Method and apparatus for reconstructing magnetic resonance imaging (MRI) images

Abstract

Systems, methods, and apparatuses associated with reconstructing magnetic resonance (MR) images based on multi-slice imaging undersampled MRI data (e.g., k-space data) are disclosed herein. Multi-slice imaging MRI data can be acquired using simultaneous multi-slice (SMS) techniques, and MRI information associated with multiple MRI slice images can be entangled in the multi-slice imaging MRI data. A neural network can be trained and used to untangle the MRI information and reconstruct MRI images for different slice images. A data consistency component can be used to estimate k-space data based on estimates made by the neural network, from which respective MRI images associated with the multiple MRI slice images can be obtained by applying a Fourier transform to the k-space data.

Description

Technical Field

[0001] This application relates to the field of magnetic resonance imaging. Background Technology

[0002] Magnetic resonance imaging (MRI) techniques are used to collect data in a spatial frequency space (e.g., often referred to as k-space), and images generated based on the collected data can provide insights into the characteristics of anatomical structures that are important for clinical research and diagnosis. The collection of k-space data can be a slow process, thus undersampling can be applied to accelerate the operation. The undersampled k-space data can then be reconstructed (e.g., reconstructed into MRI images) to obtain results with similar quality to a fully sampled dataset (e.g., a fully sampled MRI image). Conventional MRI acceleration techniques (such as compressed sensing (CS) and parallel imaging (PI)) may require good coil configurations (e.g., multiple coils with differentiated power) and / or iterative computations, making them unsuitable or unsatisfactory for certain use cases (such as those involving the collection of multi-slice imaging data, e.g., simultaneous multi-slice imaging data collection). Therefore, systems, methods, and apparatuses for reconstructing undersampled MRI information (e.g., k-space data and / or MR images) in a manner that meets the practical requirements and constraints of clinical practice are highly desirable. Summary of the Invention

[0003] This document describes systems, methods, and apparatuses associated with reconstructing magnetic resonance imaging (MRI) images based on simultaneous multi-slice imaging (e.g., two or more) datasets including undersampled MRI data (e.g., MRI images or k-space data). Such SMS datasets can include multiple MRI slice images acquired (e.g., excited) simultaneously during an MRI scan. For example, an SMS dataset can include first undersampled MRI data associated with a first MRI slice image, second undersampled MRI data associated with a second MRI slice image, and so on. According to one or more embodiments described herein, an artificial neural network (ANN) can be trained and used to acquire (e.g., receive) the SMS dataset and generate a first reconstructed MRI image corresponding to the first MRI slice image and a second reconstructed MRI image corresponding to the second MRI slice image. The ANN can be trained to perform these tasks through a training process that can include: processing first undersampled MRI training data of the SMS training dataset through instances of the ANN to obtain a first estimated MRI image; and processing second undersampled MRI training data of the SMS training dataset through instances of the ANN to obtain a second estimated MRI image, wherein the first and second undersampled MRI training data can correspond to the first and second MRI slice images of the SMS training dataset, respectively. Training may also include: determining a combined training loss (e.g., average loss, triplet loss, etc.) by jointly considering a first training loss associated with a first estimated MRI image and a second training loss associated with a second estimated MRI image; and adjusting the parameters of the ANN instances based on gradient descent using the combined training loss.

[0004] In the example, the first undersampled MRI data may include a first artifact associated with the second undersampled MRI data, and the second undersampled MRI data may include a second artifact associated with the first undersampled MRI data, but the first reconstructed MRI image and the second reconstructed MRI image may be substantially free of the first and second artifacts, respectively. The first and second reconstructed MRI images may also have a quality (e.g., resolution) substantially similar to that of a fully sampled MRI image. In the example, the ANN described herein may include a first subnetwork and a second subnetwork sharing substantially similar structures and execution parameters. The first subnetwork may be configured to process the first undersampled MRI data, and the second subnetwork may be configured to process the second undersampled MRI data. During training of the ANN, the first subnetwork (e.g., for an instance of the ANN used for training) may be configured to process the first undersampled MRI training data, and the second subnetwork (e.g., for an instance of the ANN used for training) may be configured to process the second undersampled MRI training data, and mirror updates may be applied to the corresponding parameters of the first and second subnetworks based on the training loss described herein.

