Self-supervised deep learning magnetic resonance image reconstruction method and device and electronic equipment

Abstract

The invention relates to the technical field of medical imaging, in particular to a self-supervised deep learning magnetic resonance image reconstruction method and device and electronic device.The method comprises the steps that multi-channel k-space data are acquired, the multi-channel k-space data comprise multiple pieces of parallel magnetic resonance acquired k-space data, and the k-space data comprise magnetic resonance images; and based on the multi-channel k space data, carrying out self-supervised training on a pre-established expanded neural network. In the training process, a single-channel image is generated, an under-sampling mask is applied to a discarded coil, the single-channel image and the under-sampling mask serve as network input, and a network is supervised and trained through an under-sampling matrix. And reconstructing a magnetic resonance image by using the trained expanded neural network. Therefore, the problems of complicated calculation, sensitivity to noise, dependence on a large amount of full sampling data and the like in related technologies are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical imaging, in particular to a self-supervised deep learning magnetic resonance image reconstruction method and device and electronic equipment. BACKGROUND

[0002] Magnetic resonance imaging is non-invasive and radiation-free, but the slow imaging speed limits its clinical application. In the related art, parallel magnetic resonance imaging uses multiple receiving coils to accelerate sampling, but the traditional reconstruction relies on artificial priori and iterative solution, which is slow and has poor effect under low signal-to-noise ratio. Deep learning methods (such as variable networks and model-driven deep learning) can speed up reconstruction and reduce noise, but still require a large amount of full-sampling data. Self-supervised methods (such as deep undersampling under self-supervision) only use undersampling data for training, but the performance is still inferior to that of supervised methods under high undersampling or low signal-to-noise ratio. SUMMARY

[0003] The present application provides a self-supervised deep learning magnetic resonance image reconstruction method, device and electronic equipment to solve the problems of complex calculation, sensitivity to noise and dependence on a large amount of full-sampling data in the related art.

[0004] The first aspect of the present application provides a self-supervised deep learning magnetic resonance image reconstruction method, comprising the following steps: acquiring multi-channel k-space data, wherein the multi-channel k-space data comprises k-space data of multiple parallel magnetic resonance acquisitions, and the k-space data comprises a magnetic resonance image; performing self-supervised deep learning iterative training on a pre-established unfolded neural network according to the multi-channel k-space data, wherein in the iterative training process, a single-channel image is generated according to the multi-channel k-space data, a discarded coil is extracted from the multi-channel k-space data, an undersampling mask is applied to the discarded coil, an undersampling matrix is generated according to the data of the non-discarded coil, the single-channel image and the undersampling mask are used as inputs of the unfolded neural network, and the training of the unfolded neural network is supervised by using the undersampling matrix; and reconstructing the magnetic resonance image by using the trained unfolded neural network.

[0005] Optionally, in an embodiment of the present application, the iterative reconstruction process of the unfolded neural network comprises: in each round of iterative reconstruction, the single-channel image is denoised by a denoising network; after denoising, the single-channel image is unfolded into a first multi-channel image by using a pre-calculated coil sensitivity matrix; the first multi-channel image, the multi-channel k-space data and the undersampling mask are used as inputs of a data consistency module, and the data consistency module generates a second multi-channel image until the number of iterations of the unfolded neural network reaches a maximum number of iterations, and outputs a reconstructed image of the magnetic resonance image.

[0006] Optionally, in an embodiment of the present application, the calculation formula of the data consistency module is:

[0007]

[0008] wherein, is the coil number calibration data, is the inverse Fourier transform operator, is the undersampling mask, is the Fourier transform operator, is the hyperparameter for balancing regularization and data consistency, is the inverse Fourier transform of the calibration kernel, is the identity transform operator, , is the hyperparameter for balancing data calibration and regularization, is the undersampled k-space data, is the output of the neural network denoising step in this iteration, denotes the element-wise multiplication of the output of the data consistency module and the coil sensitivity matrix followed by a summation over the coil dimension, denotes the element-wise multiplication of the coil sensitivity matrix and its conjugate followed by a summation over the coil dimension, is the coil sensitivity matrix, is the coil sensitivity conjugate.

[0009] Optionally, in an embodiment of the present application, the optimization problem of the unrolled neural network is:

[0010] wherein, is the calibration data, is the convolutional neural network; is the inverse Fourier transform of the calibration kernel , the calibration kernel is the calibration kernel fitted on the self-calibration data set located in the center of k-space:

[0011] wherein, the calibration kernel is the coefficient interpolated from coil j to coil i, is the k-space data point coordinate, denotes the operator for selecting data points around a certain coordinate.

[0012] Optionally, in an embodiment of the present application, before taking the single-channel image and the undersampling mask as the input of the unrolled neural network, further comprising: applying the undersampling matrix on the multi-channel k-space data and the undersampling mask, taking the overall 0 of the discarded coil as the input of the unrolled neural network.

[0013] Optionally, in an embodiment of the present application, the training of the unfolded neural network is supervised by the undersampling matrix, including: The undersampling matrix is taken inversely to obtain the discarded data, which is used to supervise the training of the unfolded neural network, and the formula of the training of the unfolded neural network is as follows:

[0014] Wherein, is the generated undersampling matrix, and and the data on the input set, represents the data on the evaluation set, represents the corresponding undersampling mask of the input set, represents the corresponding undersampling mask of the evaluation set.

[0015] Optionally, in an embodiment of the present application, the training target of the unfolded neural network is:

[0016] Wherein, represents the mean square error loss function, is the model parameter.

[0017] Optionally, in an embodiment of the present application, the unfolded neural network includes an input layer, a plurality of network layers and an output layer, other network layers except the last network layer are composed of convolution, batch normalization and activation function, the last network layer is composed of convolution and batch normalization, and residual connection is used between the input layer and the output layer.

[0018] The second aspect embodiment of the present application provides a self-supervised deep learning magnetic resonance image reconstruction device, including: an acquisition module configured to acquire multi-channel k-space data, wherein the multi-channel k-space data includes a plurality of k-space data of parallel magnetic resonance acquisition, and the k-space data includes a magnetic resonance image; a training module configured to perform self-supervised deep learning iterative training on a pre-established unfolded neural network according to the multi-channel k-space data, wherein, during the iterative training, a single-channel image is generated according to the multi-channel k-space data, a discarded coil is extracted from the multi-channel k-space data, an undersampling mask is applied to the discarded coil, an undersampling matrix is generated according to the data of the non-discarded coil, the single-channel image and the undersampling mask are used as inputs of the unfolded neural network, and the training of the unfolded neural network is supervised by the undersampling matrix; and a reconstruction module configured to reconstruct a magnetic resonance image by using the trained unfolded neural network.

