A fast magnetic resonance reconstruction method and system fusing depth sensitivity estimation

Abstract

The application provides a fast magnetic resonance reconstruction method and system fusing depth sensitivity estimation, and particularly relates to the technical field of magnetic resonance imaging and image reconstruction. The scheme comprises the following steps: acquiring multi-channel undersampling K-space data; performing inverse Fourier transform on the multi-channel undersampling K-space data to obtain multi-channel undersampling magnetic resonance images; inputting the multi-channel undersampling magnetic resonance images into a trained convolution reconstruction network for reconstruction to obtain reconstructed magnetic resonance images, wherein the convolution reconstruction network is constructed based on a deep neural network and trained by fusing depth sensitivity estimation. The scheme adopts an undersampling mode to acquire magnetic resonance K-space data as a data set, which can shorten the scanning time, is beneficial to improving the reconstruction efficiency, and can effectively improve the reconstruction precision and robustness of the system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of magnetic resonance imaging and image reconstruction, and particularly relates to a fast magnetic resonance reconstruction method and system fusing deep sensitivity estimation. BACKGROUND

[0002] The prior art magnetic resonance imaging (MRI) is one of important imaging technologies in modern medicine, and has very important clinical value. In the MRI imaging process, a static magnetic field, a gradient magnetic field and a specific frequency radio frequency (RF) pulse magnetic field are applied, so that the imaging object generates a magnetic resonance (MR) signal associated with the spatial position.

[0003] At present, the mainstream deep learning fast magnetic resonance imaging method often needs to estimate the sensitivity information of each coil, i.e. a sensitivity map, in the reconstruction process. The sensitivity map reflects the sensitivity of the coil to the magnetic resonance signal of different spatial positions, and is a key step in the magnetic resonance reconstruction process. However, due to the high computational complexity of the existing sensitivity estimation algorithm, such as the Soft Sense algorithm, the reconstruction efficiency is seriously restricted, and the sensitivity estimation accuracy is not high due to the influence of various noises and uncertainty factors in the magnetic resonance imaging process, resulting in low accuracy of the reconstructed magnetic resonance image. SUMMARY

[0004] In view of the above problems of the prior art, the present application aims to provide a fast magnetic resonance reconstruction method and system fusing deep sensitivity estimation, which aims to solve the problem of low reconstruction efficiency and accuracy caused by the performance of the sensitivity estimation algorithm in the prior art.

[0005] In order to achieve the above-mentioned purpose, the first aspect of the present application provides a fast magnetic resonance reconstruction method fusing deep sensitivity estimation, comprising:

[0006] Acquiring multi-channel undersampled K-space data;

[0007] Performing inverse Fourier transform on the multi-channel undersampled K-space data to obtain a multi-channel undersampled magnetic resonance image;

[0008] Inputting the multi-channel undersampled magnetic resonance image into a trained convolution reconstruction network for reconstruction to obtain a reconstructed magnetic resonance image, wherein the convolution reconstruction network is constructed based on a deep neural network and trained by fusing deep sensitivity estimation.

[0009] Optionally, the trained convolutional reconstruction network comprises a coil sensitivity estimation network, a channel synthesis module and a cascaded reconstruction network, and the multi-channel undersampled magnetic resonance image is input into the trained convolutional reconstruction network for reconstruction to obtain a reconstructed magnetic resonance image, comprising:

[0010] coil sensitivity estimation is performed on the multi-channel undersampled magnetic resonance image by using the coil sensitivity estimation network to obtain coil sensitivity;

[0011] Based on the coil sensitivity and the multi-channel undersampled magnetic resonance image, multi-channel merging is performed by using the channel synthesis module to obtain a merged single-channel image;

[0012] Based on the coil sensitivity, the merged single-channel image is reconstructed by using the cascaded reconstruction network to obtain a reconstructed magnetic resonance image.

[0013] Optionally, the coil sensitivity estimation performed on the multi-channel undersampled magnetic resonance image by using the coil sensitivity estimation network to obtain coil sensitivity comprises:

[0014] The feature map is obtained by using the residual block and the pooling layer in the coil sensitivity estimation network to extract features from the multi-channel undersampled magnetic resonance image;

[0015] The coil sensitivity is obtained by using the up-sampling block and the residual block in the coil sensitivity estimation network to decode the feature map.

[0016] Optionally, the cascaded reconstruction network is formed by alternately stacking a plurality of convolutional networks and data consistency layers, and the merged single-channel image is reconstructed by using the cascaded reconstruction network based on the coil sensitivity to obtain a reconstructed magnetic resonance image, comprising:

[0017] The initial optimized single-channel image is obtained by using the convolutional network to remove undersampling artifacts from the merged single-channel image;

[0018] Based on the coil sensitivity, the initial optimized single-channel image and the multi-channel undersampled K-space data are aligned by using the data consistency layer to obtain an initial reconstructed magnetic resonance image, and the initial reconstructed magnetic resonance image is input into the next convolutional network and the data consistency layer for iterative processing in sequence to obtain a reconstructed magnetic resonance image.

[0019] Optionally, the initial reconstructed magnetic resonance image is obtained by aligning the initial optimized single-channel image and the multi-channel undersampled K-space data based on the coil sensitivity by using the data consistency layer, comprising:

[0020] constructing an encoding operator based on the coil sensitivity, an undersampling mask and a Fourier transform;

[0021] aligning the initial optimized single-channel image and the multi-channel undersampling K-space data using the data consistency layer based on the merged single-channel image and the encoding operator, to obtain an initial reconstructed magnetic resonance image.

