Diffusion magnetic resonance data repairing method based on E-GAN network

Abstract

The application belongs to the field of intelligent information processing, and particularly relates to a diffusion magnetic resonance data repairing method based on an E-GAN network. The method comprises the following steps: S1, acquiring paired n data sets of a brain, each data set comprising dMRI image data and T1 weighted image data; S2, pre-processing the dMRI image data and the T1 weighted image data to obtain two high-b-value images and T1 weighted images along a coronal section respectively; S3, constructing a neural network for training the dMRI image data set and the T1 weighted image data set; S4, acquiring paired high-b-value images and T1 weighted images by using the dMRI image data set and the T1 weighted image data set, and training on the neural network constructed in S3 to obtain a final model; and S5, selecting the final model to predict three-dimensional image data of a target mode, and evaluating the data by using an evaluation index. The application effectively retains the combination of low-dimensional morphological information and high-dimensional texture features in medical images, and fully utilizes different features of multi-modal data.

Description

Technical Field

[0001] This invention belongs to the field of intelligent information processing, specifically a method for restoring diffuse magnetic resonance data based on E-GAN networks. Background Technology

[0002] Diffusion magnetic resonance imaging (dMRI) is a crucial tool for exploring brain connectivity and conducting cross-species comparative studies, holding significant value for understanding biological evolution and the diagnosis and treatment of human diseases. Numerous researchers utilize computational neuroscience to analyze dMRI data of the brain, aiming to reflect the structural connectivity characteristics of neurons. However, some dMRI sequences suffer from missing sequence b0 images after acquisition and transcription, affecting the accuracy of computational neuroscience. Therefore, restoring dMRI sequences is of positive significance for current brain science research.

[0003] Generating b0 images from existing modal images and restoring dMRI sequences through image transformation is one solution. As a classic network in the field of deep learning image generation, generative adversarial networks (GANs) have the ability to generate images that are difficult to distinguish from real ones, which plays an important role in magnetic resonance image transformation. Among the many GANs, CycleGAN, representing unsupervised learning, and pix2pix, representing supervised learning, have performed best on computer vision tasks.

[0004] Among these, the unsupervised CycleGAN method, which does not require paired images, has been widely applied to magnetic resonance imaging (MRI) image conversion tasks, such as converting T1-weighted MRI to diffusion tensor images; converting structural MRI T1 and T2 images, and synthesizing fMRI from T1 images; and synthesizing CT, T2, and dMRI from structural MRI. However, its application is limited to unsupervised learning tasks. Compared to unsupervised learning, pix2pix-based methods with paired images exhibit higher accuracy. Some researchers have proposed the Synb0-DisCo method using pix2pix for correcting distortion b0; and networks such as pGAN, MedGAN, Ea-GAN, and ResViT have been used for MRI brain image conversion. However, both of these methods are based on single-modal image data and cannot fully utilize the structural information of multimodal brain images. The resulting data still differs significantly from the real data and cannot meet the requirements of connectivity computation. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method for restoring diffuse magnetic resonance data based on an E-GAN network.

[0006] This invention adopts the following technical solution: a method for restoring diffuse magnetic resonance data based on an E-GAN network, comprising: S1: Obtain n paired datasets of the brain, each dataset including dMRI image data and T1-weighted image data; S2: Preprocess the dMRI image data and T1-weighted image data to obtain two high b-value images and T1-weighted images along the coronal slices, respectively; S3: Construct a neural network to train the dMRI image dataset and the T1-weighted image dataset; S4: Using the dMRI image dataset and the T1-weighted image dataset, obtain pairs of high b-value images and T1-weighted images, train them on the neural network constructed in S3, and obtain the final model; S5: Select the 3D image data of the target modality predicted by the final model, and evaluate the data using evaluation metrics.

[0007] In some embodiments, step S2 includes: S21: Preprocessing the dMRI image dataset: The skull was separated using a depth model on the dMRI image dataset to obtain brain tissue images; Perform b-value image extraction to extract paired b0 images and high b-value images; Intensity normalization is performed on paired image datasets so that the intensity value of each image data is between 0 and 1; The image data samples are resampled into 3D tensors, where the sampling dimensions of the data samples are rescaled to 256×256×256. Slice the scaled image along the coronal plane; S22: Preprocessing the T1 weighted image dataset: T1-weighted images were used to separate the skull using a depth model to obtain brain tissue images; Register the T1-weighted image to the corresponding high b-value image to align the morphology and number of slices; perform intensity normalization on the T1-weighted image so that the intensity value of each image data is between 0 and 1; The image data samples are resampled into 3D tensors with dimensions of 256×256×256; The resampled image is sliced ​​along the coronal plane.