[0005] In the example, the ANN described herein may further include a data consistency (DC) component configured to estimate k-space data based on a first intermediate image generated by the ANN using first undersampled MRI data and a second intermediate image generated by the ANN using second undersampled MRI data. The first and second reconstructed MRI images can then be generated by applying an inverse Fourier transform (e.g., a 3D Fast Fourier Transform (FFT)) to the estimated k-space data. In the example, at least a portion of the estimated k-space data may be replaced with a corresponding portion of the SMS dataset before applying the inverse Fourier transform to the estimated k-space data.

[0006] In the examples, the first undersampled MRI data included in the SMS dataset may include MRI data acquired using one or more coils from a first set. The second undersampled MRI data included in the SMS dataset may include MRI data acquired using one or more coils from a second set. In these examples, corresponding coil sensitivity maps associated with the first and second sets of coils can be determined and used to estimate the k-space data described above. Attached Figure Description

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

[0008] Figure 1 This is a block diagram illustrating an example system for processing undersampled Simultaneous Multi-Slice Imaging (SMS) datasets.

[0009] Figure 2 This is a block diagram illustrating example operations that can be associated with multi-slice imaging MRI data processing.

[0010] Figure 3 This is a block diagram illustrating an example implementation of a multi-slice imaging MRI data processing system according to one or more embodiments described herein.

[0011] Figure 4 This is a flowchart illustrating example operations that can be performed by a multi-slice imaging MRI data processing system according to one or more embodiments described herein.

[0012] Figure 5 This is a flowchart illustrating an example process for training a neural network to perform the image reconstruction task described herein.

[0013] Figure 6 This is a block diagram illustrating example components of a device that can be used to perform the image reconstruction tasks described herein. Detailed Implementation

[0014] The present disclosure is illustrated by way of example rather than limitation in the figures.

[0015] Figure 1 This is a block diagram illustrating an example system 100 for processing a simultaneous multi-slice imaging (SMS) dataset 102 collected by a magnetic resonance imaging (MRI) apparatus (e.g., an MRI scanner). The SMS dataset 102 may include undersampled MRI information associated with multiple MRI slice images of an anatomical structure (e.g., two or more MRI slice images including a first MRI slice image and a second MRI slice image). The anatomical structure may be, for example, a human heart or a human brain, and the multiple MRI slice images of the anatomical structure may be acquired simultaneously by the MRI apparatus during the scanning process (e.g., using simultaneous multi-slice imaging (SMS) excitation techniques). The SMS dataset 102 may include different types of data or information. For example, the SMS dataset 102 may include undersampled k-space data (e.g., raw MRI data) indicating the frequency, phase, and / or intensity of a signal captured by the MRI apparatus during the scanning process. Such undersampled k-space data may be characterized by Cartesian or non-Cartesian trajectories and / or may be collected using uniform, random, or pseudo-random undersampled techniques. The SMS dataset 102 may also include image data (e.g., one or more MRI images) that visually depict the anatomical structure based on the k-space data collected by the MRI apparatus. These images can include a single static image or multiple dynamic images (e.g., multi-contrast images), which can be derived, for example, by applying a Fourier transform (e.g., inverse fast Fourier transform (IFFT)) to the collected k-space data. In the example, undersampled MRI information associated with multiple MRI slice images may be entangled in SMS dataset 102 (e.g., multiple slice images may overlap in the images included in SMS dataset 102, and one slice image may include artifacts associated with another slice image, etc.). Furthermore, since multiple slice images may be spatially separated, discontinuities may also exist in the MRI slice image information included in SMS dataset 102.

[0016] Figure 1The illustrated system 100 may include an artificial neural network (ANN) 104 configured (e.g., trained) to remove noise (e.g., artifacts) and / or unravel multiple slice images included in the SMS dataset 102. For example, the ANN 104 may be pre-trained to denoise the SMS dataset 102 (e.g., by learning the similarity and / or dissimilarity of multiple slice images included in the dataset), reconstruct denoised multi-slice imaging MRI data in k-space, and derive corresponding reconstructed MRI data 106a, 106b (e.g., MRI images) corresponding to the multiple slice images. Each of the reconstructed MRI data 106a and 106b may be substantially free of artifacts introduced by other MRI slice images, thus being able to unravel from other MRI slice images. The reconstructed MRI data 106a and 106b may also have a quality (e.g., resolution) similar to that of fully sampled MRI data (e.g., fully sampled MRI images). As will be described in more detail below, ANN 104 may include components associated with reducing noise levels in the reconstructed MRI data, components associated with compensating for estimation / prediction bias, and components for incorporating other acceleration techniques into the unraveling / reconstruction process.