[0022] Optionally, the training process of the convolutional reconstruction network comprises:

[0023] obtaining a plurality of groups of undersampling magnetic resonance image samples and corresponding fully-sampled magnetic resonance images based on a fully-sampled K-space data set;

[0024] constructing a convolutional reconstruction network and initializing network parameters, the convolutional reconstruction network being constructed based on a coil sensitivity estimation network, a channel synthesis module and a cascaded reconstruction network;

[0025] inputting the undersampling magnetic resonance image samples into the convolutional reconstruction network to obtain coil sensitivity estimation values and network reconstruction images;

[0026] acquiring coil sensitivity true values, calculating network loss based on the network reconstruction images and corresponding fully-sampled magnetic resonance images, and the coil sensitivity estimation values and corresponding coil sensitivity true values;

[0027] iteratively optimizing parameters of the convolutional reconstruction network using a back propagation algorithm until the network loss converges, to obtain a trained convolutional reconstruction network.

[0028] Optionally, the process of calculating network loss based on the network reconstruction images and corresponding fully-sampled magnetic resonance images, and the coil sensitivity estimation values and corresponding coil sensitivity true values comprises:

[0029] calculating an image reconstruction loss based on the network reconstruction images and corresponding fully-sampled magnetic resonance images;

[0030] calculating a coil sensitivity loss based on the coil sensitivity estimation values and corresponding coil sensitivity true values;

[0031] calculating network loss based on the image reconstruction loss and the coil sensitivity loss.

[0032] The second aspect of the present application provides a fast magnetic resonance reconstruction system fused with deep sensitivity estimation, the system comprising:

[0033] a data acquisition module configured to acquire multi-channel undersampling K-space data;

[0034] A multi-channel image generation module is configured to perform inverse Fourier transform on the multi-channel undersampled K-space data to obtain a multi-channel undersampled magnetic resonance image;

[0035] An image reconstruction module is configured to input the multi-channel undersampled magnetic resonance image into the trained convolutional reconstruction network to reconstruct a reconstructed magnetic resonance image, wherein the convolutional reconstruction network is constructed based on a deep neural network and fused with the sensitivity estimation training.

[0036] The third aspect of the present application provides a magnetic resonance device, which comprises a memory, a processor, and a fast magnetic resonance reconstruction program of fusion deep sensitivity estimation stored on the memory and executable on the processor, and the fast magnetic resonance reconstruction program of fusion deep sensitivity estimation implements the steps of any one of the above-mentioned fast magnetic resonance reconstruction methods of fusion deep sensitivity estimation when executed by the processor.

[0037] The fourth aspect of the present application provides a computer readable storage medium, which stores a fast magnetic resonance reconstruction program of fusion deep sensitivity estimation, and the fast magnetic resonance reconstruction program of fusion deep sensitivity estimation implements the steps of any one of the above-mentioned fast magnetic resonance reconstruction methods of fusion deep sensitivity estimation when executed by a processor.

[0038] Compared with the prior art, the present application has the following advantages:

[0039] The convolutional reconstruction network is constructed based on a coil sensitivity estimation network, a channel synthesis module and a cascaded reconstruction network, the coil sensitivity estimation network is configured to perform coil sensitivity estimation on the multi-channel undersampled magnetic resonance image to obtain coil sensitivity, the channel synthesis module is configured to perform multi-channel merging on the multi-channel undersampled magnetic resonance image based on the coil sensitivity to obtain a merged single-channel image, and the cascaded reconstruction network is configured to reconstruct the merged single-channel image based on the coil sensitivity to obtain a reconstructed magnetic resonance image, which can effectively improve the reconstruction accuracy and robustness of the system. Meanwhile, the magnetic resonance K-space data is collected in an undersampling mode as a data set, which can shorten the scanning time and is conducive to improving the reconstruction efficiency. Furthermore, the convolutional reconstruction network adopts a multi-task learning framework, which can simultaneously perform iterative optimization on the cascaded reconstruction network and the coil sensitivity estimation network in the training stage, so that the cascaded reconstruction network and the coil sensitivity estimation network can be associated with each other, the reconstruction process can effectively guide the accurate estimation of the coil sensitivity, and the sensitivity estimation process can continuously provide optimization feedback for the reconstruction process, thereby improving the model performance and accelerating the convergence speed. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0041] Figure 1 Flow chart of the fusion deep sensitivity estimation fast magnetic resonance reconstruction method of the present application;

[0042] Figure 2 Schematic diagram of the network structure of the fusion deep sensitivity estimation fast magnetic resonance reconstruction of the present application;

[0043] Figure 3 Schematic diagram of the process of acquiring multi-channel undersampled K-space data and undersampled magnetic resonance image of the present application;

[0044] Figure 4 Schematic diagram of the coil sensitivity estimation network structure constructed based on ResUNet of the present application;

[0045] Figure 5 Schematic diagram of the structure of the cascaded reconstruction network of the present application;

[0046] Figure 6 Schematic diagram of the system module of the fusion deep sensitivity estimation fast magnetic resonance reconstruction of the present application;

[0047] Figure 7 Schematic diagram of the magnetic resonance device structure of the present application. DETAILED DESCRIPTION

[0048] In the following description, for the purpose of explanation and not limitation, specific details are set forth, such as particular system configurations, techniques, etc., in order to provide a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application can be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present application with unnecessary detail.

