A hippocampal subregion segmentation method and system based on ultra-high field magnetic resonance imaging reconstruction

Abstract

The invention belongs to the field of medical image processing, and specifically relates to a hippocampal subregion segmentation method and system based on ultra-high field magnetic resonance image reconstruction, comprising: preprocessing magnetic resonance image data, inputting the preprocessed image into an ultra-high field magnetic resonance image reconstruction network for feature extraction, and fusing and reconstructing the extracted features to obtain a reconstructed MRI; inputting the reconstructed MRI into a discriminator network for discrimination, and inputting the discriminated MRI into a hippocampal subregion segmentation network to extract multi-scale feature information to obtain a hippocampal subregion segmentation probability map; performing hippocampal subregion segmentation on the magnetic resonance image according to the hippocampal subregion segmentation probability map; the invention introduces a self-attention mechanism, feature matching loss and ROI loss in the ultra-high field magnetic resonance image reconstruction process to restore more anatomical details of the hippocampal region, introduces multi-scale deep supervision in the hippocampal subregion segmentation process, and designs an anatomical weighted cross entropy loss to utilize rich prior knowledge, thereby improving the accuracy of the segmentation result.

Description

Technical Field

[0001] The present invention belongs to the field of medical image processing, and in particular relates to a hippocampal subregion segmentation method and system based on ultra-high field magnetic resonance image reconstruction. Background Art

[0002] The hippocampus is an important cortical area closely related to long-term episodic memory and spatial encoding. This is reflected in the different degrees of influence of anatomical changes in different hippocampal subregions on various neurodegenerative diseases, such as Alzheimer's disease and schizophrenia. Therefore, it is necessary to quantitatively analyze the morphological and structural changes of hippocampal subregions. The structure of the hippocampus is irregular, and its intensity is relatively close to that of the surrounding brain tissue. The boundaries of the hippocampal subregions are relatively complex, and the brightness differences between different subregions are not obvious. If the hippocampal subregions are manually segmented on brain MRI, there are problems such as long time consumption, low reproducibility, and easy introduction of human errors. It also requires experienced experts to complete, which is difficult to meet the current needs of medical big data. Therefore, a high-precision automatic segmentation method for hippocampal subregions is urgently needed. In the existing automatic segmentation methods for hippocampal subregions, the brain MRI data used are mainly 3T MRI and 7TMRI. Compared with 7T MRI, the tissue contrast and boundary positioning of 3T MRI are relatively poor, and the texture details of the internal and surrounding tissues of the hippocampus cannot be clearly presented, and the segmentation effect of hippocampal subregions is poor. However, due to the high price of 7T MRI scanners, extremely high imaging costs, and longer imaging time than 3T MRI, it is not available in clinical testing. Therefore, ultra-high field MRI (7T MRI) can be synthesized from conventional MRI (3T MRI) to improve the problem of insufficient input image resolution, and the synthesized ultra-high field MRI can be used to improve the accuracy of conventional MRI segmentation of hippocampal subregions. However, since high-quality segmentation strategies are limited by the resolution of the input image, when 7T MRI is not available, how to integrate the ultra-high field MRI reconstruction network into the mature hippocampal subregion segmentation network to improve segmentation accuracy is an urgent problem to be solved. Summary of the invention

[0003] In order to solve the problems existing in the above prior art, the present invention proposes a hippocampal subregion segmentation method based on ultra-high field magnetic resonance image reconstruction, the method comprising: obtaining a brain nuclear magnetic resonance image to be segmented, preprocessing the brain nuclear magnetic resonance image, and inputting the preprocessed brain nuclear magnetic resonance image into a joint network model to obtain a segmentation result; wherein the joint network model comprises an ultra-high field magnetic resonance image reconstruction network and a hippocampal subregion segmentation network; the ultra-high field magnetic resonance image reconstruction network comprises a generation network and a discrimination network;

[0004] The process of training the joint network model includes:

[0005] S1: Obtain a magnetic resonance imaging dataset and preprocess the dataset; divide the preprocessed dataset into a training set and a validation set;

[0006] S2: Input the images in the training set into the ultra-high field magnetic resonance image reconstruction network, use the generative network to extract the deep and shallow features of the input image, and fuse and reconstruct the deep and shallow features to obtain the initial reconstructed MRI;

[0007] S3: Input the initial reconstructed MRI and the conventional MRI in the training set into the discriminator network for comparison with the real MRI. If the output is false, return to step S2. If the output is true, output the reconstructed MRI.