[0019] Optionally, in an embodiment of the present application, the reconstruction module is further configured to, in each iteration of the reconstruction process, denoise the single-channel image by the denoising network; after the denoising, unfold the single-channel image into a first multi-channel image using the pre-computed coil sensitivity matrix; input the first multi-channel image, the multi-channel k-space data and the undersampling mask into the data consistency module, and generate a second multi-channel image by the data consistency module, until the number of iterations of the unfolded neural network reaches a maximum number of iterations, and output a reconstructed image of the magnetic resonance image.

[0020] Optionally, in an embodiment of the present application, the calculation formula of the data consistency module is:

[0021]

[0022] wherein, is the coil number is the calibration data, is an inverse Fourier transform operator, is an undersampling mask, is a Fourier transform operator, is a hyperparameter for balancing regularization and data consistency, is a Fourier inverse transform of the calibration kernel, is an identity transform operator, , is a hyperparameter for balancing data calibration and regularization, is the undersampled k-space data, is the output of the neural network denoising step in the current iteration, represents the element-wise multiplication of the output of the data consistency module and the coil sensitivity matrix, and then summing in the coil dimension, represents the element-wise multiplication of the coil sensitivity matrix and its conjugate, and then summing in the coil dimension, is the coil sensitivity matrix, is the coil sensitivity conjugate.

[0023] Optionally, in an embodiment of the present application, the optimization problem of the unfolded neural network is:

[0024] wherein, is the calibration data, is a convolutional neural network; is a Fourier inverse transform of the calibration kernel , the calibration kernel is a calibration kernel fitted on a self-calibration data set located at the center of the k-space:

[0025] wherein the calibration kernel the coefficient interpolating from coil j to coil i, is the k-space data point coordinate, denotes the operator that selects data points around a certain coordinate.

[0026] Optionally, in an embodiment of the present application, before taking the single-channel image and the undersampling mask as the input of the unrolled neural network, the input module is further configured to apply the undersampling matrix to the multi-channel k-space data and the undersampling mask to take the overall 0 of the discarded coil as the input of the unrolled neural network.

[0027] Optionally, in an embodiment of the present application, the training module is further configured to take the inverse of the undersampling matrix to obtain the discarded data for supervising the training of the unrolled neural network, and the formula for supervising the training of the unrolled neural network is as follows:

[0028] wherein, is the generated undersampling matrix, and and the data on the input set, denotes the data on the evaluation set, denotes the corresponding undersampling mask of the input set, denotes the corresponding undersampling mask of the evaluation set.

[0029] Optionally, in an embodiment of the present application, the training target of the unrolled neural network is:

[0030] wherein, denotes the mean square error loss function, is the model parameter.

[0031] Optionally, in an embodiment of the present application, the unrolled neural network comprises an input layer, a plurality of network layers and an output layer, other network layers except the last network layer are composed of convolution, batch normalization and activation function, the last network layer is composed of convolution and batch normalization, and the residual connection is used between the input layer and the output layer.

[0032] The third aspect embodiment of the present application provides an electronic device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor executes the program to implement the self-supervised deep learning magnetic resonance image reconstruction method as described above.

[0033] Therefore, the present application includes the following beneficial effects: Firstly, multi-channel k-space data is acquired, wherein the multi-channel k-space data comprises a plurality of k-space data acquired in parallel by magnetic resonance, and the k-space data contains a magnetic resonance image; then, a pre-established unfolded neural network is iteratively trained in a self-supervised deep learning manner according to the multi-channel k-space data, in the iterative training process, a single-channel image is generated according to the multi-channel k-space data, a discarded coil is extracted from the multi-channel k-space data and an undersampling mask is applied to the discarded coil, an undersampling matrix is generated according to data of non-discarded coils, the single-channel image and the undersampling mask are taken as inputs of the unfolded neural network, and the training of the unfolded neural network is supervised by using the undersampling matrix; finally, the magnetic resonance image is reconstructed by using the trained unfolded neural network. Through the above method, the embodiments of the present application can realize high-quality image reconstruction without a large amount of full-sampling labeled data, significantly reducing the cost and difficulty of training data acquisition; at the same time, the neural network training process is constrained by using the undersampling matrix, so that the generalization ability of the model under the condition of undersampling is improved, the reconstruction artifacts and noise amplification can be effectively suppressed, the restoration degree of image structure details is improved, the reconstruction speed is improved, and finally the high-fidelity magnetic resonance image reconstruction under the condition of low sampling rate is realized. Thus, the problems of related art, such as complex calculation, sensitivity to noise, and dependence on a large amount of full-sampling data, are solved.

[0034] Additional aspects and advantages of the application will be set forth in part in the description that follows, and in part will become apparent to those skilled in the art upon examination of the following and / or can be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0035] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the following description, taken in conjunction with the accompanying drawings, in which: Figure 1 A flowchart of a self-supervised deep learning magnetic resonance image reconstruction method according to an embodiment of the present application; Figure 2 A schematic diagram of a coil discarding self-supervised training strategy according to an embodiment of the present application; Figure 3 A convolutional neural network structure diagram in an unfolded neural network according to an embodiment of the present application; Figure 4 An unfolded neural network according to an embodiment of the present application; Figure 5 A comparison diagram of reconstruction effects of a self-supervised deep learning magnetic resonance image reconstruction method according to an embodiment of the present application and related art; Figure 6 A SSIM comparison histogram of a self-supervised deep learning magnetic resonance image reconstruction method according to an embodiment of the present application and related art; Figure 7A PSNR comparison column chart of a self-supervised deep learning magnetic resonance image reconstruction method according to an embodiment of the present application and related technologies; Figure 8 An example diagram of a self-supervised deep learning magnetic resonance image reconstruction device according to an embodiment of the present application. Figure 9 A structural schematic diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0036] Embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings, in which the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the accompanying drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0037] The self-supervised deep learning magnetic resonance image reconstruction method, device and electronic device of the embodiments of the present application are described below with reference to the accompanying drawings. In view of the problems mentioned in the above background art, the present application provides a self-supervised deep learning magnetic resonance image reconstruction method, in which, first, multi-channel k-space data is obtained, wherein the multi-channel k-space data includes k-space data of multiple parallel magnetic resonance acquisitions, and the k-space data contains a magnetic resonance image; then, a pre-established unfolded neural network is iteratively trained according to the self-supervised deep learning, in which, during the iterative training process, a single-channel image is generated according to the multi-channel k-space data, a discarded coil is extracted from the multi-channel k-space data and an undersampling mask is applied to it, an undersampling matrix is generated according to the data of the non-discarded coil, the single-channel image and the undersampling mask are used as inputs of the unfolded neural network, and the training of the unfolded neural network is supervised by using the undersampling matrix; finally, the unfolded neural network is used to reconstruct a magnetic resonance image. Through the above method, the embodiments of the present application can realize high-quality image reconstruction without the need for a large amount of full-sampling labeled data, significantly reducing the cost and difficulty of obtaining training data; at the same time, the neural network training process is constrained by using the undersampling matrix, so that the generalization ability of the model under the condition of undersampling is improved, which can effectively suppress the reconstruction artifacts and noise amplification, improve the restoration degree of image structural details, and improve the reconstruction speed, ultimately realizing high-fidelity magnetic resonance image reconstruction under low sampling rate. Thus, the problems of related technologies, such as complex calculation, sensitivity to noise, and dependence on a large amount of full-sampling data, are solved.