[0049] It should be understood that, when used in the specification and the appended claims, the term "comprising" indicates the presence of the described features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0050] It should also be understood that the terms used in the specification of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the specification and the appended claims of the present application, the singular forms "a," "an," and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0051] It should be further understood that the term "and / or" used in the specification and the appended claims of the present application means any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

[0052] The technical solutions in the embodiments of the present application are described below in conjunction with the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0053] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the scope of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0054] Currently, mainstream deep learning fast magnetic resonance imaging methods, such as Alternating Direction Method of Multipliers-Network (ADMM-Net) and the like, often need to estimate the sensitivity information of each coil, i.e. the sensitivity map, in the reconstruction process, which reflects the sensitivity of the coil to different spatial positions of the magnetic resonance signal. In order to obtain the sensitivity map, it is usually necessary to collect calibration sequences or calibration data sets for calibration operation, which consumes time several times of the model reconstruction time, greatly affecting the reconstruction efficiency. In addition, the estimation accuracy of the sensitivity map directly affects the quality and accuracy of the reconstructed image, but due to various noise and uncertainty factors in the magnetic resonance imaging process, there is often a problem of low accuracy of the sensitivity map estimation, especially in complex imaging environments.

[0055] Based on this, the application provides a fast magnetic resonance reconstruction method fusing deep sensitivity estimation, in the training stage, a convolution reconstruction network is constructed based on a coil sensitivity estimation network, a channel synthesis module and a cascade reconstruction network, paired input data and reference data are used, a multi-task joint training is performed on the convolution reconstruction network through a back propagation algorithm to determine all parameters in the network structure. In the test stage, only test data (multi-channel undersampled image) is provided as input to the network, and the convolution neural network learned through training is used to reconstruct the test data through a forward propagation process. The method uses a multi-task learning framework, that is, the optimization processes of the cascade reconstruction network and the coil sensitivity estimation network are performed simultaneously, so that the cascade reconstruction network and the coil sensitivity estimation network can be associated with each other, the reconstruction process can effectively guide the accurate estimation of the coil sensitivity, and the sensitivity estimation process continuously provides optimization feedback for the reconstruction process, thereby improving the model performance and accelerating the convergence speed, and the trained convolution reconstruction network has good reconstruction accuracy and robustness.

[0056] The embodiment of the application provides a fast magnetic resonance reconstruction method fusing deep sensitivity estimation, which is deployed on a computer, a server or the like electronic device, applied to a scene of reconstructing a magnetic resonance image, and aims at the case of reconstructing an undersampled magnetic resonance image. Figure 1 and Figure 2 As shown in the embodiment, the steps of the method include:

[0057] Step S100: acquiring multi-channel undersampled K-space data;

[0058] Specifically, since undersampling is a sparse sampling means of reducing sampling frequency in the case that the bandwidth capability of the test equipment is insufficient or in order to shorten the scanning time. A multi-channel magnetic resonance imaging system usually includes multiple receiving coils, each of which captures different parts or features of an image. In magnetic resonance imaging, multiple receiving coils are used to simultaneously acquire K-space partial data points in an undersampling manner, which is beneficial to improve the imaging speed and image quality. Therefore, in the process of data acquisition based on the multi-channel magnetic resonance system, first, the signal selection layer is realized by using the selected layer gradient field and the radio frequency excitation, the imaging layer and position are determined, and the layer thickness and position are controlled by adjusting the strength and direction of the selected layer gradient field and the bandwidth and center frequency of the radio frequency field. Then, data is simultaneously acquired by each receiving coil, and the acquired data is filled into the K-space according to the order of phase encoding and frequency encoding to form a two-dimensional complex matrix, that is, the multi-channel undersampled K-space data K u , wherein the multi-channel undersampled K-space data K uThe K-space data includes an under-sampling region and a full-sampling region, and full-sampling data in the full-sampling region is used as self-calibration data. The under-sampling mode used in this embodiment can be a random under-sampling mode or an equidistant under-sampling mode, as shown in FIG. 2. Figure 3 FIG. 3 shows multi-channel under-sampling K-space data constructed by an equidistant under-sampling mode, and the data size can be represented as Nx*Ny*Nc, where Nx represents the number of rows of the collected data, Ny represents the number of columns of the data, and Nc represents the number of receiving channels.

[0059] Step S200: performing inverse Fourier transform on the multi-channel under-sampling K-space data to obtain a multi-channel under-sampling magnetic resonance image;

[0060] Specifically, the collected multi-channel under-sampling K-space data K u is subjected to two-dimensional Fourier transform (2D FFT) to be converted into a multi-channel under-sampling magnetic resonance image I u , and each channel represents an image captured by a receiving coil.

[0061] Step S300: inputting the multi-channel under-sampling magnetic resonance image into a trained convolutional reconstruction network for reconstruction to obtain a reconstructed magnetic resonance image, where the convolutional reconstruction network is constructed based on a deep neural network and fused with deep sensitivity estimation training.

[0062] Specifically, in the training stage, the convolutional reconstruction network is constructed based on a coil sensitivity estimation network, a channel synthesis module, and a cascaded reconstruction network, and all parameters in the network structure are determined by using paired input data and full-sampling data to perform multi-task joint training on the convolutional reconstruction network through a back propagation algorithm, where the coil sensitivity estimation network is used to perform coil sensitivity estimation on the multi-channel under-sampling magnetic resonance image to obtain coil sensitivity; the channel synthesis module is used to perform multi-channel merging on the multi-channel under-sampling magnetic resonance image based on the coil sensitivity to obtain a merged single-channel image; and the cascaded reconstruction network is used to reconstruct the merged single-channel image based on the coil sensitivity to obtain a reconstructed magnetic resonance image; the input data is generated by artificial under-sampling based on the full-sampling data. Through training, the convolutional reconstruction network can realize mapping of the input data (artificial under-sampling data) to full-sampling data; in this embodiment, each coil is represented in a matrix form, and the dimensionality thereof matches the number of image channels. In the test stage, only test data (multi-channel under-sampling image) needs to be provided as input to the network, and the convolutional neural network which has been trained and learned through the forward propagation process reconstructs the test data.

