize various network architectures. For example, the ANN 402 can be implemented as a 2D convolutional neural network (CNN), 3D CNN, cascaded CNN, recurrent neural network (RNN), generative adversarial network (GAN), and / or combinations thereof. In an example, the specific structure of the ANN 402 can be determined by performing a neural architecture search (NAS) that aims to learn the network topology that can achieve the best performance (e.g., in terms of computational efficiency) in the image reconstruction tasks described herein. Such a search can be performed using at least the following modules or components: a search space, a search algorithm (or optimization method), and an evaluation strategy. The search space can define the types of ANNs that can be designed and optimized, the search strategy can specify how to explore the search space, and the evaluation strategy can be used to evaluate the performance of candidate ANNs. Various methods can be utilized to sample the search space and find the architecture that yields the best performance. These methods can include, for example, random search, reinforcement learning, gradient descent, etc.

[0029] In one or more suitable architectures of the ANN 402, the network can include multiple convolutional layers, one or more pooling layers, and / or one or more fully connected layers. In an example, each convolutional layer can include multiple convolutional kernels or filters that are configured to extract specific features from the input image through one or more convolutional operations (e.g., the input image can be obtained by applying an inverse Fourier transform to the corresponding MR data set such as x in the above equation z ). In an example, the convolutional layer can include one or more depthwise separable convolutional layers (e.g., 3D depthwise separable convolutional layers) that are configured to perform a convolutional operation on an input that includes multiple individual channels corresponding to different interpretations or reproductions of the input. For example, the real and imaginary parts of the input complex values can be transformed into two separate channels and provided to the network. Then, each input channel can be convolved with a corresponding filter, and the convolved outputs can be stacked together to derive the target output.

[0030] The convolution operation described herein (e.g., using regular or depthwise convolutional layers) may be followed by batch normalization and / or linear or non-linear activation, and the features extracted by the convolutional layer may be downsampled by one or more pooling layers (e.g., using a 2×2 window and a stride of 2) to reduce the redundancy and / or size of the features (e.g., by a factor of 2). As a result of the convolution and / or downsampling operations, a corresponding feature representation of the input image may be obtained, for example, in the form of one or more feature maps or feature vectors.

[0031] The ANN 402 may also include a plurality of transposed convolutional layers and / or one or more unpooling layers. Through these layers, the ANN 402 may perform a series of upsampling and / or transposed convolution operations based on the feature maps or feature vectors generated by the above-mentioned downsampling operations. For example, the ANN 402 may upsample the feature representation based on the pooling indices stored during the downsampling phase (e.g., using a 3×3 transposed convolutional kernel with a stride of 2) to restore the features extracted from the input image to a size or resolution corresponding to a fully sampled MR dataset or image (e.g., the MR dataset may be transformed into an MR image by Fourier transform and vice versa).

[0032] Various loss functions may be employed to facilitate the training of the ANN 402. Such loss functions may be based on, for example, mean squared error (MSE), L1 norm, L2 norm, structural similarity index measure (SSIM) loss, adversarial loss, etc. Training data may be acquired from an actual MRI procedure with parallel imaging (e.g., a 3D multi-contrast, multi-coil procedure) to obtain a gold standard image. For example, a 3-fold and / or 5-fold Poisson disk undersampling scheme may be used to undersample the data in the phase encoding and / or slice image directions.

[0033] Figure 5An example cascaded network structure that can be used to implement the ANN 402 is illustrated. As shown, the network can include a plurality of blocks (e.g., sub-networks) 502a-c, each block including one or more (e.g., five) convolutional layers (e.g., 2D or 3D convolutional layers), one or more pooling layers, and / or one or more batch normalization layers. One or more (e.g., each) of the convolutional layers can be associated with a corresponding activation function (e.g., rectified linear unit (ReLU)), and each of the blocks or sub-networks 502b-c located after the first block or sub-network 502a in the cascaded structure can learn to reconstruct an MR image based on the output of the previous block or sub-network. The blocks or sub-networks 502a-c can be configured to perform residual learning (e.g., the network layers can be trained to learn a residual mapping with reference to the layer input). Thus, each of the blocks or sub-networks 502a-c can include a corresponding residual connection 504a-c configured to sum the output of the block or sub-network with its input. All the blocks or sub-networks 502a-c can be jointly trained as a large network in an end-to-end manner.

[0034] One or more (e.g., all) of the blocks or sub-networks can also include corresponding data consistency layers 506a-c (e.g., data consistency functions or modules), which are configured to ensure that the values predicted by the sub-network or block in the image domain are consistent with the acquired k-space samples. To this end, the operations performed by the data consistency layers 506a-c can include transforming the image reconstructed by the sub-network or network block (e.g., via Fourier transform) into k-space data, and performing a comparison (e.g., element-wise comparison) of the network predicted values with the gold-standard k-space samples.