[0017] Figure 2 Examples are illustrated by multi-slice imaging MRI data processing systems (e.g., according to one or more embodiments described herein) Figure 1 The system 100 performs an example operation. As shown, the data processing system can acquire (e.g., receive) an MRI dataset that includes undersampled MRI data associated with multiple MRI slice images (e.g., simultaneously acquired MRI data Kz=0 and MRI data Kz=1, which may correspond to corresponding phase codes along the slice image orientation). In the simultaneously acquired multi-slice imaging MRI dataset, the corresponding data associated with multiple slice images may be entangled (e.g., as shown in 1), and the data processing system described herein can be configured to use an artificial neural network (e.g., pre-trained to learn a multi-slice imaging MRI data processing model) to untangle the data associated with the multiple slice images (e.g., at 2) and reconstruct the data (e.g., at 3) to a higher quality (e.g., similar to the quality of a fully sampled MRI dataset).

[0018] Figure 3 This shows a configuration for processing SMS dataset 302 (e.g., Figure 1 A multi-slice imaging MRI data processing system 300 (e.g., SMS dataset 102) Figure 1An example implementation of System 100. As described herein, SMS dataset 302 may include MRI data (e.g., undersampled MRI data) associated with multiple MRI slice images (e.g., two or more slice images). For example, multiple MRI slice images may be acquired using SMS excitation techniques, thus MRI data associated with each of the multiple MRI slice images may be entangled in MRI dataset 302. System 300 may include artificial neural network 304 (e.g., Figure 1 The ANN (104) can be pre-trained to receive the SMS dataset 302 and reconstruct the corresponding MRI data (e.g., MRI images) of multiple MRI slice images based on the SMS dataset 302. The reconstructed MRI data of the slice images can be decomposed (e.g., slice image 1 is 306a, slice image 2 is 306b, etc.) and can have a quality similar to that of a fully sampled MRI image.

[0019] In the example, ANN 304 can include multiple (e.g., two or more) subnetworks (e.g., Figure 3 The subnetworks 308a and 308b shown have the same or substantially similar structure (e.g., in terms of the number of layers, the type of layers, the number of feature maps or vectors generated by the respective networks, etc.) and / or the same or substantially similar execution parameters (e.g., the weights associated with the kernels or filters of the respective networks). Each of the multiple subnetworks can be trained to process corresponding MRI slice images included in the SMS dataset 302, and the multiple subnetworks together are able to learn (e.g., identify) the similarity and / or dissimilarity of different MRI slice images included in the SMS dataset 302, and denoise the SMS dataset 302 (e.g., remove artifacts from it) based on the learned (e.g., identified) similarity and / or dissimilarity. Figure 3 The examples illustrate that subnetworks (e.g., 308a and 308b) can be configured to form a conjoined neural network; however, those skilled in the art will understand that more than two subnetworks can be used to process the SMS dataset 302 (e.g., N subnetworks can be used to process N slice images, where N can be greater than two), and the subnetworks can share similar structures and / or execution parameters. Those skilled in the art will also understand that in some implementations, ANN 304 may include only one network (e.g., instead of the multiple subnetworks described herein), and the ANN can be trained to sequentially process multiple MRI slice images included in the SMS dataset 302, learn (e.g., identify) the similarity and / or dissimilarity of the multiple MRI slice images through processing, and denoise (e.g., remove artifacts) the individual MRI slice images of the SMS dataset 302 based on the learned (e.g., identified) similarity and / or dissimilarity.