[0038] Specifically, Figure 1 A flowchart of a self-supervised deep learning magnetic resonance image reconstruction method according to an embodiment of the present application.

[0039] As Figure 1 shown, the self-supervised deep learning magnetic resonance image reconstruction method includes the following steps: In step S101, multi-channel k-space data is acquired, wherein the multi-channel k-space data comprises k-space data of multiple parallel magnetic resonance acquisitions, and the k-space data comprises a magnetic resonance image.

[0040] The k-space data is a frequency domain signal of a magnetic resonance acquisition, and in the present application, the magnetic resonance image is generated after inverse Fourier transform and channel combination, and is simultaneously used as a model input for reconstruction training. The multi-channel k-space data is magnetic resonance frequency domain raw data acquired in parallel by multiple receiving coils, and in the present application, is used for self-supervised deep learning model training. The magnetic resonance image is a spatial domain image reconstructed from the k-space data, and in the present application, is used as a model output for evaluating reconstruction performance.

[0041] It can be understood that acquiring the multi-channel k-space data can make full use of the information redundancy and noise independence of different receiving coils to provide reliable input for the self-supervised deep learning model. In the training process, a supervision signal is formed by discarding part of the coils, and the network can be trained stably without full sampling reference data, effectively suppressing noise amplification and aliasing artifacts, and significantly improving the reconstruction accuracy, signal-to-noise ratio and structural fidelity of the magnetic resonance image under the condition of undersampling.

[0042] In step S102, a pre-established unfolded neural network is iteratively trained according to the multi-channel k-space data. In the iterative training process, a single-channel image is generated according to the multi-channel k-space data, the discarded coils are extracted from the multi-channel k-space data, an undersampling mask is applied to the discarded coils, an undersampling matrix is generated according to the data of the non-discarded coils, and the single-channel image and the undersampling mask are used as inputs of the unfolded neural network, and the training of the unfolded neural network is supervised by the undersampling matrix.

[0043] The unfolded neural network is a structure that maps the steps of a traditional iterative reconstruction algorithm into neural network layers. In this application, the physical model constraints and deep learning capabilities are combined to reconstruct magnetic resonance images. Self-supervised deep learning iterative training is a multi-round training process in which the network itself generates a supervision signal in the absence of full-sampling reference data. In this application, the neural network is trained to reconstruct the k-space signals of the discarded coils by dynamically discarding part of the coil data and using the remaining coil data to supervise the network. A single-channel image is a single complex image obtained by root square and merging of multi-channel k-space data, which is used as input for the unfolded neural network and for denoising processing. The discarded coil represents a situation in which part of the received data from the multi-channel k-space data is selectively and randomly shielded during self-supervised training, so as to simulate the absence of coil signals and improve the robustness of the unfolded neural network under different coil combinations. The undersampling mask is a binary matrix used to indicate the sampled and unsampled data points in the k-space, which is used to impose undersampling constraints on the discarded coil in this application. The undersampling matrix is a binary matrix generated from the data of the non-discarded coils, which represents the effective sampling positions for supervising the network, and is used to calculate the training loss in this application.

[0044] It can be understood that by using multi-channel k-space data to iteratively train the unfolded neural network, the deep learning model can be efficiently trained without full-sampling reference data. Discarding part of the coils and using the remaining coils to generate a supervision signal can make full use of the redundancy and noise independence of multi-coil information, and improve the reconstruction capability of the network for undersampling data. The introduction of the undersampling mask and the undersampling matrix enables the network to accurately learn the data distribution at different sampling positions during training.

[0045] Figure 2 Figure 1 is a schematic diagram of a coil discarding self-supervised training strategy. A portion of the coils is randomly selected from the original multi-channel k-space data, and the data in these coils is subjected to an undersampling mask as input to the network. During training, the single-channel image output by the network is subjected to channel unfolding and Fourier transform to regenerate multi-channel k-space data, and a loss function is calculated with the actual acquired data on the coils subjected to the mask, and the convolutional neural network parameters are updated accordingly.

[0046] During training, a portion of the coils is first selected from the multi-coil k-space data, denoted as , which is discarded; the other part of the coils is used as input to the network, denoted as . During training, the network is encouraged to use the input part of the coils to fit the data of the discarded coils. The specific steps are as follows: First, the discarded coils are extracted:

[0047] Then an undersampling matrix is generated: ,

[0048] wherein, is the total number of coils, is the two-dimensional size of the k-space data. The undersampling matrix is used to set the data of the discarded coils to 0 as a whole while keeping the data of other coils for network input, so as to realize self-supervised training.

[0049] In an embodiment of the present application, before taking the single-channel image and the undersampling mask as the input of the unrolled neural network, further comprising: applying the undersampling matrix to the multi-channel k-space data and the undersampling mask, and setting the data of the discarded coils to 0 as a whole as the input of the unrolled neural network.

[0050] It can be understood that the embodiments of the present application can artificially manufacture part of the missing data in the self-supervised training, so as to prompt the unrolled neural network to reconstruct the signal of the discarded coil by using the remaining coil information. By setting the discarded coil to 0, the network must rely on the data of the remaining coil to recover the missing information in the training, so as to enhance the accuracy of the network in processing the partial coil missing or undersampling data, and help to reconstruct the image to more completely retain the detailed structure.