[0063] In this embodiment, the trained convolutional reconstruction network can directly and efficiently and accurately obtain the coil sensitivity through the coil sensitivity estimation network, thereby effectively improving the reconstruction efficiency and accuracy of the system; at the same time, the magnetic resonance K-space data is collected in an undersampling manner, which can shorten the scanning time and improve the reconstruction efficiency.

[0064] In a preferred embodiment, the trained convolutional reconstruction network comprises a coil sensitivity estimation network, a channel synthesis module and a cascaded reconstruction network, and the input of the multi-channel undersampled magnetic resonance image into the trained convolutional reconstruction network in step S300 for reconstruction to obtain the reconstructed magnetic resonance image comprises:

[0065] Step S310: estimating the coil sensitivity of the multi-channel undersampled magnetic resonance image by using the coil sensitivity estimation network to obtain the coil sensitivity.

[0066] Specifically, the coil sensitivity refers to the response degree of each coil when capturing the image, which is affected by multiple factors such as the physical position, shape of the coil and magnetic field distribution. The coil sensitivity estimation network is used to process the multi-channel undersampled magnetic resonance image to calculate the coil sensitivity, which specifically includes: inputting the multi-channel undersampled magnetic resonance image into the coil sensitivity estimation network, the network adaptively extracts self-calibration data features from the multi-channel undersampled magnetic resonance image, and nonlinearly maps them to the coil sensitivity of each channel. The coil sensitivity will be continuously optimized in the network training process, so as to gradually approach the true value.

[0067] In this embodiment, the coil sensitivity is directly estimated from the multi-channel undersampled magnetic resonance image by using the coil sensitivity estimation network, which has higher computational efficiency than other existing sensitivity estimation methods, and has better robustness to complex clinical scenarios, laying a good foundation for subsequent optimization of reconstruction efficiency and image quality.

[0068] In a preferred embodiment, the coil sensitivity estimation of the multi-channel undersampled magnetic resonance image by using the coil sensitivity estimation network in step S310 comprises:

[0069] Step S311: using the residual block and the pooling layer in the coil sensitivity estimation network to extract features from the multi-channel undersampled magnetic resonance image to obtain a feature map.

[0070] Step S312: decoding the feature map by using the up-sampling block and the residual block in the coil sensitivity estimation network to obtain the coil sensitivity.

[0071] Specifically, the coil sensitivity estimation network in this embodiment adopts the ResUNet structure, as shown in Figure 4As shown, first, the multi-channel undersampled magnetic resonance image is preprocessed, including but not limited to normalization, denoising, etc., to ensure that the input data is within a suitable range and to reduce unnecessary computational burden. Further, in order to improve the generalization ability of the coil sensitivity estimation network, random rotation, flipping and other data enhancement operations can also be performed on the multi-channel undersampled magnetic resonance image. Then, the preprocessed multi-channel undersampled magnetic resonance image is processed alternately through a series of residual blocks (RB) and pooling (such as 2x2Pooling) in the encoder. After each residual block, the depth (i.e. the number of channels) of the feature map may increase, and after the pooling layer, the size of the feature map will decrease, and finally the feature map is obtained. Each residual block contains one or more convolution layers and a skip connection (C) for directly passing the intermediate feature map in the encoding layer to the decoding layer, which helps to alleviate the gradient vanishing problem in deep networks and improve the training efficiency and performance of the network. The pooling layer is used to reduce the size of the feature map, reduce the amount of calculation, and extract higher-level features. For example, 2x2 pooling is used to select the maximum pooling or average pooling in a 2x2 region as the output.

[0072] Then, the feature map is decoded alternately through a series of upsampling (such as 2x2 Upsampling) and residual blocks (RB) in the decoder. After each upsampling layer, the size of the feature map increases, and after the residual block, the depth of the feature map may be further adjusted. Finally, a 1x1 convolution is used to adjust the number of channels to match the number of channels of the coil sensitivity, so as to output the feature map, i.e. the predicted coil sensitivity map. The upsampling layer is used to increase the size of the feature map to restore the resolution similar to the input image. Upsampling can be achieved by interpolation or transposed convolution. The residual block is also used to extract and refine features. The 1x1 convolution not only changes the number of channels, but also fuses information in different channels. It should be noted that the number of residual blocks, the number of convolution layers and the depth (number of channels) of the feature map will affect the performance and computational complexity of the network, so they can be flexibly adjusted according to actual application requirements.

[0073] The embodiment constructs a deep learning model based on the ResUNet network, and uses a multi-level encoder and decoder structure to iteratively optimize the multi-channel undersampled magnetic resonance image, which is beneficial to improve the accuracy of the predicted coil sensitivity of the multi-channel undersampled magnetic resonance image.

[0074] It should be noted that the coil sensitivity estimation network in the embodiment adopts the ResUNet structure. As other preferred embodiments, neural network structures such as an artificial neural network (ANN), a recurrent neural network (RNN), a convolutional neural network (CNN), and a Transformer can also be selected to construct the coil sensitivity estimation network. Any coil sensitivity estimation network constructed based on the coil sensitivity estimation concept of the embodiment is within the protection scope of the present application.