[0008] In some embodiments, the neural network in step S3 adopts an E-GAN network, which includes an E-Net generator architecture and a PatchGAN discriminator architecture. The E-Net generator architecture includes: two identical convolutional encoding structures and a shared convolutional decoding structure; The convolutional coding structure consists of eight downsampling blocks. The first block comprises a convolutional module and a Leaky ReLU activation function, with a 4×4 kernel, a stride of 2, and an edge padding pixel value of 1. The output dimension is set to half the input dimension. The second to seventh blocks are constructed in the order of convolutional module, BatchNorm, and Leaky ReLU activation function, with a 4×4 kernel, a stride of 2, an edge padding pixel value of 1, and an output dimension set to half the input dimension. The last block consists of a convolutional module and a ReLU activation function, with a 4×4 kernel, a stride of 2, an edge padding pixel value of 1, and an output dimension y that is half the input dimension x. The convolutional decoding structure consists of 8 upsampling blocks. The first seven blocks are constructed in the order of transposed convolution, BatchNorm operation, and ReLU activation function. The kernel size of the transposed convolutional layer is 4×4, the stride is 2, the edge padding pixel value is 1, and the output dimension y is twice the input dimension x. The last block is composed of transposed convolution and Tanh activation function. The feature map output by the i-th downsampling block of the two encoders and the input feature map of the (9-i)-th upsampling block are fused through long connections, where 1≤i≤8. The PatchGAN discriminator architecture consists of five blocks. The first and second blocks consist of convolutional modules and LeakyReLU activation functions, with a kernel size of 4×4, a stride of 2, and an edge padding pixel value of 1. The output dimension y is set to half the input dimension x. The third and fourth blocks consist of convolutional modules, BatchNorm operations, and LeakyReLU activation functions, with a kernel size of 4×4, a stride of 2, and an edge padding pixel value of 1. The output dimension y is set to half the input dimension x. The fifth block consists of convolutional modules and BatchNorm operations, with a kernel size of 4×4, a stride of 2, and an edge padding pixel value of 1. The output dimension y is set to half the input dimension x.

[0009] In some embodiments, the mathematical form of the convolutional module in an E-GAN network is: This represents the input to each convolutional layer in the network. express The output after passing through a convolutional layer; This represents a sequence of convolution operations, batch normalization, and activation functions. Defined as: and Represents the weight matrix, the This indicates that a convolution operation is being performed. Indicates batch normalization. This is the activation function.

[0010] In some embodiments, step S4 includes: S41: Train on E-GAN using a coronal slice dataset of high b-value images, b0 images, and T1 images paired with n datasets; S42: Use a model trained for multiple epochs as a pre-trained model, then train on E-GAN on n datasets respectively, and merge the images generated by the E-GAN network along the coronal plane to obtain the three-dimensional predicted b0 image. S43: Rescale the image to have the same dimensions as the original b0 image; S44: By calculating the structural similarity between the predicted result and the real b0 image in each iteration, the n models with the highest similarity index are finally selected as the final models.

[0011] In some embodiments, in step S42, the number of models in an epoch is 20 to 200.

[0012] In some embodiments, step S5, which predicts the three-dimensional image data of the target modality, includes: using the n models obtained in S4 as the final E-GAN models for the n datasets respectively; performing intensity normalization on the images to be generated so that the intensity value of each image data is between 0 and 1; resampling the image data samples into a 3D intensity matrix, wherein the highest sampling dimension of the data samples is rescaled to 256; resampling the scaled images into slices along the coronal plane; obtaining the average three-dimensional probability matrix through the E-GAN model; and rescaling and cropping the matrix to obtain a b0 image with the same size as the original image, which is the generated result.

[0013] In some embodiments, the evaluation metrics in step S5 include peak signal-to-noise ratio (PSNR), structural similarity, and mutual information. A higher PSNR indicates higher image quality; a structural similarity closer to 1 indicates a more similar structure between the generated image and the target image; and higher mutual information indicates a higher correlation between the generated image and the target image.