[0008] S4: Input the reconstructed MRI and conventional MRI into the hippocampal subregion segmentation network to extract multi-scale feature information and obtain the hippocampal subregion segmentation probability map;

[0009] S5: Obtain the segmentation truth images corresponding to the reconstructed MRI and conventional MRI according to the hippocampal subregion segmentation probability map, set the initial category average dice to 0, and use the validation set data to verify the model performance after each epoch, and calculate the category average dice. If it is higher than the category average dice value of the previous epoch, save the model, otherwise continue training;

[0010] S6: Calculate the loss function of the joint network model, continuously adjust the model parameters, and complete the model training when the loss function converges.

[0011] A hippocampal subregion segmentation system based on ultra-high field magnetic resonance image reconstruction, the system comprising: a joint network model training module, a test set magnetic resonance image receiving module, a test set magnetic resonance image segmentation processing module and a segmentation result output module;

[0012] The joint network model training module includes a training image receiving unit, a training image analysis and processing unit, an ultra-high field magnetic resonance imaging synthesis unit, and a hippocampal subregion segmentation unit;

[0013] The training image receiving unit is used to receive a training image;

[0014] The training image analysis and processing unit is used to perform feature analysis on the training image and fuse the feature information of the image;

[0015] The ultra-high field magnetic resonance image synthesis unit reconstructs the ultra-high field MRI image by using an adversarial learning method according to the fused feature information;

[0016] The hippocampal subregion segmentation unit is used to perform feature analysis on the training image and the synthesized ultra-high field MRI image, fuse effective feature information, and perform segmentation processing on the fused feature information until the loss functions of the ultra-high field MRI image synthesis unit and the hippocampal subregion segmentation unit in the joint network model converge to a preset threshold, thereby completing the deep learning training;

[0017] The test set magnetic resonance image receiving module is used to receive magnetic resonance image data information to be segmented;

[0018] The test set magnetic resonance image segmentation processing module is used to perform segmentation processing on the magnetic resonance image data information to be segmented received by the magnetic resonance image receiving module (2);

[0019] The segmentation result output module is used to output the hippocampal subregion segmentation result.

[0020] Beneficial effects of the present invention:

[0021] The present invention introduces self-attention mechanism, feature matching loss and ROI loss in the process of ultra-high field magnetic resonance imaging reconstruction to restore more anatomical details of the hippocampus region; the generated ultra-high field MRI and the original conventional MRI are input into the segmentation network, a multi-scale deep supervision mechanism is introduced in the segmentation process, and an anatomical weighted cross entropy loss is designed to utilize rich anatomical prior knowledge. The obtained hippocampal subregion segmentation results can significantly improve the segmentation accuracy of conventional MRI, providing a more reliable measurement for the detection and diagnosis of neurodegenerative diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 The overall framework diagram of the hippocampal subregion segmentation network reconstructed by combined ultra-high field magnetic resonance imaging of the present invention;

[0023] Figure 2 This is a network framework diagram for ultra-high field magnetic resonance imaging reconstruction of the present invention;

[0024] Figure 3 This is a framework diagram of the self-attention module of the present invention;

[0025] Figure 4 It is a network framework diagram of hippocampal subregion segmentation of the present invention;

[0026] Figure 5 An example of a 7T MRI generated for the left hippocampus of the present invention and a corresponding real 7T MRI result;

[0027] Figure 6 A comparison diagram of the segmentation results of the left hippocampal subregion by the embodiment of the present invention and the segmentation results by other methods;