[0075] Step S320: based on the coil sensitivity and the multi-channel undersampled magnetic resonance image, multi-channel merging is performed by using the channel synthesis module to obtain a merged single-channel image.

[0076] Specifically, the channel synthesis module performs weighted merging on the undersampled magnetic resonance image of each channel according to the coil sensitivity information of each channel, and further removes the redundant information in each channel to finally obtain a high-quality synthesized single-channel image.

[0077] In the embodiment, the channel synthesis module is used to merge the multi-channel undersampled magnetic resonance image and the coil sensitivity into a synthesized single-channel image. The synthesized single-channel image integrates the information of the multi-channel undersampled magnetic resonance image, reduces the information redundancy, and retains important details and spatial structure features in the original image. In addition, the data dimension of the synthesized single-channel image is greatly reduced compared with the multi-channel undersampled magnetic resonance image, which further reduces the consumption of computing and storage resources of the subsequent cascade reconstruction network.

[0078] Step S330: based on the coil sensitivity, the cascade reconstruction network is used to reconstruct the merged single-channel image to obtain a reconstructed magnetic resonance image.

[0079] Specifically, the cascaded reconstruction network is a deep learning model cascaded by multiple sub-networks, specifically, n ResUNets and data consistency layers (DC) are alternately stacked. In each round of processing, first, the ResUNet is used to extract features from the merged single-channel image, and the feature representation is enhanced through a nonlinear activation function (such as ReLU). Then, based on the coil sensitivity, the undersampling mask and the Fourier transform, an encoding operator is constructed, which is used to accurately reflect the physical characteristics and acquisition strategy of the magnetic resonance image scanner; based on the merged single-channel image and the encoding operator, the data consistency layer is used to align the initial optimized single-channel image and the multi-channel undersampled K-space data, and an initial reconstructed magnetic resonance image is obtained, the specific implementation process is: using the encoding operator to encode the initial optimized single-channel image into K-space to obtain the corresponding K-space data; then minimizing the difference between the K-space data and the multi-channel undersampled K-space data in the sampling region, realizing the consistency calibration of the initial reconstructed magnetic resonance image. It can be seen that the data consistency layer combined with the physical model and the deep learning method can not only fully exert the advantages of deep learning in image feature learning and representation, but also ensure the consistency between the reconstructed image and the real sampling data by means of the constraint of the physical model. By alternately stacking ResUNet and data consistency layer, the network can extract more rich image features in each round of processing, and optimize the reconstruction result under the physical constraint. This combination not only improves the quality of the reconstructed image and reduces the artifacts, but also enhances the robustness of the network to complex data.

[0080] In a preferred embodiment, the training process of the convolutional reconstruction network comprises:

[0081] Step M100: Based on the full-sampling K-space data set, a plurality of groups of undersampled magnetic resonance image samples and corresponding full-sampling magnetic resonance images are obtained.

[0082] Specifically, based on the multi-channel full-sampling K-space data, undersampled K-space data is generated by retrospective undersampling to obtain a plurality of pairs of undersampled K-space data and corresponding full-sampling K-space data. Then, the inverse Fourier transform is used to generate the undersampled magnetic resonance image samples corresponding to the undersampled K-space data, and the full-sampling magnetic resonance images corresponding to the full-sampling K-space data.

[0083] Step M200: Construct a convolutional reconstruction network and initialize network parameters, the convolutional reconstruction network is constructed based on a coil sensitivity estimation network, a channel synthesis module and a cascaded reconstruction network;

[0084] Specifically, the convolution reconstruction network is constructed based on a coil sensitivity estimation network, a channel synthesis module and a cascaded reconstruction network, paired undersampled K-space data and corresponding fully sampled K-space data are used as training data sets, and all parameters in the network structure are determined by multi-task joint training of the convolution reconstruction network through a back propagation algorithm. The coil sensitivity estimation network is used for coil sensitivity estimation of the multi-channel undersampled magnetic resonance image to obtain the coil sensitivity; the channel synthesis module is used for multi-channel merging of the multi-channel undersampled magnetic resonance image based on the coil sensitivity to obtain a merged single-channel image; and the cascaded reconstruction network is used for reconstructing the merged single-channel image based on the coil sensitivity to obtain a reconstructed magnetic resonance image.

[0085] For example, the coil sensitivity estimation network is constructed based on ResUNet, and its specific structure can be found in steps S311 and S312, which will not be repeated here. Similar to the classic segmentation network UNet, ResUNet also uses a large number of concatenation operations to aggregate multi-scale features at different levels, thereby enhancing the expression ability and detail recovery ability of the model. The process of calculating the coil sensitivity by the coil sensitivity estimation network constructed in this embodiment can be briefly described as follows:

[0086]

[0087] wherein S is the coil sensitivity, represents the coil sensitivity estimation network, and the parameters thereof are w.