[0014] Compared with the prior art, the present invention has the following beneficial effects: This invention employs an E-type E-Net generator based on a multimodal data encoding / decoding structure and a long-connection structure. This effectively preserves the combination of low-dimensional morphological information and high-dimensional texture features in medical images, fully utilizing the different features of multimodal data to improve model training performance and enrich the details of the generated images. Simultaneously, it utilizes PatchGAN as the discriminator of the E-GAN, where the output convolution replaces a specific region of the image with local features, allowing the discriminator to better focus on the features of each region. This results in the generator more closely resembling the source image in terms of detail. The peak signal-to-noise ratio (PSNR) reaches 32.2587, structural similarity 0.8822, and mutual information 1.3828 on five datasets, respectively, and the visual effect is also close to the source image. Attached Figure Description

[0015] Figure 1 This is a structural diagram of the E-GAN model of the present invention; Figure 2 This is a flowchart of the present invention; Figure 3 This is a diagram illustrating the prediction results of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] A method for inpainting diffusion magnetic resonance (DMR) data based on an E-GAN network includes: S1: Obtain paired dMRI image datasets and T1-weighted image datasets of the brain.

[0018] The data sample of this invention comprises dMRI images and T1-weighted images of primates from five data centers: the University of Wisconsin-Madison (UWM), the University of California, Davis (UC Davis), Mount Sinai School of Medicine (Mount Sinai-P, Mount Sinai-S), and Axelsen University (AMU). The UWM data center contains 66 datasets, with 53 used for training and 13 for testing; the UDCavis data center contains 19 datasets, with 15 used for training and 4 for testing; the Mount Sinai-P data center contains 6 datasets, with 4 used for training and 2 for testing; the Mount Sinai-S data center contains 5 datasets, with 3 used for training and 2 for testing; and the AMU data center contains 4 datasets, with 3 used for training and 1 for testing.

[0019] S2: Preprocess the dMRI image dataset and the T1-weighted image dataset to obtain high b-value images and T1-weighted images along the coronal slices of the two datasets, respectively.

[0020] Step S2 includes: S21: Preprocessing the dMRI image dataset: Specifically, a depth model was used to separate the skull from the dMRI image dataset to obtain brain tissue images. Then, b-value image extraction was performed to extract pairs of b0 images and high b-value images (b-values ​​around 1000). The paired image datasets were then normalized to ensure that the intensity value of each image data was between 0 and 1. The image data samples were then resampled into 3D tensors, with the sampling dimension of the data samples rescaled to 256×256×256. The scaled images were then sliced ​​along the coronal plane.

[0021] S22: Preprocessing the T1 weighted image dataset: Specifically, the T1-weighted images are used to separate the skull using a depth model to obtain brain tissue images. Then, the T1-weighted images are registered to corresponding high-b-value images, aligning their morphology and slice count. The T1-weighted images are then normalized to ensure that the intensity value of each image data point is between 0 and 1. The image data samples are then resampled into 3D tensors with dimensions of 256×256×256, and the resampled images are sliced ​​along the coronal plane.

[0022] S3: Construct a neural network to train the dMRI image dataset and the T1-weighted image dataset.

[0023] In step S3, the neural network uses E-GAN, which includes the E-Net generator architecture and the PatchGAN discriminator architecture.

[0024] The E-Net generator architecture is an E-type multimodal convolutional neural network based on an encoder-decoder structure and long-connection modules. It contains two identical convolutional encoder structures and a shared convolutional decoder structure. The convolutional encoder structure comprises eight downsampling blocks. The first block consists of a convolutional module and a LeakyReLU activation function, with a 4×4 kernel, a stride of 2, and edge padding of 1 pixel, setting the output dimension to half the input dimension. The second to seventh blocks are constructed in the order of convolutional module, BatchNorm, and LeakyReLU activation function, with a 4×4 kernel, a stride of 2, edge padding of 1 pixel, and the output dimension set to half the input dimension. The last block consists of a convolutional module and a ReLU activation function, with a 4×4 kernel, a stride of 2, edge padding of 1 pixel, and the output dimension y being half the input dimension x. The decoding structure comprises eight upsampling blocks. The first seven blocks are constructed in the order of transposed convolution, BatchNorm operation, and ReLU activation function. The kernel size of the transposed convolution layer is 4×4, the stride is 2, the edge padding pixel value is 1, and the output dimension y is twice the input dimension x. The last block is constructed using transposed convolution and Tanh activation function. The feature map output from the i-th (1≤i≤8) downsampling block of the two encoders is fused with the input feature map from the (9-i)-th upsampling block through a long connection. This decoding structure can fuse the two encoders, making full use of multimodal data. Adding a new decoder would not only increase the model complexity but also prevent the fusion of multimodal features.