[0051] In an embodiment of the present application, the training of the unrolled neural network is supervised by using the undersampling matrix, comprising: taking the inverse of the undersampling matrix to obtain the discarded data, which is used to supervise the training of the unrolled neural network, and the formula for supervising the training of the unrolled neural network is as follows:

[0052] wherein, is the generated undersampling matrix, and and the data on the input set, indicates the data on the evaluation set, indicates the undersampling mask corresponding to the input set, indicates the undersampling mask corresponding to the evaluation set.

[0053] It can be understood that by taking the inverse of the undersampling matrix to obtain the discarded coil data as a supervision signal, the unrolled neural network can utilize the input coil data to predict the data of the missing coil in the training process, so as to realize noise-independent self-supervised learning. The embodiments of the present application utilize the redundancy of multi-coil data, so that the network can still be stably trained in the case of lacking full-sampling reference data.

[0054] In an embodiment of the present application, the training target of the unrolled neural network is:

[0055] wherein, represents the mean squared error loss function, is a model parameter.

[0056] It can be understood that, under the self-supervised condition, the embodiments of the application compare the corresponding k-space signal reconstruction result of only retaining part of the received coil data as the network input with the real k-space signal of the discarded coil, so as to take the k-space data of the discarded coil as the supervision signal to update the network parameters in reverse.

[0057] In an embodiment of the application, the unfolded neural network includes an input layer, a plurality of network layers, and an output layer, other network layers except the last network layer are composed of convolution, batch normalization, and activation function, the last network layer is composed of convolution and batch normalization, and residual connection is used between the input layer and the output layer.

[0058] wherein, the network layer is a basic unit for building the network, composed of convolution, batch normalization, and activation function, for feature extraction and nonlinear mapping, in the application, the last network layer only contains convolution and batch normalization, for generating the final output. The residual connection is a skip connection established between the input layer and the output layer, for maintaining the input information and accelerating the network convergence.

[0059] It can be understood that, through the multi-layer structure of the input layer, the plurality of network layers, and the output layer, the unfolded neural network can extract multi-scale features of the image layer by layer and realize complex nonlinear mapping; the combination of convolution, batch normalization, and activation function helps to enhance the feature expression ability and training stability; the convolution and batch normalization of the last layer ensure the accuracy of the output, and the residual connection between the input layer and the output layer retains the original information, alleviates the gradient vanishing problem, speeds up the network convergence, and improves the detail fidelity and overall quality of the reconstructed image.

[0060] As shown in Figure 3 , the convolutional neural network structure used in the unfolded neural network of the embodiments of the application is composed of ten layers, the first nine layers all include convolution, batch normalization, and activation function, and no pooling layer is set to ensure that the feature information can be completely retained at each layer; the tenth layer only contains convolution and batch normalization, without activation function, so as to avoid the truncation of negative value output. The network introduces residual connection between the input layer and the output layer to effectively prevent gradient vanishing, improve the stability of training, and at the same time ensure that the network can retain the original information under multi-layer nonlinear transformation to realize high-quality image reconstruction.

[0061] In step S103, the unfolded neural network is used to reconstruct the magnetic resonance image.

[0062] It can be understood that the unfolded neural network trained can be used to quickly map the undersampled or multi-coil k-space data to high-quality magnetic resonance images. The embodiments of the present application can significantly improve the reconstruction speed, enhance the structural fidelity and noise suppression capability of the image, and avoid the problems of slow convergence and undersampling noise amplification of the traditional iterative method.

[0063] In an embodiment of the present application, the iterative reconstruction process of the unfolded neural network includes: in each round of iterative reconstruction process, the single-channel image is denoised by the denoising network; after the denoising processing, the single-channel image is unfolded into a first multi-channel image by using a pre-calculated coil sensitivity matrix; the first multi-channel image, the multi-channel k-space data and the undersampling mask are taken as inputs of the data consistency module, and the data consistency module generates a second multi-channel image until the number of iterations of the unfolded neural network reaches a maximum number of iterations, and outputs a reconstructed image of the magnetic resonance image.

[0064] The denoising network is a neural network module for suppressing noise and artifacts in the single-channel image and improving the image reconstruction quality. The coil sensitivity matrix represents the response characteristics of each receiving coil in space and is used to unfold the single-channel image into a multi-channel image. The first multi-channel image is a multi-coil complex image obtained by unfolding the single-channel image by the coil sensitivity matrix, which contains the signal components of each receiving coil. The undersampling mask is used to identify the actual acquisition and missing data positions in the k-space and guide the data consistency module to correct the reconstruction result. The data consistency module receives the first multi-channel image, the multi-channel k-space data and the undersampling mask, and optimizes the reconstruction result by iteration to keep the sampled positions consistent with the original acquisition data, for example, SPIRiT (Iterative Self-consistent Parallel Imaging Reconstruction) can be used to realize the module. The second multi-channel image is a multi-coil complex image output by the data consistency module, which is used as the input of the next iteration or the multi-coil representation of the final reconstructed image. The reconstructed image is the final single-channel magnetic resonance image generated by the unfolded neural network after multiple iterations, which is used for clinical diagnosis or subsequent analysis.

[0065] It can be understood that through the iterative reconstruction process, the unfolded neural network can alternately perform denoising and data consistency optimization in each iteration, so that the influence of undersampling or noise is gradually suppressed; the single-channel image is unfolded into a multi-channel image by using the coil sensitivity matrix, which can fully utilize the spatial redundancy of multi-coil information and improve the data fidelity; and the reconstructed image output after the final iteration to the maximum number of steps is significantly better than the single processing or the traditional method in terms of signal-to-noise ratio, structural details and visual quality.

[0066] A schematic diagram of the unfolded neural network of the embodiments of the present application is shown in FIG. 1, which takes multi-channel undersampled Figure 4 spatial raw acquisition data as input. First, the inverse Fourier transform is performed on the raw data and a single-channel undersampled image is obtained through a channel merging operation, which is taken as the initial input of the iterative reconstruction . .

[0067] In each iteration, the single-channel image is first denoised by a convolutional neural network, and the calculation formula is as follows:

[0068] wherein, x is the input image of the current iteration, y is the output image after denoising, and f is the convolutional neural network.