[0020] In the example, each or a single network among the multiple sub-networks described herein may include a convolutional neural network (CNN) configured to receive a corresponding input MRI image and extract features from the input image. Such an input MRI image may be part of a multi-slice imaging MRI dataset 302, or the input MRI image may be obtained by applying an IFFT to k-space data included in the multi-slice imaging MRI dataset 302 (e.g., when the multi-slice imaging MRI dataset 302 includes raw MRI data instead of MRI images). The CNN may include multiple layers, such as one or more convolutional layers, one or more pooling layers, and / or one or more fully connected layers. Each convolutional layer may include multiple convolutional kernels or filters configured to extract specific features from the input MRI image. Following the convolutional operation may be batch normalization and / or non-linear activation, and the features extracted by the convolutional layers (e.g., in the form of twin feature maps or feature vectors) may be downsampled (e.g., using a 2×2 window and a stride of 2) by pooling layers and / or fully connected layers to reduce feature redundancy and / or size (e.g., reduced by a factor of 2). In the example (e.g., when the input includes dynamic images), a recurrent convolutional neural network structure can be used to extract certain hidden states of the input using one or more convolutional layers and to pass these hidden states through different image frames.

[0021] The corresponding feature maps or vectors generated by the CNN can be used to determine the similarity (or dissimilarity) between different slice images of the SMS dataset 302, and noise (e.g., artifacts caused by other slice images) can be removed from the individual MRI slice images based on the determined similarity or dissimilarity. For example, for any one of multiple MRI slice images, noise (e.g., artifacts) may correspond to MRI information introduced by one or more other MRI slice images, and since the slice images may be acquired simultaneously, this influence may be mutual between MRI slice images. Thus, if a first MRI slice image influences a second MRI slice image at a certain spatial frequency, the first MRI slice image may also be influenced by the second MRI slice image at the same spatial frequency. Therefore, the similarity between multiple MRI slice images can indicate noise or artifact patterns, and by learning these similarities, the CNN is able to remove noise from the individual MRI slice images.

[0022] In the example, the CNN associated with each of the subnetworks 308a, 308b may also include one or more non-pooling layers and one or more transposed convolutional layers. Through the non-pooling layers, the CNN can upsample the previously extracted features and further process the upsampled feature representation through one or more transposed convolutional operations (e.g., deconvolutional operations), followed by one or more batch normalization operations, to derive one or more dense feature maps (e.g., which may be scaled up by a factor of 2). These dense feature maps can then be used to generate MRI data of quality with a fully sampled MRI dataset.

[0023] In the example, ANN 304 may also include a data consistency (DC) checker 310 (e.g., as a layer of ANN 304) configured to check and / or improve the fidelity of the MRI data predicted by ANN 304. For example, DC checker 310 may be configured to receive MRI data generated by the conjoined network (e.g., denoised first and second intermediate MRI images predicted by subnetworks 308a and 308b based on the input), process the data (e.g., the first and second intermediate MRI images) to derive the corresponding MRI (e.g., k-space data), and obtain corresponding MRI images (e.g., unwrapped MRI images 306a and 306b) corresponding to multiple slice images of SMS dataset 302 based on the derived MRI data (e.g., by applying an inverse Fourier transform, such as IFFT, to the derived MRI data).

[0024] Various techniques can be used to derive MRI data based on estimates generated by subnetworks 308a and 308b. As an example, the desired MRI data, denoted as s(k) (e.g., the desired SMS dataset), can be determined based on the following equation:

[0025] s(k) = f(s1(k),s2(k)) 1), where k can represent a readout in k-space, s1(k) and s2(k) can represent two MRI slice images, and f can represent a function used to combine the two MRI slice images to obtain s(k). The function f can take the following form:

[0026] f(a,b)=u*a+v*b 2),

[0027] The values ​​of u and v can be manipulated (e.g., adjusted) to simulate the modulation applied to MRI slice images a and b. For example, for all even-numbered readouts (e.g., where k is even), the values ​​of u and v can be set to u = 1 and v = 1, such that s(k) = s1(k) + s2(k), and for all odd-numbered readouts (e.g., where k is odd), the values ​​of u and v can be set to u = 1 and v = -1, such that s(k) = s1(k) - s2(k). On the other hand, if the MRI images I1(x) and I2(x) (e.g., intermediate MRI images) predicted by subnetworks 308a and 308b are stacked together along the z-direction, the corresponding image space can be represented by I(x,z) = [I1(x), I2(x)]. Applying a Fourier transform (e.g., a fast Fourier transform) along the z-direction, the following equation can be derived (e.g., based on the properties of the Fourier transform):