[0025] The PatchGAN discriminator architecture consists of five blocks. The first and second blocks consist of convolutional modules and LeakyReLU activation functions, with a 4×4 kernel size, a stride of 2, and an edge padding pixel value of 1. The output dimension y is set to half the input dimension x. The third and fourth blocks consist of convolutional modules, a BatchNorm operation, and a LeakyReLU activation function, with a 4×4 kernel size, a stride of 2, and an edge padding pixel value of 1. The output dimension y is set to half the input dimension x. The fifth block consists of a convolutional module and a BatchNorm operation, with a 4×4 kernel size, a stride of 2, and an edge padding pixel value of 1. The output dimension y is set to half the input dimension x. Through adversarial learning between the generator and the discriminator, suitable generator parameters are finally obtained to achieve high-quality b0 modality image generation.

[0026] The mathematical form of the basic convolutional module is: This represents the input to each convolutional layer in the network. express The output after passing through a convolutional layer.

[0027] The This represents a sequence of convolution operations, batch normalization, and activation functions. Defined as: The and Represents the weight matrix, the This indicates that a convolution operation is being performed. Indicates batch normalization, the This is the activation function.

[0028] The constructed E-GAN network model is trained using the training set. The generator adversarial loss, discriminator adversarial loss, and pixel loss are used as the total loss of the model. Backpropagation is performed with a learning rate of 0.002 to update the weights. The loss calculation formula for the model is as follows: Generate adversarial loss: Discriminator against loss: Pixel loss: Where E represents expectation, G represents generator, D represents discriminator, x and z represent input images of two modalities, and y represents target image.

[0029] The total loss obtained in summary is as follows: The , , These are the weight constants obtained during model training.

[0030] The models used in this invention were trained uniformly for 80 epochs, and then trained separately on five different datasets. The test data consisted of brain dMRI images and T1-weighted images from different data centers that were not used in the training. All images were first intensity-normalized so that the intensity value of each image data point was between 0 and 1. The image data samples were then resampled into a 3D tensor, specifically 256×256×256. The scaled images were then resampled into slices along the coronal plane to construct the final test dataset. The E-GAN model was then used to obtain the predicted b0 image, which was then rescaled and cropped to obtain the predicted b0 image with the same size as the original image. Some brain MRI prediction results are shown below. Figure 3As shown, the leftmost column of the image represents the dataset. For each dataset, the first column is the original high-b-value image, the second column is the prediction result, and the third column is the true b0 image. To better display details, red rectangles frame some image details.

[0031] The E-type generative adversarial network constructed in step S3 has flexible data modality acceptance capabilities, specifically manifested as follows: (1) When the user cannot provide T1 modal data, the input of T1 modal data of E-Net can be replaced with other high B value modal data to extract compensation features without modifying the network structure and training parameters.

[0032] (2) When users want to add more modalities, they only need to copy the E-Net data input encoding part to add a new encoding channel without modifying any other network structure and training parameters, and the new modal data can be connected.

[0033] S4: Using the dMRI image dataset and the T1-weighted image dataset, obtain pairs of high b-value images and T1-weighted images, train them on the neural network constructed in S3, and obtain the final model; Step S4 includes: S41: Train on E-GAN using a dataset of five datasets: coronal slices of paired high b-value images, b0 images, and T1 images.

[0034] S42: To find the appropriate epoch size for this model, models were trained for 10 different epochs at intervals of 20 epochs, ranging from 20 to 200 epochs. The model achieved optimal performance at 80 epochs. Therefore, the model trained for 80 epochs was used as the pre-trained model, and then trained on E-GAN on 5 datasets respectively. The images generated by the E-GAN network were merged along the coronal plane to obtain the 3D predicted b0 image.