[0028] Figure 7This is a comparison diagram of the segmentation results of the right hippocampal subregions of the embodiment of the present invention and the segmentation results of other methods. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] A hippocampal subregion segmentation method based on ultra-high field magnetic resonance imaging reconstruction includes: obtaining a brain nuclear magnetic resonance image to be segmented, preprocessing the brain nuclear magnetic resonance image, and inputting the preprocessed brain nuclear magnetic resonance image into a joint network model to obtain a segmentation result. The joint network model includes an ultra-high field magnetic resonance imaging reconstruction network and a hippocampal subregion segmentation network; the ultra-high field magnetic resonance imaging reconstruction network includes a generation network and a discrimination network.

[0031] The process of training the joint network model includes:

[0032] S1: Obtain a magnetic resonance imaging dataset and preprocess the dataset; divide the preprocessed dataset into a training set and a validation set;

[0033] S2: Input the images in the training set into the ultra-high field magnetic resonance image reconstruction network, use the generative network to extract the deep and shallow features of the input image, and fuse and reconstruct the deep and shallow features to obtain the initial reconstructed MRI;

[0034] S3: Input the initial reconstructed MRI and the conventional MRI in the training set into the discriminator network for comparison with the real MRI. If the output is false, return to step S2. If the output is true, output the reconstructed MRI.

[0035] S4: Input the reconstructed MRI and conventional MRI into the hippocampal subregion segmentation network to extract multi-scale feature information and obtain the hippocampal subregion segmentation probability map;

[0036] S5: Obtain the segmentation truth images corresponding to the reconstructed MRI and conventional MRI according to the hippocampal subregion segmentation probability map, set the initial category average dice to 0, and use the validation set data to verify the model performance after each epoch, and calculate the category average dice. If it is higher than the category average dice value of the previous epoch, save the model, otherwise continue training;

[0037] S6: Calculate the loss function of the joint network model, continuously adjust the model parameters, and complete the model training when the loss function converges.

[0038] The process of preprocessing brain MRI images includes: denoising and bias field correction for all MRI data, including 3T T2 MRI, 7T T2 MRI, 3T T1MRI and 7T T1 MRI; aligning 3T T2 MRI with 3T T1 MRI, and aligning 7T T2 MRI with 7T T1 MRI; linearly registering 7T T1 MRI to 3T T1MRI, and linearly registering 3T T1 MRI to an unbiased template using ANTS deformable registration; extracting two cubic ROIs containing the left and right hippocampi from the unbiased template; for each side, mapping the generated binary segmentation into the template space by merging the anatomical labels of that side into a single label, and extracting a rectangular box covering all hippocampal regions in the template space to obtain a cubic ROI; upsampling the cubic ROIs of the left and right templates; mapping MRI data and manual labels of all modalities to the corresponding hemispheric template ROI to obtain the preprocessed brain MRI images.

[0039] In this embodiment, Figure 1 As shown in the figure, the joint network model includes an ultra-high field magnetic resonance imaging reconstruction network (Ultra_SynNet) and a hippocampal subregion segmentation network (HS_SegNet). First, the preprocessed 3T T1 and 3T T2 MRI are sent to the generator together, and the generated 7T T1 / T2 MRI is output. Then, the 3T image is spliced ​​with the generated 7T image and the real 7T image and sent to the discriminator. Through adversarial training, the generator is prompted to generate better quality 7T images. Then the generated 7T image and the 3T image are input into the segmentation network together to obtain the final hippocampal subregion segmentation probability map.

[0040] In this embodiment, if Figure 2 As shown in the figure, the preprocessed 3T T1 and 3T T2 MRI are input into the generator after channel splicing. The generator outputs realistic 7T T1 / T2 MRI. The original input 3T image, the generated 7T image and the corresponding true value image are respectively sent to the discriminator. Through adversarial training, the generator is prompted to synthesize high-quality 7T images, thereby better improving the accuracy of hippocampal subregion segmentation of conventional MRI.