[0069] Subsequently, the single-channel image after denoising is unfolded into a first multi-channel image using a pre-calculated coil sensitivity matrix, and is input into a data consistency module together with the original multi-channel spatial data and the corresponding undersampling mask to generate a second multi-channel image, thereby ensuring that the reconstructed image is consistent with the actual acquisition data under physical constraints. The algorithm alternately cycles between the convolutional neural network denoising step and the data consistency step until the number of iterations reaches a preset maximum value. Finally, the multi-channel image output by the network is merged through channels to generate a reconstructed magnetic resonance image, thereby realizing high-quality iterative reconstruction.

[0070] In an embodiment of the present application, the calculation formula of the data consistency module is as follows:

[0071]

[0072]

[0073] wherein, x is the input image of the current iteration, y is the output image after denoising, and f is the convolutional neural network. ​​​​​​​​​​​​​​​​​is the output of the neural network denoising step in this iteration, is the output of the data consistency module, and is element-wise multiplied with the coil sensitivity matrix and summed over the coil dimension, is the output of the data consistency module, and is element-wise multiplied with the coil sensitivity matrix and summed over the coil dimension, is the coil sensitivity matrix, is the coil sensitivity conjugate.

[0074] It can be understood that, through the coil sensitivity matrix The weighted synthesis of the multi-channel image can effectively integrate the data of each receiving coil into a high-quality single-channel image, thereby maintaining the spatial consistency and detail integrity of the signal; meanwhile, the average effect of the independent noise of each coil is utilized to significantly reduce the noise level, improve the signal-to-noise ratio of the image, and improve the clarity and usability of the reconstructed image.

[0075] In an embodiment of the present application, the iterative reconstruction process of the unrolled neural network is as follows.

[0076] First, the zero-padded reconstructed image is input, and the neural network denoising step is performed on the single-channel image in each iteration:

[0077] The denoised single-channel image is unfolded into a multi-coil image using the pre-calculated coil sensitivity matrix . Subsequently, the multi-coil image, the multi-channel spatial data and the undersampling mask are input into the data consistency module for iterative data consistency calculation:

[0078] The output multi-coil image of this step is synthesized into a single-channel complex image using the coil sensitivity matrix for the next iteration:

[0079] The algorithm loops between the neural network denoising step and the data consistency step until the number of iterations reaches the maximum value, and finally outputs the reconstructed magnetic resonance image . By alternately performing denoising and data consistency operations, the feature extraction capability of the neural network and the physical model constraint are effectively combined.

[0080] In an embodiment of the present application, the optimization problem of the unrolled neural network is:

[0081] wherein, for calibration data, for a convolutional neural network; is a calibration kernel , the inverse Fourier transform of the calibration kernel is a calibration kernel fitted on a self-calibration data set located at the center of k-space:

[0082] where the calibration kernel is a coefficient interpolating from coil j to coil i, is a k-space data point coordinate, denotes an operator that selects data points around a certain coordinate.

[0083] It can be understood that the embodiments of the application simultaneously consider data consistency, calibration consistency and neural network regularization through a unified optimization framework, so as to realize high-quality magnetic resonance image reconstruction.

[0084] Specifically, the data consistency term guarantees that the reconstructed image is as close as possible to the actual acquisition value in the under-sampled k-space data, so as to ensure that the reconstructed result is true and reliable; the calibration consistency term utilizes the spatial correlation between the multi-channel coils, constrains the multi-channel information through the calibration kernel, and improves the accuracy and consistency of the reconstruction; and the regularization term suppresses noise and artifacts through neural network denoising constraint, and simultaneously enhances the structural details and contrast of the reconstructed image by using the learned image prior.

[0085] For example, in order to better maintain the spatial data consistency in the reconstruction process and enhance the robustness to noise, the SPIRiT algorithm under the MoDL (Model-Based Deep Learning) framework can be used as the data consistency constraint to replace the image domain reconstruction method SENSE (SENSitivity Encoding) used in the traditional physical guided reconstruction algorithm.

[0086] Specifically, first, the convolution kernel is fitted on a self-calibration data set located at the center of k-space, which is pre-acquired, so as to obtain a calibration kernel used for multi-coil data reconstruction. Since the k-space data has certain stationarity and redundancy in the spatial frequency, it can be a priori considered that the convolution kernel fitted on the self-calibration data maintains similar properties in the entire k-space, that is, when the convolution kernel acts on the complete k-space data, it does not change the original Spatial information. This property is called calibration consistency, which ensures that the reconstructed image can maintain the consistency of the physical acquisition data and learn the redundant information between multiple coils through convolution during the convolution operation with the calibration kernel in the reconstruction process, thereby improving the quality and stability of the reconstructed image.

[0087] Figure 5 The reconstruction effect of the self-supervised deep learning magnetic resonance image reconstruction method proposed in the present application is compared with that of the traditional reconstruction method, the supervised deep learning method and the existing self-supervised reconstruction method SSDU (Self-Supervised Learning via Data Undersampling). The selected data is a Weighted image raw data, in which 4 repeated full sampling The spatial data reconstructed by inverse Fourier transform is averaged as a reference image. The remaining images are all reconstructed from 2-fold accelerated, variable density sampled Spatial data. Each reconstructed image is provided with an enlarged view of a partial region below, which is used to show the fidelity of image details under different reconstruction methods, and the SSIM (Structural Similarity Index Measure) of each image is marked in the upper left corner of the image in white small letters.

[0088] In order to verify the effectiveness of the self-supervised deep learning magnetic resonance image reconstruction method proposed in the present application, the public M4Raw data set (multi-contrast, multi-repetition, multi-channel low-field nuclear magnetic resonance Spatial data set, https: / / www.humanconnectome.org) is selected as the experimental data source.

[0089] 80 examples of T1 weighted k-space data are selected for the experiment, of which 50 examples are used for training, 10 examples are used as a validation set for the verification of the supervised algorithm, and the remaining 20 examples are used to test the performance of the algorithm.

[0090] In the data preprocessing process, first, the inverse Fourier transform is performed on the Spatial data of each acquisition of each data to obtain a multi-coil image; the root sum square of the multi-coil image is synthesized to obtain a single-channel amplitude image, and the amplitude images of multiple repeated acquisitions are averaged to obtain a reconstruction gold standard. Further, in Twenty columns of phase-encoded data were selected as self-calibration data in the spatial domain, and a calibration kernel was fitted. The coil sensitivity matrix was estimated based on the calibration data using ESPIRiT (Eigenvalue-based SENSitivity maps for Parallel Imaging Reconstruction), which was used for channel merging.