[0035] S43: Rescale the image to have the same dimensions as the original b0 image; S44: By calculating the structural similarity between the predicted result and the real b0 image in each iteration, the five models with the highest similarity index are selected as the final models.

[0036] S5: Select the 3D image data of the target modality predicted by the final model, and evaluate the data using evaluation metrics.

[0037] The method for predicting the 3D image data of the target modality using the E-GAN model in step S5 is as follows: the five models obtained in S4 are used as the final E-GAN models for the five datasets respectively. The images to be generated are intensity-normalized so that the intensity value of each image data is between 0 and 1. The image data samples are resampled into a 3D intensity matrix, where the highest sampling dimension of the data samples is rescaled to 256. The scaled images are resampled into slices along the coronal plane. The average 3D probability matrix is ​​obtained through the E-GAN model. The matrix is ​​rescaled and cropped to obtain a b0 image with the same size as the original image. This image is the generated result.

[0038] Evaluation indicators include: Peak signal-to-noise ratio: in This represents the maximum intensity value of a voxel in the y-image. This represents the maximum intensity value of the voxels in the image generated after x enters the generator. Represents y and Calculate the mean square error. ; Structural similarity: in This represents the mean of y. This represents the mean of the images generated by the generator when x enters the image. The variance of y is represented by represent variance Representing y and covariance; Mutual information: Consider two random variables X and Y, whose joint probability density function is: Their marginal probability density functions are respectively and ,in and Representing y and The probability of each pixel appearing in an image. The joint probability for each pixel.

[0039] Evaluation metrics were calculated using the prediction results and the real b0 image to assess the model's performance. Among them, a higher peak signal-to-noise ratio indicates higher image quality; a structural similarity closer to 1 indicates that the generated image's structure is more similar to the target image; and a higher mutual information indicates a higher correlation between the generated image and the target image.

[0040] The evaluation results are shown in Table 1. Compared with the four widely used deep learning models without other preprocessing operations and using the same dataset, the prediction results obtained by the model used in this invention have good performance in all evaluation indicators. The average peak signal-to-noise ratio, structural similarity and mutual information reached 32.2587, 0.8822 and 1.3828 respectively, indicating that the model of this invention has high accuracy.

[0041] Table 1 Comparison of Models The evaluation results are shown in Table 2. Compared with E-GAN, which has only one encoder and reads single-modal data, the E-type generator of this invention has a significant advantage in reading multimodal data using two encoders.

[0042] Table 2 Comparison of Single Encoder and Multi Encoder Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for restoring diffuse magnetic resonance data based on E-GAN networks, characterized in that, include: S1: Obtain n paired datasets of the brain, each dataset including dMRI image data and T1-weighted image data; S2: Preprocess the dMRI image data and T1-weighted image data to obtain two high b-value images and T1-weighted images along the coronal slices, respectively; S3: Construct a neural network to train the dMRI image dataset and the T1-weighted image dataset; In step S3, the neural network adopts an E-GAN network, which includes an E-Net generator architecture and a PatchGAN discriminator architecture. The E-Net generator architecture includes: two identical convolutional encoding structures and a shared convolutional decoding structure; The convolutional coding structure consists of eight downsampling blocks. The first block is composed of a convolutional module and a LeakyReLU activation function, with a kernel size of 4×4, a stride of 2, and an edge padding pixel value of 1. The output dimension is set to half the input dimension. The second to seventh blocks are composed of a convolutional module, a BatchNorm, and a LeakyReLU activation function in that order, with a kernel size of 4×4, a stride of 2, and an edge padding pixel value of 1. The output dimension is set to half the input dimension. The last block is composed of a convolutional module and a ReLU activation function, with a kernel size of 4×4, a stride of 2, and an edge padding pixel value of 1. The output dimension y is half the input dimension x. The convolutional decoding structure consists of 8 upsampling blocks. The first seven blocks are constructed in the order of transposed convolution, BatchNorm operation, and ReLU activation function. The kernel size of the transposed convolutional layer is 4×4, the stride is 2, the edge expansion pixel value is 1, and the output dimension y is twice the input dimension x. The last block is composed of transposed convolution and Tanh activation function. The feature map output by the i-th downsampling block of the two encoders and the input feature map of the (9-i)-th upsampling block are fused through long connections, where 1≤i≤8. The PatchGAN discriminator architecture consists of five blocks. The first and second blocks consist of convolutional modules and LeakyReLU activation functions, with a kernel size of 4×4, a stride of 2, and an edge padding pixel value of 1. The output dimension y is set to half the input dimension x. The third and fourth blocks consist of convolutional modules, BatchNorm operations, and LeakyReLU activation functions, with a kernel size of 4×4, a stride of 2, and an edge padding pixel value of 1. The output dimension y is set to half the input dimension x. The fifth block consists of convolutional modules and BatchNorm operations, with a kernel size of 4×4, a stride of 2, and an edge padding pixel value of 1. The output dimension y is set to half the input dimension x. S4: Using the dMRI image dataset and the T1-weighted image dataset, obtain pairs of high b-value images and T1-weighted images, train them on the neural network constructed in S3, and obtain the final model; S5: Select the 3D image data of the target modality predicted by the final model, and evaluate the data using evaluation metrics.