[0028] J(x,kz)=FFTz(I(x,z))=[I1(x)+I2(x),I1(x)-I2(x)] 3),

[0029] Where J(x,kz=0) can be equal to I1(x)+I2(x), and J(x,kz=1) can be equal to I1(x)-I2(x). Comparing this to equations 1) and 2) shown above, s(k)=s1(k)+s2(k) can correspond to kz=0, s(k)=s1(k)-s2(k) can correspond to kz=1, and the predictions (e.g., MRI images) made by subnetworks 308a and 308b can be converted into MRI data based on, for example, s(k,kz)=[s1(k)+s2(k),s1(k)-s2(k)]. Using the MRI data thus obtained, corresponding reconstructed MRI images (e.g., 306a and 306b) corresponding to multiple slice images can be obtained, for example by applying IFFT to the MRI data, as illustrated in equation 4) below:

[0030] iFFT_k_kz(s(k,kz))=I(x,z)=[I1(x),I2(x)] 4)

[0031] In the example, DC checker 310 can be configured to compare denoised MRI data (e.g., obtained using the techniques described herein) with actually acquired MRI data (e.g., represented by SMS dataset 302) to ensure data consistency or fidelity. For example, DC checker 310 can be configured to compensate for estimation bias introduced by ANN 304 (e.g., in the case of actually acquired data) by replacing one or more portions of the estimated MRI data with corresponding portions of SMS dataset 302. This fidelity-compensated MRI data can then be transformed to image space (e.g., via IFFT) to obtain fidelity-compensated MRI images for multiple slice images.

[0032] The technique described in this paper outperforms other deep learning-based techniques used for processing SMS data. One reason may be that in SMS datasets, the acquired slice images can be spatially separated by a greater distance (e.g., to avoid slice image interference), where other deep learning-based techniques (such as those utilizing 3D convolutional kernels) may fail to produce satisfactory results. Conversely, since correlations may still exist between slice images (e.g., slice images may all be associated with the same physiological process such as heart contraction), the neural network architecture described in this paper can be adapted to learn the similarities or dissimilarity between slice images and remove artifacts from SMS based on the determined similarities or dissimilarity.

[0033] It should be noted that although examples are described in this document in the context of two MRI slice images, two MRI datasets, or two subnetworks, those skilled in the art will understand that the techniques disclosed herein can also be applied to more than two (e.g., three or more) MRI slice images, more than two MRI datasets, or more than two subnetworks.

[0034] In the example, the reconstruction results obtained using ANN 304 can be further improved by utilizing the features of the coils used to acquire the multi-slice imaging MRI dataset 302. These features may include, for example, the correlation of multiple coils, which may be indicated by one or more coil sensitivity maps associated with the coils. These coil sensitivity maps may be estimated, for example, from a reference scan or based on one or more calibration regions of the multi-slice imaging MRI dataset 302. The coil sensitivity maps may include, for example, a first coil sensitivity map associated with a first group of one or more coils used to acquire information associated with a first MRI slice image and a second coil sensitivity map associated with a second group of one or more coils used to acquire information associated with a second MRI slice image (e.g., the first group of coils and the second group of coils may overlap (e.g., include the same coils)). Once obtained, the coil sensitivity maps associated with the coils can be applied together with a Fourier transform (e.g., via ANN 304 and / or DC inspector 310) to reconstruct the multi-slice imaging MRI data. For example, MRI data (e.g., MRI images) associated with multiple coils can be multiplied by the corresponding complex conjugate of the coil sensitivity maps and then summed to obtain MRI images of coil combinations, which can then be provided to subnetworks 308a, 308b for denoising. For example, the MRI data of coil combinations can be redistributed to multiple coils by multiplying the combined data by the coil sensitivity maps of each group of coils. The redistributed multi-coil images can then be transformed to k-space to perform fidelity checks and / or compensation, as described herein. Coil compression layers (e.g., as part of ANN 304) can be used to compress MRI data between coil orientations.