[0035] Figure 6 An example process 600 for training a neural network (e.g., Figure 4 the ANN 402 shown) to perform the image reconstruction operations described herein is illustrated. The training can be performed using data collected from an actual MRI process with parallel imaging (e.g., a 3D multi-contrast, multi-coil MRI process) to obtain a gold-standard image. Then, the training data can be retrospectively undersampled, for example, using a 3-fold and / or 5-fold Poisson disk undersampling scheme in the phase encoding and / or slice image directions to obtain an undersampled training data set that includes multiple contrasts and multiple coils. The undersampled training data set can be further divided into smaller parts (e.g., subsets), each smaller part corresponding to a specific set of one or more contrast settings, one or more coils, and / or one or more segments along the readout direction (e.g., the readout can be randomly chopped into 128 segments for training purposes). Then, the smaller parts or subsets of the training data set can train the neural network to learn a model for reconstructing an MR image based on the undersampled data (e.g., k-space data).

[0036] Process 600 may begin at 602, and at 604, initial parameters of a neural network (e.g., weights associated with various filters or kernels of the neural network) may be initialized. The parameters may be initialized, for example, based on samples collected from one or more probability distributions or parameter values of another neural network having a similar architecture. At 606, the neural network may receive an input MR image associated with a portion or subset of the MR data described herein (e.g., the input MR image may be generated by applying an inverse Fourier transform to the input MR data), and reconstruct an output image through the various layers of the neural network. At 608, the reconstructed image may be compared with a gold standard image to determine the adjustments needed to the currently assigned neural network parameters. The adjustments may be determined based on a loss function (e.g., MSE, L1, L2, etc.) and gradient descent associated with the loss function (e.g., stochastic gradient descent).

[0037] At 610, the neural network may apply the adjustments to the currently assigned network parameters, for example, through a backpropagation process. At 612, the neural network may determine whether one or more training termination criteria are met. For example, if the neural network has completed a predetermined number of training iterations, if the difference between the prediction result and the gold standard value is below a predetermined threshold, or if the change in the value of the loss function between two training iterations is below a predetermined threshold, then the neural network may determine that the training termination criteria are met. If it is determined at 612 that the training termination criteria are not met, the neural network may return to 606. If it is determined at 612 that the training termination criteria are met, the neural network may end the training process 600 at 614.

[0038] For simplicity of illustration, the training steps are depicted and described herein in a particular order. However, it should be understood that the training operations may occur in various orders, simultaneously, and / or in conjunction with other operations not presented or described herein. Additionally, it should be noted that not all operations that may be included in the training process are depicted and described herein, and not all of the illustrated operations need to be performed.

[0039] The systems, methods, and / or apparatuses described herein may be implemented using one or more processors, one or more storage devices, and / or other suitable auxiliary devices (such as display devices, communication devices, input / output devices, etc.). Figure 7FIG. 0 is a block diagram illustrating an example device 700 that may be configured to perform the image reconstruction operations described herein. As shown, device 700 may include a processor (e.g., one or more processors) 702, which may be a central processing unit (CPU), a graphics processing unit (GPU), a microcontroller, a reduced instruction set computer (RISC) processor, an application specific integrated circuit (ASIC), an application specific instruction set processor (ASIP), a physics processing unit (PPU), a digital signal processor (DSP), a field programmable gate array (FPGA), or any other circuit or processor capable of performing the functions described herein. Device 700 may also include a communication circuit 704, a memory 706, a mass storage device 708, an input device 710, and / or a communication link 712 (e.g., a communication bus) through which one or more of the components shown in the figure may exchange information.

[0040] The communication circuit 704 may be configured to send and receive information using one or more communication protocols (e.g., TCP / IP) and one or more communication networks, including local area networks (LANs), wide area networks (WANs), the Internet, wireless data networks (e.g., Wi-Fi, 3G, 4G / LTE, or 5G networks). The memory 706 may include a storage medium (e.g., a non-transitory storage medium) configured to store machine-readable instructions that, when executed, cause the processor 702 to perform one or more of the functions described herein. Examples of machine-readable media may include volatile or non-volatile memory, including but not limited to semiconductor memories (e.g., electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory, etc.). The mass storage device 708 may include one or more disks, such as one or more internal hard disks, one or more removable disks, one or more magneto-optical disks, one or more CD-ROM or DVD-ROM disks, etc., on which instructions and / or data may be stored to facilitate operation of the processor 702. The input device 710 may include a keyboard, a mouse, a voice control input device, a touch-sensitive input device (e.g., a touch screen), etc., for receiving user input to the device 700.

[0041] It should be noted that device 700 may operate as a stand-alone device or may be connected (e.g., networked or clustered) with other computing devices to perform the functions described herein. And even though only one instance of each component is shown in Figure 7 FIG. 0, those skilled in the art will understand that device 700 may include multiple instances of one or more of the components shown in the figure.