[0088] The channel synthesis module accepts a multi-channel image and a coil sensitivity as input and outputs a merged single-channel image. The specific functions of the channel synthesis module can be found in step S320, which will not be repeated here. In short, the coil sensitivity S and the multi-channel undersampled image I u are subjected to channel synthesis to obtain a synthesized single-channel image x 0 The process can be described as:

[0089] x 0 = SI u (2)

[0090] The cascaded reconstruction network is a deep learning model formed by cascading a plurality of sub-networks, specifically n ResUNets and data consistency layers are alternately stacked, and the structure is as shown in Figure 5 The specific functions of each module of the cascaded reconstruction network can be found in step S330, which will not be repeated here. Assuming that the cascaded reconstruction network is iterated n times, the calculation process of the kth iteration can be described as:

[0091]

[0092] where k denotes the iteration number, k ∈ [1, n]. Equation (3) represents the computation process of the neural network, which is used to remove the undersampling artifacts in the undersampled image, ResUNet k , where θ denotes the network parameters, z k-1 is the output of ResUNet k . Equation (4) represents the data consistency process, which is used to calibrate the image and prevent the output image from deviating too much from the original undersampled image, denotes the encoding operator, where M denotes the undersampling mask, F denotes the image Fourier transform, S is the coil sensitivity, denotes the conjugate operation, I denotes the identity matrix with the same size as x 0 , and λ is a learnable consistency weight, which is usually initialized to a value greater than 0 and optimized together with the network training process. Since is usually irreversible, equation (4) is further transformed into solving the following linear equation system:

[0093] Ax = b (5)

[0094] where b = x 0 + λz k-1 , and x is the to-be-solved x k . Equation (5) can be solved iteratively by the conjugate gradient descent algorithm. Since the conjugate gradient descent algorithm is completely derivable in the computation process, it can be directly embedded in the neural network training process and participate in the backpropagation process. In addition, the conjugate gradient descent algorithm has the characteristics of second-order convergence, which means that it can approach the optimal solution of the problem in fewer iterations. Compared with the traditional gradient descent method, the conjugate gradient descent can select an optimal search direction at each iteration, avoiding repeated calculation of the direction, thereby improving the convergence speed.

[0095] After n times of repeated execution of the neural network processing process shown in equation (3) and the data consistency process shown in equation (4), a cascade reconstruction network with n levels can be obtained. Specifically, the network is iterated alternately n times, extracts image features in each round, and performs physical constraints through the data consistency layer, thereby gradually optimizing the reconstructed image. The output of each layer is not only affected by the result of the previous layer, but also ensures the consistency of the reconstructed image with the actual sampling data through the data consistency process. The cascade structure enables each iteration to be corrected based on the previous layer, thereby gradually refining the image reconstruction. Generally, the reconstruction accuracy of the cascade reconstruction network will increase as the value of n increases, but too large n will also increase the computational load of the network and reduce the reconstruction efficiency. Therefore, n needs to be selected comprehensively according to the actual situation.

[0096] Further, each ResUNet can be configured to share weights, which can effectively reduce the parameter amount of the model, avoid overfitting, and improve the generalization ability of the model.

[0097] Step M300: inputting the undersampled magnetic resonance image sample into the convolution reconstruction network to obtain a coil sensitivity estimation value and a network reconstruction image;

[0098] Specifically, the undersampled magnetic resonance image sample is input into the convolution reconstruction network constructed above, the coil sensitivity estimation network is used to estimate the coil sensitivity of the multi-channel undersampled magnetic resonance image, and the coil sensitivity is obtained; the multi-channel undersampled magnetic resonance image is merged by the channel synthesis module to obtain a merged single-channel image, and the merged single-channel image is reconstructed by the cascaded reconstruction network to obtain a reconstructed magnetic resonance image.

[0099] Step M400: acquiring a coil sensitivity true value, calculating a network loss based on the network reconstruction image and the corresponding fully sampled magnetic resonance image, and the coil sensitivity estimation value and the corresponding coil sensitivity true value;

[0100] Specifically, the coil sensitivity estimation algorithm is used to process the fully sampled data to calculate the coil sensitivity true value. In this embodiment, a multi-task learning framework is adopted, that is, the optimization processes of the cascaded reconstruction network and the coil sensitivity estimation network are performed simultaneously, so that the cascaded reconstruction network and the coil sensitivity estimation network can be associated with each other, the reconstruction process can effectively guide the accurate estimation of the coil sensitivity, and the sensitivity estimation process can continuously provide optimization feedback for the reconstruction process, thereby improving the model performance and accelerating the convergence speed. Therefore, in this embodiment, the loss between the network reconstruction image and the corresponding fully sampled magnetic resonance image is calculated, and the loss between the coil sensitivity estimation value and the corresponding coil sensitivity true value is calculated, and the network loss is calculated by simultaneously considering the two losses.

[0101] Step M500: iteratively optimizing the parameters of the convolution reconstruction network by using a back propagation algorithm to obtain a trained convolution reconstruction network.

[0102] In a preferred embodiment, the process of calculating the network loss based on the network reconstruction image and the corresponding fully sampled magnetic resonance image, and the coil sensitivity estimation value and the corresponding coil sensitivity true value in step M400 includes:

[0103] Step M410: calculating an image reconstruction loss based on the network reconstruction image and the corresponding fully sampled magnetic resonance image;

[0104] Step M420: calculating a coil sensitivity loss based on the coil sensitivity estimation value and the corresponding coil sensitivity true value;

[0105] Step M430: calculating a network loss based on the image reconstruction loss and the coil sensitivity loss.

[0106] Specifically, the image reconstruction loss is calculated based on a network reconstructed image and a corresponding fully sampled magnetic resonance image The coil sensitivity loss is calculated based on a coil sensitivity estimation value and a corresponding coil sensitivity true value Thus, the network loss is obtained, that is:

[0107]

[0108] Wherein, S p and S g respectively represent a coil sensitivity map output by the coil sensitivity estimation network and a true coil sensitivity map; R p and R g respectively represent a reconstruction result output by the reconstruction network and a fully sampled single-channel image synthesized based on the coil sensitivity; and may be a common image difference measurement loss such as a mean square error loss, a structure similarity index (SSIM) loss, etc.; α is a weight parameter for balancing the coil sensitivity loss and the reconstruction loss, and the value range is related to the loss function type of and , and is usually a real number greater than 0.