[0041] The generator uses the 3D variant of Pix2pixGAN, which is a symmetrical 6-layer encoder-decoder structure, including 6 3D convolutional layers, 6 3D deconvolutional layers, and 5 feature splicing layers; each 3D convolutional layer passes through batch normalization and ReLU activation function; each 3D deconvolutional layer passes through the splicing layer, batch normalization and ReLU activation function, and the first three 3D deconvolutional layers pass the Dropout operation after the splicing layer, batch normalization and ReLU activation function; the convolution kernel size of the first two 3D convolutional layers and the last two 3D deconvolutional layers is 5, and the remaining convolution kernel size is 3 ; After each 3D deconvolution layer passes through batch normalization, ReLU activation function or batch normalization, ReLU activation function, and Dropout operation, it is spliced ​​with the convolution layer with the same number of feature maps and passed through batch normalization and ReLU activation function to obtain the feature splicing layer; each 3D convolution layer and 3D deconvolution layer uses spectral normalization; the fourth 3D convolution layer and the fourth 3D deconvolution layer pass through a 3D self-attention module after passing through batch normalization and ReLU activation function, allowing the generator to effectively model the relationship between separated spatial regions, so that the generated image contains real anatomical details.

[0042] The generator input is a set of preprocessed images, which are 3T T1-weighted MRI and 3TT2-weighted MRI; the generator output passes through a three-dimensional convolution layer with a convolution kernel size of 3, a three-dimensional self-attention module, and a hyperbolic tangent function in sequence, and the final generator output is a realistic 7TT1-weighted MRI or 7T T2-weighted MRI.

[0043] In this embodiment, the discriminator is PatchGAN, which is a fully convolutional network including 5 three-dimensional convolutional layers; each three-dimensional convolutional layer passes through batch normalization, ReLU activation function and Dropout operation; the third three-dimensional convolutional layer passes through a three-dimensional self-attention module after passing through batch normalization, ReLU activation function and Dropout operation; each three-dimensional convolutional layer uses spectral normalization; the discriminator input is the result of splicing the generator output and the generator input channel, or the result of splicing the real magnetic resonance image corresponding to the generator output and the generator input channel.

[0044] In another embodiment, the discriminator output passes through a three-dimensional convolution layer with a convolution kernel size of 3, batch normalization, and a sigmoid operation in sequence; the discriminator determines whether the input image is a real image, and incorporates the result into the final objective function in the form of a loss, so as to promote the generator to generate a more realistic 7T MRI.

[0045] The loss function of the ultra-high field magnetic resonance imaging reconstruction network is as follows:

[0046] The adversarial loss is as follows:

[0047] L cGAN (G,D)=E (x,y) [logD(x,y)]+E (x) [log(1-D(x,G(x)))]

[0048] Among them, x is the input image, i.e., 3T T1-weighted MRI and 3T T2-weighted MRI, y is the corresponding ground truth image, i.e., the real 7T MRI, G is the generator, and D is the discriminator.

[0049] The image matching loss is as follows:

[0050] L im (G) = E (x,y) (‖yG(x)‖1)

[0051] The feature matching loss is as follows:

[0052]

[0053] Among them, D i is the feature map of the i-th layer, T is the total number of discriminator layers, N i is the element sum of the feature map of the i-th layer; the ROI loss is as follows:

[0054]

[0055] N is the total number of voxels;

[0056] Objective function:

[0057]

[0058] Among them, μ, λ and γ are hyperparameters that control the importance of feature matching loss, ROI loss and image matching loss. L2 norm regularization is introduced in the generator and discriminator to prevent overfitting.