[0042] Although the present disclosure has been described in terms of certain embodiments and generally associated methods, variations and modifications of the embodiments and methods will be apparent to those skilled in the art. Accordingly, the foregoing description of the exemplary embodiments does not limit the present disclosure. Other changes, substitutions, and alterations are also possible without departing from the spirit and scope of the present disclosure. Additionally, unless otherwise specifically stated, discussions using terms such as "analyze", "determine", "enable", "identify", "modify", etc., refer to actions and processes of a computer system or similar electronic computing device that manipulate and transform data represented as physical (e.g., electronic) quantities within the registers and memories of the computer system into other data represented as physical quantities within the memories or other such information storage, transmission, or display devices of the computer system.

[0043] It should be understood that the foregoing description is intended to be illustrative, not restrictive. After reading and understanding the above description, many other embodiments will be apparent to those skilled in the art. Accordingly, the scope of the present disclosure should be determined with reference to the appended claims and the full scope of equivalents to which such claims are entitled.

Claims

1. A method for reconstructing magnetic resonance (MR) images, the method comprising: Obtaining a set of undersampled MR data associated with an anatomical structure, wherein the set of undersampled MR data includes data associated with multiple contrast settings, multiple coils, and readout directions; Selecting a first portion of the undersampled MR data, the first portion corresponding to a first subset of the multiple contrast settings, a first subset of the multiple coils, or a first segment in the readout direction; Using one or more neural networks to reconstruct a first MR image of the anatomical structure based on the first portion of the undersampled MR data; Selecting a second portion of the undersampled MR data, the second portion corresponding to a second subset of the multiple contrast settings, a second subset of the multiple coils, or a second segment in the readout direction; and Using the one or more neural networks to reconstruct a second MR image of the anatomical structure based on the second portion of the undersampled MR data; Generating a target MR image of the anatomical structure by combining at least the first MR image and the second MR image; Wherein the first portion and the second portion of the undersampled MR data correspond to the same contrast setting, and correspond to different coils and / or different segments in the readout direction.

2. The method according to claim 1, wherein The first MR image is reconstructed independently of the second MR image, or, the second MR image is reconstructed based on the first MR image.

3. The method according to claim 1, wherein The one or more neural networks include a cascaded convolutional neural network (CNN), the CNN including one or more data consistency layers, or, the one or more neural networks include multiple depthwise separable convolutional layers.

4. The method according to claim 1, wherein The set of undersampled MR data includes two-dimensional or three-dimensional MR data.

5. A device, comprising: One or more processors, configured to: Obtain a set of undersampled MR data associated with an anatomical structure, wherein the set of undersampled MR data includes data associated with multiple contrast settings, multiple coils, and readout directions; Select a first portion of the undersampled MR data, the first portion corresponding to a first subset of the multiple contrast settings, a first subset of the multiple coils, or a first segment in the readout direction; Use one or more neural networks to reconstruct a first MR image of the anatomical structure based on the first portion of the undersampled MR data; Select a second portion of the undersampled MR data, the second portion corresponding to a second subset of the multiple contrast settings, a second subset of the multiple coils, or a second segment in the readout direction; and Use the one or more neural networks to reconstruct a second MR image of the anatomical structure based on the second portion of the undersampled MR data; Generate a target MR image of the anatomical structure by combining at least the first MR image and the second MR image; Wherein the first portion and the second portion of the undersampled MR data correspond to the same contrast setting, and correspond to different coils and / or different segments in the readout direction.

6. The device according to claim 5, wherein Each of the one or more neural networks has a structure determined via neural architecture search.

7. A method of training a neural network to learn a model for reconstructing magnetic resonance (MR) images, the method comprising: The neural network receives a first portion of an MR training dataset, wherein the MR training dataset includes undersampled MR data associated with an anatomical structure, a plurality of contrast settings, a plurality of coils, and a readout direction, and wherein the first portion of the training dataset corresponds to a first subset of the plurality of contrast settings, a first subset of the plurality of coils, or a first segment in the readout direction; The neural network reconstructs a first MR image of the anatomical structure based on the first portion of the training dataset; The neural network adjusts one or more execution parameters based on a difference between the first MR image and a first gold standard image; The neural network receives a second portion of the training dataset, wherein the second portion of the training dataset corresponds to a second subset of the plurality of contrast settings, a second subset of the plurality of coils, or a second segment in the readout direction; The neural network reconstructs a second MR image of the anatomical structure based on the second portion of the training dataset; and The neural network further adjusts the one or more execution parameters based on a difference between the second MR image and a second gold standard image; wherein the first portion and the second portion of the undersampled MR data correspond to the same contrast setting, and to different coils and / or different segments in the readout direction.

Images