[0091] To verify the performance of the method proposed in the application, a plurality of benchmark models were selected as controls in the embodiment, including: (1) Traditional reconstruction method: SENSE and SPIRiT reconstruction using Tikhonov regularization and solved by conjugate gradient method; (2) Supervised deep learning reconstruction method: MoDL; (3) Self-supervised reconstruction method: SSDU.

[0092] The above methods respectively use SENSE and SPIRiT as data consistency constraints to ensure the fairness of the comparison results.

[0093] The deep learning reconstruction model proposed in the embodiment is implemented using the PyTorch DL framework; the calculation of the SPIRiT calibration convolution kernel is implemented through MATLAB software; and the estimation of the coil sensitivity matrix is implemented using the Berkeley Advanced Reconstruction Toolbox.

[0094] The denoising network is a 10-layer convolutional neural network without pooling layers, all intermediate layer channel numbers are 64, and a linear rectifier function is used as an activation function, and no activation function is added to the output layer to avoid negative numbers being truncated. Batch normalization is performed after each convolution operation, and a residual connection is introduced between the input and the output. The input of the convolutional neural network and the data consistency module is a complex image, and the real part and the imaginary part are input into the convolutional neural network as two channels, and the output format is consistent with the input. The number of iterations of the unrolled network is set to 10, and the multi-coil complex image output in the last iteration is combined into a single-channel amplitude image by root sum.

[0095] In the SPIRiT reconstruction link, the calibration kernel fitting step includes a hyperparameter , which is set to 0.1; the optimization solving step includes two hyperparameters, which are used to balance the weights of calibration data consistency, acquisition data consistency and regularization. In the application, the above two parameters are set as learnable parameters, the coefficient of calibration data consistency is initialized to 1.0, and the coefficient of the regularization term is initialized to 0.05.

[0096] During the model training process, the adaptive momentum estimation method is used as the optimizer, and the initial learning rate is set to and the ReduceLROnPlateau learning rate scheduler of PyTorch is adopted to reduce the learning rate to 80% of the original value when the training loss does not decrease for two consecutive rounds. Each deep learning reconstruction algorithm is trained for 30 rounds on a NVIDIA A800 graphics card with 80 GB of memory. For the supervised method, the checkpoint with the lowest validation set loss is selected as the final model; for the unsupervised method, the checkpoint with the lowest training loss is selected as the final model.

[0097] To further verify the effectiveness of the self-supervised deep learning magnetic resonance image reconstruction method proposed in the application, the application embodiments are compared with related technologies under the image domain and k-space data consistency constraints, respectively. Table 1 shows the quantitative evaluation indicators of different methods when using the image domain reconstruction algorithm SENSE as the data consistency constraint. Among them, the comparison of the self-supervised deep learning magnetic resonance image reconstruction method of the application, the Tikhonov regularization reconstruction method, the supervised deep learning reconstruction method MoDL, and the self-supervised method SSDU in terms of SSIM and PSNR (Peak Signal-to-Noise Ratio) is as follows: the application is better than Tikhonov and SSDU in SSIM, close to MoDL, and also has an improvement in the PSNR indicator.

[0098] Table 1

[0099] Table 2 further shows the experimental results when the k-space reconstruction algorithm SPIRiT is used as the data consistency constraint. It can be seen that the self-supervised deep learning magnetic resonance image reconstruction method of the application also achieves better reconstruction results than Tikhonov and SSDU, and is closer to the results of the supervised MoDL in SSIM.

[0100] Table 2

[0101] In addition, Figure 6 With Figure 7 respectively intuitively show the comparison results of each embodiment in the structural similarity index (SSIM) and the peak signal-to-noise ratio (PSNR). Among them, the blue-gray column represents the reconstruction result based on the image domain reconstruction algorithm SENSE as the data consistency constraint, and the red-brown column represents the reconstruction result based on the k-space reconstruction algorithm SPIRiT as the data consistency constraint.

[0102] As can be seen from the figure, the self-supervised deep learning magnetic resonance image reconstruction method proposed in the application can significantly improve the reconstruction quality under two constraint conditions, which is that the SSIM and PSNR are better than Tikhonov and SSDU methods, and close to or partially better than MoDL reconstruction.

[0103] According to the self-supervised deep learning magnetic resonance image reconstruction method proposed in the embodiment of the application, first, the multi-channel k-space data is obtained, wherein the multi-channel k-space data includes k-space data of multiple parallel magnetic resonance acquisitions, and the k-space data contains a magnetic resonance image; then, the pre-established unfolded neural network is iteratively trained according to the self-supervised deep learning of the multi-channel k-space data. In the iterative training process, a single-channel image is generated according to the multi-channel k-space data, the discarded coils are extracted from the multi-channel k-space data and an undersampling mask is applied to the discarded coils, an undersampling matrix is generated according to the data of the non-discarded coils, the single-channel image and the undersampling mask are used as the input of the unfolded neural network, and the training of the unfolded neural network is supervised by using the undersampling matrix; finally, the unfolded neural network is used to reconstruct the magnetic resonance image. Through the above method, the embodiment of the application can realize high-quality image reconstruction without a large amount of full-sampling labeled data, significantly reducing the cost and difficulty of training data acquisition; at the same time, the neural network training process is constrained by using the undersampling matrix, so that the generalization ability of the model under the condition of undersampling is improved, which can effectively suppress the reconstruction artifacts and noise amplification, improve the restoration degree of image structure details, and improve the reconstruction speed, finally realizing high-fidelity magnetic resonance image reconstruction under low sampling rate. Thus, the problems of related art, such as complex calculation, sensitivity to noise, and dependence on a large amount of full-sampling data, are solved.

[0104] Secondly, the self-supervised deep learning magnetic resonance image reconstruction device proposed in the embodiment of the application is described with reference to the accompanying drawings.

[0105] Figure 8 is a block schematic diagram of the self-supervised deep learning magnetic resonance image reconstruction device of the embodiment of the application.

[0106] As Figure 8 shown, the self-supervised deep learning magnetic resonance image reconstruction device 10 includes an acquisition module 100, a training module 200, and a reconstruction module 300.

[0107] The acquisition module 100 is configured to acquire multi-channel k-space data, wherein the multi-channel k-space data comprises k-space data of multiple parallel magnetic resonance acquisitions, and the k-space data comprises a magnetic resonance image. The training module 200 is configured to perform self-supervised deep learning iterative training on a pre-established unfolded neural network according to the multi-channel k-space data, wherein in the iterative training process, a single-channel image is generated according to the multi-channel k-space data, a discarded coil is extracted from the multi-channel k-space data, an undersampling mask is applied to the discarded coil, an undersampling matrix is generated according to data of non-discarded coils, the single-channel image and the undersampling mask are taken as inputs of the unfolded neural network, and the training of the unfolded neural network is supervised by using the undersampling matrix.