[0059] In this embodiment, if Figure 3 As shown in the figure, the self-attention module framework converts the feature map x of the middle hidden layer of the network into two feature spaces f(x) = W f x and g(x) = W g x, used to calculate attention. Specifically includes:

[0060] Step 1: Convert the feature map x of the middle hidden layer of the network into two feature spaces. The calculation formula is:

[0061] f(x)=W f x

[0062] g(x)=W g x

[0063] Where f(x) represents the feature space obtained after feature extraction of the input image; W f represents the learning weight matrix, which is represented by a 1×1×1 convolution; g(x) represents the feature space obtained after feature extraction of the input image; W g Represents the learning weight matrix, represented by a 1×1×1 convolution;

[0064] Step 2: Calculate the attention in the feature space. The calculation formula is:

[0065]

[0066] S j,i =f(x j ) T g(x i )

[0067] Among them, β j,i represents the attention of voxel j to voxel i, and N represents the total number of voxels in the input image;

[0068] Step 3: Calculate the third feature space h(x) = W h x∈R C×(h×w×d) , where W h express,

[0069] R C×(h×w×d) express;

[0070] Step 4: Calculate the attention feature map O of the third feature space j , and its calculation expression is:

[0071]

[0072] h(x i )=W h x,v(x i )=W v x

[0073] Where v represents β j,i After matrix multiplication with the third feature space h(x), information is extracted from the fused features; β j,i represents the attention of voxel j to voxel i; W v Represents the learning weight matrix, represented by a 1×1×1 convolution;

[0074] Step 5: The attention feature map finally output by the attention calculation module based on the three feature spaces is expressed as:

[0075] y j=αO j +x j

[0076] Among them, α is a learnable scalar initialized to 0.

[0077] In this embodiment, the hippocampal subregion segmentation network framework is shown in FIG. Figure 4 As shown in the figure, the process of processing the input data by the network structure includes: the network is a symmetrical 3D-Unet structure, which includes 10 three-dimensional convolutional layers, 4 three-dimensional deconvolutional layers and 4 feature splicing layers; each three-dimensional convolutional layer before the fifth three-dimensional convolutional layer passes through batch normalization, ReLU activation function, and 3D MaxPooling operation; the fifth three-dimensional convolutional layer is input into the first three-dimensional deconvolutional layer through batch normalization and ReLU function; each three-dimensional deconvolutional layer passes through the feature splicing layer, three-dimensional convolutional layer, batch normalization and ReLU function in turn; each three-dimensional deconvolutional layer and the three-dimensional convolutional layer with the same number of feature maps before the 3D MaxPooling operation and through batch normalization and ReLU function are channel spliced ​​to obtain the feature splicing layer; a multi-scale deep supervision module is introduced in each three-dimensional deconvolutional layer. The input of the segmenter is the result of splicing the output of the generator and the input channel of the generator. The top output of the segmenter passes through a three-dimensional convolutional layer with a convolution kernel size of 3 and an output feature map number of 8, and finally outputs the final hippocampal sub-area segmentation probability map through softmax.

[0078] The segmentation loss function used by the segmenter is the anatomical weighted cross entropy loss designed by the present invention based on anatomical prior knowledge. In order to alleviate the problem of class imbalance during training, class weights are added to the cross entropy loss, and higher weights are set for classes that need attention and are more difficult to segment. The weighted cross entropy loss formula is as follows:

[0079] L WCE =-w c ∑Y gt log Y pred

[0080] Among them, Y gt is the segmentation criterion, Y pred is the predicted probability map, w c is the class weight. The expression of class weight is:

[0081]

[0082] Among them, PA is the probability map, σ1 and σ2 are hyperparameters, and c∈C is the segmentation category.

[0083] The anatomically weighted cross entropy loss is:

[0084]

[0085] Among them, h, w, d are the height, width, and depth of the input image.

[0086] In this embodiment, the multi-scale deep supervision mechanism specifically includes: first, upsampling the output of the three middle layers of the extended path in the segmentation network to the top layer output size, and the number of output feature maps is the number of categories. Then, the side outputs of different scales are passed through the softmax layer to obtain the segmentation probability map. The segmentation loss of the output layer of different scales is calculated:

[0087]

[0088] Among them, L i is the segmentation loss of the i-th side output layer, and N is the number of side output layers.

[0089] The objective function of the hippocampal subregion segmentation network is:

[0090] L seg =L mds +L AWC +v S L2(s)

[0091] Among them, v S It is to control the weight and add L2 norm regularization to the segmentation network to prevent overfitting.