[0035] Figure 4 This illustrates a multi-slice imaging MRI data processing system (e.g., according to one or more embodiments described herein) that can be used by a multi-slice imaging MRI data processing system. Figure 1 System 100 and / or Figure 3 The flowchart illustrates an example operation performed by system 300. As shown, the operation can begin at 402, and at 404, the MRI data processing system can receive a multi-slice imaging MRI dataset, such as an SMS dataset comprising undersampled MRI data (e.g., k-space data or MRI images) associated with multiple MRI slice images of anatomical structures. At 406, the MRI data processing system can use a pre-trained artificial neural network (e.g., Figure 3 The interconnected network shown is used to process multi-slice imaging MRI data to remove noise (e.g., artifacts) from the input MRI data. At 408, a data consistency component (e.g., Figure 3The DC inspector 310 further processes the denoised MRI data (e.g., denoised intermediate MRI images estimated by an artificial neural network). During this process, the fidelity of the estimated MRI data can be verified and / or improved, and MRI data corresponding to the input multi-slice imaging MRI dataset can be reconstructed (e.g., by converting the estimated intermediate MRI images into MRI data via Fourier transform). Based on the reconstructed MRI data, the MRI data processing system can generate an MRI image including multiple slice images from the input MRI dataset by applying IFFT to the reconstructed MRI data at 410. The resulting MRI image can be deconstructed and has higher quality (e.g., similar to the quality of a fully sampled MRI image) compared to the input multi-slice imaging MRI dataset. The operation of the multi-slice imaging MRI data processing system can then terminate at 412.

[0036] Figure 5 Examples are given for training neural networks (e.g., Figure 1 ANN 104 and / or Figure 3 Example procedure 500 (using an instance of ANN 304) to perform the multi-slice imaging MRI data processing operations described herein. Training can be performed using data collected from actual MRI procedures (e.g., undersampled multi-slice imaging MRI data acquired using SMS technology) and / or computer-simulated or computer-enhanced MRI data. For example, fully sampled multi-slice imaging 2D or 3D MRI images collected from a real MRI procedure can be used as a starting point for simulation training data. Selected slice images can be extracted from real MRI data with sufficient slice image distances, and the modulation of slice images can be simulated by adding phase modulation terms to the individual slice images. Data associated with multiple slice images can be combined (e.g., summed) to obtain an SMS dataset directly in image space, or multiple slice images can be first transformed to k-space and then combined (e.g., summed) to obtain an SMS dataset. During the data simulation process, additional in-plane undersampling can be combined with through-plane undersampling.

[0037] like Figure 5As shown, the training process 500 can begin at 502, and at 504, the parameters of the neural network (e.g., weights associated with various filters or kernels of the neural network) can be initialized. The parameters can be initialized, for example, based on samples collected from one or more probability distributions or parameter values ​​from another neural network with a similar architecture. At 506, the neural network can receive a collection of undersampled multi-slice imaging MRI training data (e.g., SMS data) including entangled MRI information associated with multiple MRI slice images. The training data can be provided in the format of undersampled MRI images or undersampled k-space data (e.g., raw MRI data). In the latter case, the undersampled k-space data can be converted into MRI images, for example, via Fourier transform. As described herein, the neural network can include multiple structurally identical subnetworks (e.g., Figure 3 The subnetworks 308a and 308b shown are configured to process training data associated with corresponding MRI slice images, or the neural network may consist of only one network configured to sequentially process training data associated with multiple MRI slice images.