[0108] In an embodiment of the present application, the reconstruction module 300 is further configured to, in each round of iterative reconstruction process, perform denoising processing on the single-channel image by using a denoising network; after the denoising processing, expand the single-channel image into a first multi-channel image by using a pre-calculated coil sensitivity matrix; take the first multi-channel image, the multi-channel k-space data and the undersampling mask as inputs of a data consistency module, and generate a second multi-channel image by using the data consistency module, until the number of iterations of the unfolded neural network reaches a maximum number of iterations, and output a reconstructed image of the magnetic resonance image.

[0109] In an embodiment of the present application, the calculation formula of the data consistency module is as follows:

[0110]

[0111] wherein, is the calibration data of the coil number is an inverse Fourier transform operator, is an undersampling mask, is a Fourier transform operator, is a hyperparameter for balancing regularization and data consistency, is a Fourier inverse transform of the calibration kernel, is an identity transform operator, , , is a hyperparameter for balancing data calibration and regularization, is undersampled k-space data, is the output of the neural network denoising step in this iteration, denotes the sum of the element-wise multiplication of the output of the data consistency module and the coil sensitivity matrix in the coil dimension, denotes the sum of the element-wise multiplication of the coil sensitivity matrix and its conjugate in the coil dimension, is a coil sensitivity matrix, is a coil sensitivity conjugate.

[0112] In an embodiment of the present application, the optimization problem of the unrolled neural network is:

[0113] wherein, is calibration data, is a convolutional neural network; is a calibration kernel is an inverse Fourier transform of the calibration kernel is a calibration kernel fitted on a self-calibration data set located at the center of k-space:

[0114] wherein, the calibration kernel is a coefficient interpolated from coil j to coil i, is a k-space data point coordinate, denotes an operator that selects data points around a certain coordinate.

[0115] In an embodiment of the present application, before taking the single-channel image and the undersampling mask as the input of the unrolled neural network, the input module is further used to apply an undersampling matrix on the multi-channel k-space data and the undersampling mask, and take the overall 0 of the discarded coil as the input of the unrolled neural network.

[0116] In an embodiment of the present application, the training module is further used to take the inverse of the undersampling matrix to obtain the discarded data, and is used to supervise the training of the unrolled neural network. The formula for supervising the training of the unrolled neural network is as follows:

[0117] wherein, is a generated undersampling matrix, and is used to and input set, denotes data on the evaluation set, denotes the corresponding undersampling mask of the input set, denotes the corresponding undersampling mask of the evaluation set.

[0118] In an embodiment of the present application, the training target of the unrolled neural network is:

[0119] wherein, denotes a mean squared error loss function, is a model parameter.

[0120] In an embodiment of the present application, the unfolded neural network comprises an input layer, a plurality of network layers, and an output layer, other network layers except the last network layer are composed of convolution, batch normalization, and activation function, the last network layer is composed of convolution and batch normalization, and a residual connection is used between the input layer and the output layer.

[0121] It should be noted that the foregoing explanation and description of the self-supervised deep learning magnetic resonance image reconstruction method embodiment also applies to the self-supervised deep learning magnetic resonance image reconstruction device of this embodiment, which will not be described here.

[0122] According to the self-supervised deep learning magnetic resonance image reconstruction device provided in the embodiments of the present application, first, multi-channel k-space data is obtained, wherein the multi-channel k-space data comprises k-space data of a plurality of parallel magnetic resonance acquisitions, and the k-space data contains a magnetic resonance image; then, the pre-established unfolded neural network is iteratively trained according to the self-supervised deep learning, in the iterative training process, a single-channel image is generated according to the multi-channel k-space data, a discarded coil is extracted from the multi-channel k-space data and an undersampling mask is applied to the discarded coil, an undersampling matrix is generated according to the data of the non-discarded coil, the single-channel image and the undersampling mask are used as inputs of the unfolded neural network, and the training of the unfolded neural network is supervised by using the undersampling matrix; finally, the unfolded neural network after the training is used to reconstruct a magnetic resonance image. Through the above device, the embodiments of the present application can realize high-quality image reconstruction without a large amount of full-sampling labeled data, significantly reducing the cost and difficulty of obtaining training data; at the same time, the neural network training process is constrained by using the undersampling matrix, so that the generalization ability of the model under the condition of undersampling is improved, which can effectively suppress the reconstruction artifacts and noise amplification, improve the restoration degree of image structural details, improve the reconstruction speed, and finally realize high-fidelity magnetic resonance image reconstruction under low sampling rate. Thus, the problems of related technologies, such as complex calculation, sensitivity to noise, and dependence on a large amount of full-sampling data, are solved.

[0123] Figure 9 The structure schematic diagram of the electronic device provided in the embodiments of the present application is shown. The electronic device can include: The memory 901, the processor 902, and the computer program stored in the memory 901 and executable on the processor 902.

[0124] The processor 902 implements the self-supervised deep learning magnetic resonance image reconstruction method provided in the above embodiments when executing the program.

[0125] Further, the electronic device further includes: The communication interface 903 is used for communication between the memory 901 and the processor 902.

[0126] The memory 901 is configured to store a computer program capable of running on the processor 902.

[0127] The memory 901 can include a high-speed RAM (Random Access Memory) memory, and can further include a nonvolatile memory, for example, at least one disk memory.

[0128] If the memory 901, the processor 902 and the communication interface 903 are independently implemented, the communication interface 903, the memory 901 and the processor 902 can be connected with each other through a bus and complete communication between each other. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, Figure 9 Only one thick line is used in the figure to represent the bus, but it does not mean that there is only one bus or only one type of bus.

[0129] Optionally, in a specific implementation, if the memory 901, the processor 902 and the communication interface 903 are integrated on a chip, the memory 901, the processor 902 and the communication interface 903 can complete communication between each other through an internal interface.

[0130] The processor 902 can be a CPU (Central Processing Unit), or an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement one or more embodiments of the present application.

[0131] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example" or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms is not necessarily directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples without contradiction.