[0092] The objective function of the joint network Syn_SegNet is as follows:

[0093] L syn_seg =L seg +L syn

[0094] Another specific embodiment of the present invention is to perform a hippocampal subregion segmentation experiment on a private dataset 3T-7T MRI paired dataset. There are 7 segmentation categories, namely SUB, CA2, CA1, CA4-DG, ERC, CA3, and Tail. The experimental environment of this experiment: TensorFlow 1.15 is used, based on Ubuntu server and NVIDIA RTX TITAN gpu.

[0095] In order to determine the optimal MRI modality generation for hippocampal subregion segmentation, the present invention uses left hippocampal data for experiments. By generating images of three different modalities, they are sent to the segmentation network through channel splicing with the original 3T MRI input, and the obtained segmentation results are compared for accuracy by Dice similarity coefficient (DSC), Hausdorff coefficient (HD95) and Jaccard coefficient (JC). It can be found that the generated 7T T1 MRI is used for subsequent hippocampal subregion segmentation. Compared with the results obtained by generating the other two modal images, the results not only obtain the best results in the segmentation of most subregions, but also the best in the average values ​​of the three evaluation indicators. Therefore, 7T T1 MRI is used in subsequent experiments.

[0096] Figure 5 The generated 7T MRI for the left hippocampus and the corresponding real 7T MRI results are shown in the figure. The ultra-high field magnetic resonance imaging reconstruction network in the joint network model is used to generate high-quality 7T-like images, which are then used in the subsequent hippocampal subregion segmentation task to supplement the texture details of the hippocampal region missing in conventional MRI, thereby improving the accuracy of MRI hippocampal subregion segmentation. The example of generating a 7T-like image is shown in the figure below. Figure 5 As shown, it can be seen that the 7T T1 and 7T T2 MRI obtained by using the ultra-high field magnetic resonance image reconstruction network proposed in the present invention are compared with their corresponding real 7T T1 and 7T T2 MRI. The generated images can restore most of the anatomical texture details and have higher image quality.

[0097] In the present embodiment, Dice similarity coefficient (DSC), Hausdorff coefficient (HD95) and Jaccard coefficient (JC) are used to compare the hippocampal subregion segmentation accuracy of different methods on left and right hippocampal data. Wherein, 3D U-Net and ResDUnet are existing excellent hippocampal subregion segmentation methods, HS_SegNet is a joint network that removes ultra-high field magnetic resonance imaging reconstruction network, and the hippocampal subregion segmentation network of the present invention is used alone to segment, Syn+Seg_Net is to segment ultra-high field magnetic resonance imaging reconstruction network and hippocampal subregion segmentation network as two independent tasks in an offline manner, and Syn_SegNet is the end-to-end segmentation network of the multi-task joint proposed by the present invention. According to the above-mentioned segmentation result, it can be known that whether it is left hippocampus or right hippocampus, the segmentation accuracy of the method proposed by the present invention is better than the remaining four segmentation methods. Comparing the results of HS_SegNet and Syn_SegNet, it can be seen that the present invention introduces the advancedness of ultra-high field magnetic resonance imaging reconstruction in hippocampal subregion segmentation, and the generated high-quality 7T images are promoted in an end-to-end online manner to promote the hippocampal subregion segmentation of conventional MRI, achieving a significant improvement effect. Comparing the results of Syn+Seg_Net and Syn_SegNet, it can be found that the multi-task joint end-to-end online method used in the present invention is better than the offline method for segmentation.

[0098] Figure 6 The following is a comparison of the segmentation results of the left hippocampal subregions by the embodiment of the present invention and the segmentation visualization results of other methods. Among them, arrow a indicates segmentation fragments, arrow b indicates over-segmentation of subregion Tail, arrow c indicates discontinuity of subregion CA3, arrow d indicates segmentation error of subregion CA3, and arrow e indicates discontinuity of subregion SUB. Figure 6 As shown, it can be seen that the joint network proposed in the present invention obtains the best visualization result, 3D U-Net and Syn+SegNet both have segmentation fragments, ResDUnet has obvious CA3 segmentation errors, and HS_SegNet and Syn+SegNet both have obvious SUB discontinuities.