[0109] In the embodiment, a multi-task learning framework is adopted, and the network loss shown in formula (6) is used to simultaneously iteratively optimize the cascaded reconstruction network and the coil sensitivity estimation network, so that the cascaded reconstruction network and the coil sensitivity estimation network can be associated with each other, the reconstruction process can effectively guide the accurate estimation of the coil sensitivity, and the sensitivity estimation process can continuously provide optimization feedback for the reconstruction process, thereby improving the model performance and accelerating the convergence speed.

[0110] As shown in Figure 6 , corresponding to the above-mentioned fusion deep sensitivity estimation based fast magnetic resonance reconstruction method, the embodiment of the present application further provides a fusion deep sensitivity estimation based fast magnetic resonance reconstruction system, and the above-mentioned fusion deep sensitivity estimation based fast magnetic resonance reconstruction system comprises:

[0111] A data acquisition module 610 is configured to acquire multi-channel undersampled K-space data.

[0112] A multi-channel image generation module 620 is configured to perform inverse Fourier transform on the multi-channel undersampled K-space data to obtain a multi-channel undersampled magnetic resonance image.

[0113] The image reconstruction module 630 is configured to input the multi-channel undersampled magnetic resonance image into the trained convolutional reconstruction network to reconstruct the magnetic resonance image, wherein the convolutional reconstruction network is constructed based on a deep neural network and fused with deep sensitivity estimation training.

[0114] Specifically, the specific functions of the fast magnetic resonance reconstruction system fused with deep sensitivity estimation in the embodiment can also refer to the corresponding descriptions in the fast magnetic resonance reconstruction method fused with deep sensitivity estimation, which will not be described here.

[0115] Based on the above embodiment, the application further provides a magnetic resonance device, the principle block diagram of which can be shown in the figure. Figure 7 The magnetic resonance device includes a processor, a memory, a network interface and a display screen connected through a system bus. The processor of the magnetic resonance device is configured to provide computing and control capabilities. The memory of the magnetic resonance device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a fast magnetic resonance reconstruction program fused with deep sensitivity estimation. The internal memory provides an environment for the operating system and the fast magnetic resonance reconstruction program fused with deep sensitivity estimation in the non-volatile storage medium. The network interface of the magnetic resonance device is configured to communicate with an external terminal through a network connection. The fast magnetic resonance reconstruction program fused with deep sensitivity estimation is executed by the processor to implement the steps of any one of the fast magnetic resonance reconstruction methods fused with deep sensitivity estimation. The display screen of the magnetic resonance device can be a liquid crystal display screen or an electronic ink display screen.

[0116] Those skilled in the art can understand that, Figure 7 The principle block diagram shown in the figure is only a block diagram of part of the structure related to the application scheme, and does not constitute a limitation on the magnetic resonance device to which the application scheme is applied. The specific magnetic resonance device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0117] In one embodiment, a magnetic resonance device is provided, which includes a memory, a processor and a fast magnetic resonance reconstruction program fused with deep sensitivity estimation stored in the memory and executable on the processor. The fast magnetic resonance reconstruction program fused with deep sensitivity estimation is executed by the processor to implement the steps of any one of the fast magnetic resonance reconstruction methods fused with deep sensitivity estimation provided by the embodiment.

[0118] The embodiment of the present application further provides a computer readable storage medium, and the computer readable storage medium stores a fast magnetic resonance reconstruction program fused with depth sensitivity estimation. The fast magnetic resonance reconstruction program fused with depth sensitivity estimation, when executed by a processor, implements the steps of any fast magnetic resonance reconstruction method fused with depth sensitivity estimation provided by the embodiment of the present application.

[0119] It should be understood that the sequence numbers of the steps in the above embodiments do not mean the execution sequence, and the execution sequence of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiment of the present application.

[0120] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above functional units and modules is taken as an example, and in actual application, the above functions can be completed by different functional units and modules according to needs, that is, the internal structure of the above device is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit, and the integrated unit can be realized in the form of hardware or software. In addition, the specific names of each functional unit and module are only for easy distinction, and do not limit the protection scope of the present application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the above method embodiments, which will not be described here.