[0038] At 508, the neural network can process training data associated with multiple MRI slice images and generate corresponding intermediate MRI images, in which all or a subset of artifacts (e.g., noise) present in the training data can be removed. At 510, the intermediate MRI images generated by the neural network can be combined and / or transformed (e.g., through the data consistency component of the neural network, such as...). Figure 3A DC checker (310) is used to obtain reconstructed MRI data corresponding to multiple slice images included in the input training dataset. The reconstructed MRI data can then be used to derive corresponding estimated MRI images (e.g., denoised and unraveled MRI images) of the multiple slice images. At 512, a combined training loss can be determined by jointly considering the estimated MRI images associated with the multiple slice images. For example, the combined training loss can be determined based on a triplet loss or contrast loss associated with the estimated MRI images of the multiple MRI slice images. In the example, the combined training loss can be determined based on the average of the corresponding training losses associated with the multiple MRI slice images, while each corresponding training loss (e.g., corresponding to a single MRI slice image) can be determined by comparing the estimated MRI image of the single MRI slice image with the corresponding gold-standard MRI image. Various suitable loss functions can be used to compute the loss described herein, including loss functions based on, for example, mean squared error, L1 norm, L2 norm, structural similarity index (SSIM), etc. Once determined, the combined training loss (e.g., derived by jointly considering the estimated MRI images of the multiple slice images) can be used to determine whether one or more training termination criteria have been met. For example, if the combined training loss is below a predetermined threshold, or if the change in the corresponding combined training loss between two training iterations (e.g., between consecutive training iterations) is below a predetermined threshold, then the training termination criterion can be determined to be met. If the training termination criterion is determined to be met at 512, then the training process 500 can end at 514. If the training termination criterion is determined not to be met at 512, then at 516, the neural network can adjust its parameters by backpropagating the training loss through the neural network (e.g., based on gradient descent associated with the training loss). In examples such as when the neural network comprises multiple structurally similar subnetworks, mirror tuning can be applied to the corresponding parameters of the subnetworks (e.g., weights associated with the filters or kernels of the subnetworks) before the training operation returns to 508. In this way, parameter sharing can be achieved among multiple subnetworks.

[0039] For the sake of simplicity, the training steps are depicted and described in a specific order herein. However, it should be understood that training operations can occur in various orders, simultaneously, and / or with other operations not presented or described herein. Furthermore, it should be noted that not all operations that may be included in the training process are depicted and described herein, and not all exemplified operations need to be performed.

[0040] 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 6This is a block diagram illustrating an example device 600 that can be configured to perform the multi-slice imaging MRI data processing tasks described herein. As shown, device 600 may include a processor (e.g., one or more processors) 602, which may be a central processing unit (CPU), graphics processing unit (GPU), microcontroller, reduced instruction set computer (RISC) processor, application-specific integrated circuit (ASIC), application-specific instruction set processor (ASIP), physical processing unit (PPU), digital signal processor (DSP), field-programmable gate array (FPGA), or any other circuitry or processor capable of performing the functions described herein. Device 600 may also include communication circuitry 604, memory 606, mass storage device 608, input device 610, and / or communication link 612 (e.g., communication bus) through which one or more components shown in the figure exchange information.

[0041] Communication circuitry 604 can 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, and wireless data networks (e.g., Wi-Fi, 3G, 4G / LTE, or 5G networks). Memory 606 may include a storage medium (e.g., a non-transitory storage medium) configured to store machine-readable instructions that, when executed, cause processor 602 to perform one or more functions described herein. Examples of machine-readable media may include volatile or non-volatile memory, including but not limited to semiconductor memory (e.g., electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM)), flash memory, etc.). Mass storage device 608 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-ROMs or DVD-ROMs, etc., on which instructions and / or data may be stored for operation of processor 602. Input device 610 may include a keyboard, mouse, voice-controlled input device, touch-sensitive input device (e.g., touch screen), etc., for receiving user input from device 600.

[0042] It should be noted that device 600 can operate as a standalone device or can be connected to other computing devices (e.g., networked or clustered) to perform the functions described herein. And even in Figure 6 Only one example of each component is shown in the figure, and those skilled in the art will understand that device 600 may include multiple instances of one or more components shown in the figure.

[0043] Although this disclosure has been described according to certain embodiments and generally associated methods, changes and variations of the embodiments and methods will be apparent to those skilled in the art. Therefore, the above description of exemplary embodiments does not limit this disclosure. Other changes, substitutions, and modifications are possible without departing from the spirit and scope of this disclosure. Furthermore, unless specifically stated otherwise, discussions using terms such as “analyze,” “determine,” “enable,” “identify,” and “modify” refer to the actions and processes of a computer system or similar electronic computing device that manipulate and transform data representing physical (e.g., electronic) quantities within the registers and memories of the computer system into other data representing physical quantities within the computer system's memory or other such information storage, transmission, or display devices.