[0099] Figure 7This is a comparison chart of the segmentation results of the right hippocampal subregions of the embodiment of the present invention and the segmentation visualization results of other methods. Among them, arrow a indicates segmentation holes, arrows b and d indicate discontinuity in the segmentation of subregion CA3, arrow c indicates over-segmentation of subregion ERC, arrow e indicates segmentation error, and arrow f indicates insufficient segmentation. It can be seen that the joint network proposed in the present invention obtains the best visualization results, 3D U-Net has segmentation holes, 3D U-Net, ResDUnet, HS_SegNet and Syn+SegNet all have discontinuity in CA3 segmentation, and Syn+SegNet has segmentation errors.

[0100] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A hippocampal subregion segmentation method based on ultra-high field magnetic resonance imaging reconstruction, characterized in that: include: Acquire a brain magnetic resonance image to be segmented, preprocess the brain magnetic resonance image, and input the preprocessed brain magnetic resonance image into a joint network model to obtain a segmentation result; wherein the joint network model includes an ultra-high field magnetic resonance image reconstruction network and a hippocampal subregion segmentation network; the ultra-high field magnetic resonance image reconstruction network includes a generation network and a discrimination network; The process of training the joint network model includes: S1: Obtain a magnetic resonance imaging dataset and preprocess the dataset; divide the preprocessed dataset into a training set and a validation set; S2: The images in the training set are input into the ultra-high field magnetic resonance image reconstruction network, and the deep and shallow features of the input image are extracted using the generative network, and the deep and shallow features are fused and reconstructed to obtain the initial reconstructed MRI; the generative network structure includes a symmetrical 6-layer encoder-decoder structure, which includes 6 3D convolutional layers, 6 3D deconvolutional layers, and 5 feature splicing layers; S3: Input the initial reconstructed MRI and the conventional MRI in the training set into the discriminator network for comparison with the real MRI. If the output is false, return to step S2. If the output is true, output the reconstructed MRI. The discriminant network is a PatchGAN network, which includes a fully convolutional network with 5 layers of 3D convolutional layers. S4: The reconstructed MRI and conventional MRI are input into the hippocampal subregion segmentation network to extract multi-scale feature information and obtain the hippocampal subregion segmentation probability map; the hippocampal subregion segmentation network is a symmetrical 3D-Unet structure, including 10 three-dimensional convolutional layers, 4 three-dimensional deconvolutional layers and 4 feature splicing layers; S5: Obtain the segmentation truth images corresponding to the reconstructed MRI and conventional MRI according to the hippocampal subregion segmentation probability map, set the initial category average dice to 0, and use the validation set data to verify the model performance after each epoch, and calculate the category average dice. If it is higher than the category average dice value of the previous epoch, save the model, otherwise continue training; S6: Calculate the loss function of the joint network model, continuously adjust the model parameters, and complete the model training when the loss function converges.

2. The method for hippocampal subregion segmentation based on ultra-high field magnetic resonance imaging reconstruction according to claim 1, characterized in that: The process of preprocessing brain MRI images includes: denoising and bias field correction for all MRI data, aligning 3T T2 and 7T T2 MRI with the corresponding 3T T1 and 7T T1 MRI, linearly registering 7T T1 MRI to 3T T1 MRI, and linearly registering 3TT1 MRI to an unbiased template using ANTS deformable registration; extracting two cubic ROIs containing the left and right hippocampi from the unbiased template; for each side, mapping the resulting binary segmentation into the template space by merging the anatomical labels of that side into a single label, and extracting a rectangular box covering all hippocampal regions in the template space to obtain a cubic ROI; upsampling the cubic ROIs of the left and right templates; mapping MRI data and manual labels of all modalities to the corresponding hemispheric template ROI to obtain the preprocessed brain MRI images.