2. The diffusion magnetic resonance data restoration method based on E-GAN network according to claim 1, characterized in that, Step S2 includes: S21: Preprocessing the dMRI image dataset: The skull was separated using a depth model on the dMRI image dataset to obtain brain tissue images; Perform b-value image extraction to extract paired b0 images and high b-value images; Intensity normalization is performed on paired image datasets so that the intensity value of each image data is between 0 and 1; The image data samples are resampled into 3D tensors, where the sampling dimensions of the data samples are rescaled to 256×256×256. Slice the scaled image along the coronal plane; S22: Preprocessing the T1 weighted image dataset: T1-weighted images were used to separate the skull using a depth model to obtain brain tissue images; Register the T1-weighted image to the corresponding high b-value image to align the morphology and number of slices; perform intensity normalization on the T1-weighted image so that the intensity value of each image data is between 0 and 1; The image data samples were resampled into 3D tensors with dimensions of 256×256×256; The resampled image is sliced ​​along the coronal plane.

3. The diffusion magnetic resonance data restoration method based on E-GAN network according to claim 1, characterized in that, The mathematical form of the convolutional module in the E-GAN network is: This represents the input to each convolutional layer in the network. express The output after passing through a convolutional layer; This represents a sequence of convolution operations, batch normalization, and activation functions. Defined as: The and Represents the weight matrix, the This indicates that a convolution operation is being performed. Indicates batch normalization. This is the activation function.

4. The diffusion magnetic resonance data restoration method based on E-GAN network according to claim 2, characterized in that, Step S4 includes: S41: Train on E-GAN using a coronal slice dataset of high b-value images, b0 images, and T1 images paired with n datasets; S42: Use a model trained for multiple epochs as a pre-trained model, then train on E-GAN on n datasets respectively, and merge the images generated by the E-GAN network along the coronal plane to obtain the three-dimensional predicted b0 image. S43: Rescale the image to have the same dimensions as the original b0 image; S44: By calculating the structural similarity between the predicted result and the real b0 image in each iteration, the n models with the highest similarity index are finally selected as the final models.

5. The diffusion magnetic resonance data restoration method based on E-GAN network according to claim 4, characterized in that, In step S42, the number of models in an epoch is between 20 and 200.

6. The diffusion magnetic resonance data restoration method based on E-GAN network according to claim 1, characterized in that, The process of predicting the three-dimensional image data of the target modality in S5 includes: using the n models obtained in S4 as the final E-GAN models for the n datasets respectively; performing intensity normalization on the images to be generated so that the intensity value of each image data is between 0 and 1; resampling the image data samples into a 3D intensity matrix, wherein the highest sampling dimension of the data samples is rescaled to 256; resampling the scaled images into slices along the coronal plane; obtaining the average three-dimensional probability matrix through the E-GAN model; and rescaling and cropping the matrix to obtain a b0 image with the same size as the original image, which is the generated result.

7. The diffusion magnetic resonance data restoration method based on E-GAN network according to claim 1, characterized in that, The evaluation metrics in step S5 include peak signal-to-noise ratio (PSNR), structural similarity, and mutual information. A higher PSNR indicates higher image quality. A structural similarity closer to 1 indicates that the structure of the generated image is more similar to that of the target image. A higher mutual information indicates that the generated image is more correlated with the target image.

Images