[0121] In the above embodiments, the description of each embodiment has its own emphasis, and the parts not described or recorded in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0122] Those skilled in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0123] In the embodiments of the present application, it should be understood that the disclosed apparatus / terminal device and method can be implemented in other ways. For example, the apparatus / terminal device embodiments described above are merely schematic, and the division of the above modules or units is merely a logical function division, and there can be another division manner in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0124] The above-described embodiments are merely used to illustrate the technical solutions of the present application, but not limit the present application; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced equivalently; and these modifications or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A fast magnetic resonance reconstruction method fusing deep sensitivity estimation, characterized in that, The method comprises the following steps: acquiring multi-channel undersampled K-space data; performing inverse Fourier transform on the multi-channel undersampled K-space data to obtain multi-channel undersampled magnetic resonance images; inputting the multi-channel undersampled magnetic resonance images into a trained convolutional reconstruction network for reconstruction to obtain reconstructed magnetic resonance images, wherein the convolutional reconstruction network is constructed based on a deep neural network and fused with deep sensitivity estimation training; the trained convolutional reconstruction network comprises a coil sensitivity estimation network, a channel synthesis module and a cascaded reconstruction network, and the inputting of the multi-channel undersampled magnetic resonance images into the trained convolutional reconstruction network for reconstruction to obtain the reconstructed magnetic resonance images comprises: coil sensitivity estimation of the multi-channel undersampled magnetic resonance images is performed by using the coil sensitivity estimation network to obtain coil sensitivity; based on the coil sensitivity and the multi-channel undersampled magnetic resonance images, multi-channel merging is performed by using the channel synthesis module to obtain merged single-channel images; based on the coil sensitivity, the merged single-channel images are reconstructed by using the cascaded reconstruction network to obtain the reconstructed magnetic resonance images; the cascaded reconstruction network is formed by alternately stacking a plurality of convolutional networks and data consistency layers, the convolutional network is a ResUNet network, and the reconstruction of the merged single-channel images by using the cascaded reconstruction network based on the coil sensitivity to obtain the reconstructed magnetic resonance images comprises: the merged single-channel images are de-aliased by using the convolutional network to obtain initial optimized single-channel images; based on the coil sensitivity, the initial optimized single-channel images are aligned with the multi-channel undersampled K-space data by using the data consistency layer to obtain initial reconstructed magnetic resonance images, and the initial reconstructed magnetic resonance images are inputted into the next convolutional network and data consistency layer for iterative processing in sequence to obtain the reconstructed magnetic resonance images; the alignment of the initial optimized single-channel images with the multi-channel undersampled K-space data based on the coil sensitivity by using the data consistency layer to obtain the initial reconstructed magnetic resonance images comprises: based on the coil sensitivity, an encoding operator is constructed based on an undersampling mask and Fourier transform; based on the merged single-channel images and the encoding operator, the initial optimized single-channel images are aligned with the multi-channel undersampled K-space data by using the data consistency layer to obtain the initial reconstructed magnetic resonance images.

2. The fast magnetic resonance reconstruction method with fusion of depth sensitivity estimates of claim 1, wherein, the coil sensitivity estimation of the multi-channel undersampled magnetic resonance images by using the coil sensitivity estimation network comprises: feature extraction of the multi-channel undersampled magnetic resonance images is performed by using a residual block and a pooling layer in the coil sensitivity estimation network to obtain feature maps; coil sensitivity is obtained by decoding the feature maps by using an up-sampling block and a residual block in the coil sensitivity estimation network.

3. The fast magnetic resonance reconstruction method that fuses depth sensitivity estimates of claim 1, wherein, the training process of the convolutional reconstruction network comprises: based on a full-sampling K-space data set, a plurality of groups of undersampled magnetic resonance image samples and corresponding full-sampling magnetic resonance images are obtained; constructing a convolution reconstruction network and initializing network parameters, the convolution reconstruction network being constructed based on a coil sensitivity estimation network, a channel synthesis module and a cascaded reconstruction network; inputting the undersampled magnetic resonance image sample into the convolution reconstruction network to obtain a coil sensitivity estimation value and a network reconstruction image; acquiring a coil sensitivity true value, calculating a network loss based on the network reconstruction image and a corresponding fully sampled magnetic resonance image, and the coil sensitivity estimation value and a corresponding coil sensitivity true value; iteratively optimizing the parameters of the convolution reconstruction network using a back propagation algorithm until the network loss converges, to obtain a trained convolution reconstruction network.

4. The fast magnetic resonance reconstruction method that fuses depth sensitivity estimates of claim 3, wherein, The process of calculating the network loss based on the network reconstruction image and the corresponding fully sampled magnetic resonance image, and the coil sensitivity estimation value and the corresponding coil sensitivity true value, comprises: calculating an image reconstruction loss based on the network reconstruction image and the corresponding fully sampled magnetic resonance image; calculating a coil sensitivity loss based on the coil sensitivity estimation value and the corresponding coil sensitivity true value; calculating the network loss based on the image reconstruction loss and the coil sensitivity loss.

5. A fast magnetic resonance reconstruction system fusing depth sensitivity estimates, characterized by, The application is applied to the steps of the fast magnetic resonance reconstruction method with fused deep sensitivity estimation as claimed in any one of claims 1-4, comprising: a data acquisition module for acquiring multi-channel undersampled K-space data; a multi-channel image generation module for performing inverse Fourier transform on the multi-channel undersampled K-space data to obtain a multi-channel undersampled magnetic resonance image; an image reconstruction module for inputting the multi-channel undersampled magnetic resonance image into the trained convolution reconstruction network for reconstruction to obtain a reconstructed magnetic resonance image, the convolution reconstruction network being constructed based on a deep neural network and trained by fusing deep sensitivity estimation.

6. A magnetic resonance apparatus characterized by comprising: The magnetic resonance device comprises a memory, a processor, and a fast magnetic resonance reconstruction program with fused deep sensitivity estimation stored on the memory and executable on the processor, and the fast magnetic resonance reconstruction program with fused deep sensitivity estimation implements the steps of the fast magnetic resonance reconstruction method with fused deep sensitivity estimation as claimed in any one of claims 1-4 when executed by the processor.

7. A computer-readable storage medium, characterized in that, The computer readable storage medium stores the fast magnetic resonance reconstruction program with fused deep sensitivity estimation, and the fast magnetic resonance reconstruction program with fused deep sensitivity estimation implements the steps of the fast magnetic resonance reconstruction method with fused deep sensitivity estimation as claimed in any one of claims 1-4 when executed by the processor.

Images