[0044] It should be understood that the above description is intended to be illustrative and not restrictive. Many other embodiments will become apparent to those skilled in the art upon reading and understanding the above description. Therefore, the scope of this disclosure should be determined by reference to the appended claims and the full scope of their equivalents.

Claims

1. A method for reconstructing magnetic resonance imaging (MRI) images, the method comprising: Obtain a simultaneous multi-slice imaging (SMS) dataset, wherein the SMS dataset includes first undersampled MRI data associated with a first MRI slice image of an organ and second undersampled MRI data associated with a second MRI slice image of the organ, the first and second MRI slice images being acquired simultaneously from an MRI procedure; and An artificial neural network (ANN) is used to generate a first reconstructed MRI image of the organ corresponding to the first MRI slice image and a second reconstructed MRI image of the organ corresponding to the second MRI slice image, wherein the ANN is trained to generate the first reconstructed MRI image and the second reconstructed MRI image, and the training includes: The first estimated MRI image is obtained by processing the first undersampled MRI training data of the SMS training dataset through an instance of the ANN, the first undersampled MRI training data corresponding to the first MRI slice image of the SMS training dataset; The second undersampled MRI training data of the SMS training dataset is processed by the instance of the ANN to obtain a second estimated MRI image, the second undersampled MRI training data corresponding to a second MRI slice image of the SMS training dataset; The combined training loss is determined by jointly considering a first training loss associated with the first estimated MRI image and a second training loss associated with the second estimated MRI image; and The parameters of the instance of the ANN are adjusted based on gradient descent using the combined training loss; The ANN includes a first subnetwork and a second subnetwork, the first subnetwork being configured to process the first undersampled MRI data, and the second subnetwork being configured to process the second undersampled MRI data, wherein the first subnetwork and the second subnetwork share a structure and execution parameters.

2. The method according to claim 1, wherein, The first undersampled MRI data includes a first artifact associated with the second undersampled MRI data, wherein the second undersampled MRI data includes a second artifact associated with the first undersampled MRI data, and wherein the first reconstructed MRI image and the second reconstructed MRI image are free of the first artifact and the second artifact, respectively.

3. The method according to claim 1, wherein, The first training loss is determined based on the first estimated MRI image and the corresponding first gold standard MRI image, the second training loss is determined based on the second estimated MRI image and the corresponding second gold standard MRI image, and the combined training loss is determined based on the average of the first training loss and the second training loss.

4. The method according to claim 1, wherein, During the training of the ANN, the first subnetwork of the instance of the ANN is configured to process the first undersampled MRI training data, the second subnetwork of the instance of the ANN is configured to process the second undersampled MRI training data, and adjusting the parameters of the instance of the ANN includes applying mirror updates to the corresponding parameters of the first subnetwork and the second subnetwork.

5. The method according to claim 1, wherein, The first undersampled MRI data associated with the first MRI slice image includes a first undersampled MRI image or first undersampled k-space data, and the second undersampled MRI data associated with the second MRI slice image includes a second undersampled MRI image or second undersampled k-space data.

6. The method according to claim 1, wherein, The ANN also includes a data consistency DC component configured to estimate k-space data based on a first intermediate MRI image generated by the ANN using the first undersampled MRI data and a second intermediate MRI image generated by the ANN using the second undersampled MRI data, wherein the first reconstructed MRI image and the second reconstructed MRI image are generated by applying an inverse Fourier transform to the estimated k-space data.

7. The method according to claim 6, further comprising: Before applying the inverse Fourier transform to the estimated k-space data, at least a portion of the estimated k-space data is replaced with the corresponding portion of the SMS dataset.

8. The method according to claim 6, wherein, The first undersampled MRI data included in the SMS dataset includes MRI data acquired using one or more coils from a first group, the second undersampled MRI data included in the SMS dataset includes MRI data acquired using one or more coils from a second group, and the method further includes: determining corresponding coil sensitivity maps associated with the first group of one or more coils and the second group of one or more coils, wherein the k-space data is further estimated based on the corresponding coil sensitivity maps associated with the first group of one or more coils and the second group of one or more coils.

9. An apparatus for reconstructing magnetic resonance imaging (MRI) images, comprising: One or more processors configured to perform the method for reconstructing magnetic resonance imaging (MRI) images as described in any one of claims 1-8.

Images