[0132] Furthermore, the terms "first", "second", etc. are used herein for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly pointing to the number of technical features indicated. Thus, features defined with "first", "second" etc. can explicitly or implicitly include at least one of such features. In the description of the application, the meaning of "N" is at least two, for example two, three, etc., unless explicitly specified otherwise.

[0133] Any process or method descriptions or blocks in flow charts herein, or described elsewhere herein, can be understood as representing modules, segments, or portions of code which include one or more executable instructions for implementing specific logical functions or steps in the process. Alternate implementations are included within the scope of the preferred embodiments of the present application in which additional functionality can be added or some functionality can be removed, by adding, removing or modifying the process steps described and / or described elsewhere herein.

[0134] It should be understood that portions of the present application can be realized with a hardware, software, firmware or a combination thereof. In the above embodiments, the steps or methods can be realized with software or firmware stored in a memory and executed by a suitable instruction execution system. As such, if realized with hardware and in another embodiment, any one or a combination of the following technologies known in the art can be used: discrete logic circuitry having logic gates for implementing logic functions on data signals, application specific integrated circuits having appropriate combinational logic gates, programmable gate arrays, field programmable gate arrays, etc.

[0135] Those of ordinary skill in the art can understand that all or part of the steps carried out by the method of the above embodiments can be instructed by a program to complete the relevant hardware, and the above program can be stored in a computer readable storage medium. When the program is executed, it includes one of the steps of the method embodiments or a combination thereof.

[0136] Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present application, and those of ordinary skill in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A self-supervised deep learning magnetic resonance image reconstruction method, characterized in that, Includes the following steps: Acquire multi-channel k-space data, wherein the multi-channel k-space data includes multiple parallel k-space data acquired by magnetic resonance imaging, and the k-space data includes magnetic resonance images; Self-supervised deep learning iterative training is performed on the pre-established unfolded neural network based on the multi-channel k-space data. During the iterative training process, a single-channel image is generated based on the multi-channel k-space data, discarded coils are extracted from the multi-channel k-space data, an undersampling mask is applied to the discarded coils, an undersampling matrix is ​​generated based on the data of the non-discarded coils, and the single-channel image and the undersampling mask are used as the input of the unfolded neural network. The training of the unfolded neural network is supervised using the undersampling matrix. The unfolded neural network was trained to reconstruct the magnetic resonance image.

2. The self-supervised deep learning magnetic resonance image reconstruction method of claim 1, wherein, The iterative reconstruction process of the unfolded neural network includes: In each round of iterative reconstruction, the single-channel image is denoised using a denoising network; After denoising, the image is expanded into a first multi-channel image using a pre-calculated coil sensitivity matrix. The first multi-channel image, the multi-channel k-space data, and the undersampling mask are used as inputs to the data consistency module. The data consistency module generates a second multi-channel image until the number of iterations of the unfolded neural network reaches the maximum number of iterations, and then outputs the reconstructed image of the magnetic resonance image.

3. The self-supervised deep learning magnetic resonance image reconstruction method of claim 2, wherein, The calculation formula for the data consistency module is as follows: wherein, is the coil number, is the calibration data, is the inverse Fourier transform operator, is the undersampling mask, is the Fourier transform operator, is the hyperparameter for balancing regularization and data consistency, is the inverse Fourier transform of the calibration kernel, is the identity transform operator, , is the hyperparameter for balancing data calibration and regularization, is the undersampled k-space data, is the output of the neural network denoising step in this iteration, denotes the element-wise multiplication of the output of the data consistency module with the coil sensitivity matrix and the summation over the coil dimension, denotes the element-wise multiplication of the coil sensitivity matrix with its conjugate and the summation over the coil dimension, is the coil sensitivity matrix, is the coil sensitivity conjugate.

4. The self-supervised deep learning magnetic resonance image reconstruction method of claim 2, wherein, The optimization problem of the unfolded neural network is: in, For calibration data, It is a convolutional neural network; It is a calibration core Inverse Fourier transform, calibration kernel The calibration kernel fitted to the self-calibration dataset centered in k-space: Among them, calibration core The interpolation coefficient from coil j to coil i, Let the coordinates of the k-space data points be... This indicates an operator that selects data points around a given coordinate.

5. The self-supervised deep learning magnetic resonance image reconstruction method according to claim 1, characterized in that, Before using the single-channel image and the undersampled mask as input to the unfolded neural network, the method further includes: The undersampling matrix is ​​applied to the multi-channel k-space data and the undersampling mask, and the entire discarded coil is set to 0 as the input of the unfolded neural network.

6. The self-supervised deep learning magnetic resonance image reconstruction method according to claim 1, characterized in that, The step of using the undersampling matrix to supervise the training of the expanded neural network includes: The undersampling matrix is ​​inverted to obtain the discarded data, which is used to supervise the training of the unfolded neural network. The formula for supervising the training of the unfolded neural network is as follows: in, For the generated undersampling matrix, use and Input data from the set, This indicates the data on the evaluation set. This represents the undersampling mask corresponding to the input set. This indicates the undersampling mask corresponding to the evaluation set.

7. The self-supervised deep learning magnetic resonance image reconstruction method according to claim 1 or 6, characterized in that, The training objective of the unfolded neural network is: in, This represents the mean squared error loss function. These are the model parameters.

8. The self-supervised deep learning magnetic resonance image reconstruction method according to claim 1, characterized in that, The unfolded neural network includes an input layer, multiple network layers, and an output layer. The network layers other than the last one consist of convolution, batch normalization, and activation functions. The last network layer consists of convolution and batch normalization. The input layer and the output layer use residual connections.

9. A self-supervised deep learning magnetic resonance image reconstruction device, characterized in that, Includes the following steps: The acquisition module is used to acquire multi-channel k-space data, wherein the multi-channel k-space data includes multiple parallel k-space data acquired by magnetic resonance imaging, and the k-space data includes magnetic resonance images; The training module is used to perform self-supervised deep learning iterative training on a pre-established unfolded neural network based on the multi-channel k-space data. During the iterative training process, a single-channel image is generated based on the multi-channel k-space data, discarded coils are extracted from the multi-channel k-space data, an undersampling mask is applied to the discarded coils, an undersampling matrix is ​​generated based on the data of the non-discarded coils, the single-channel image and the undersampling mask are used as inputs to the unfolded neural network, and the training of the unfolded neural network is supervised using the undersampling matrix. A reconstruction module is used to reconstruct magnetic resonance images using the trained unfolded neural network.

10. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the self-supervised deep learning magnetic resonance image reconstruction method according to any one of claims 1-8.

Images