3. The method for hippocampal subregion segmentation based on ultra-high field magnetic resonance imaging reconstruction according to claim 1, characterized in that: Generating a network structure to process data includes: each three-dimensional convolution layer passes through batch normalization and ReLU activation function; each three-dimensional deconvolution layer passes through a splicing layer, batch normalization and ReLU activation function, and the first three three-dimensional deconvolution layers pass through a Dropout operation after the splicing layer, batch normalization and ReLU activation function; the convolution kernel sizes of the first two three-dimensional convolution layers and the last two three-dimensional deconvolution layers are both 5, and the remaining convolution kernel sizes are all 3; each three-dimensional deconvolution layer, after passing through batch normalization, ReLU activation function or batch normalization, ReLU activation function, and Dropout operation, is spliced ​​with convolution layers with the same number of feature maps and passing through batch normalization and ReLU activation function to obtain the feature splicing layer; each three-dimensional convolution layer and three-dimensional deconvolution layer uses spectral normalization; the fourth three-dimensional convolution layer and the fourth three-dimensional deconvolution layer pass through a three-dimensional self-attention module after passing through batch normalization and ReLU activation function, and the image generated by the generator contains real anatomical details.

4. The method for hippocampal subregion segmentation based on ultra-high field magnetic resonance imaging reconstruction according to claim 3, characterized in that: The process of processing the feature map by the 3D self-attention module includes: Step 1: Convert the feature map x of the middle hidden layer of the network into two feature spaces. The calculation formula is: f(x)=W f x g(x)=W g x Among them, f(x) represents the feature space obtained after feature extraction of the input image, W f represents the learning weight matrix, g(x) represents the feature space obtained after feature extraction of the input image, and W g represents the learning weight matrix; Step 2: Calculate the attention in the feature space. The calculation formula is: S j,i =f(x j ) T g(x i ) Among them, β j,i represents the attention of voxel j to voxel i, and N represents the total number of voxels in the input image; Step 3: Calculate the third feature space h(x) = W h x∈R C×(h×w×d) ; Step 4: Calculate the attention feature map O of the third feature space j , and its calculation expression is: h(x i )=W h x,v(x i )=W v x Where v represents β j,i After matrix multiplication with the third feature space h(x), information is extracted from the fused features. j,i represents the attention of voxel j to voxel i, W v represents the learning weight matrix; Step 5: The attention feature map finally output by the attention calculation module based on the three feature spaces is expressed as: the j =αO j +x j Among them, α is a learnable scalar initialized to 0.

5. The method for hippocampal subregion segmentation based on ultra-high field magnetic resonance imaging reconstruction according to claim 1, characterized in that: The discriminative network processes the data flow including each 3D convolution layer passing through batch normalization, ReLU activation function and Dropout operation; the third 3D convolution layer passes through a 3D self-attention module after passing through batch normalization, ReLU activation function and Dropout operation; each 3D convolution layer uses spectral normalization.

6. The method for hippocampal subregion segmentation based on ultra-high field magnetic resonance imaging reconstruction according to claim 1, characterized in that: Each 3D convolution layer before the fifth 3D convolution layer in the hippocampal subregion segmentation network passes through batch normalization, ReLU activation function, and 3D MaxPooling operation; the fifth 3D convolution layer is input into the first 3D deconvolution layer through batch normalization and ReLU function; each 3D deconvolution layer passes through the feature splicing layer, 3D convolution layer, batch normalization, and ReLU function in turn; each 3D deconvolution layer is channel-spliced ​​with the 3D convolution layer with the same number of feature maps before the 3D MaxPooling operation and through batch normalization and ReLU function to obtain the feature splicing layer; A multi-scale deep supervision module is introduced in each 3D deconvolution layer.

7. The method for hippocampal subregion segmentation based on ultra-high field magnetic resonance imaging reconstruction according to claim 1, characterized in that: The process of extracting multi-scale feature information from reconstructed MRI and conventional MRI using the hippocampal subregion segmentation network includes: upsampling the outputs of the three middle layers of the extended path in the segmentation network to the top layer output size, and the number of output feature maps is the number of categories; the side outputs of different scales are passed through the softmax layer to obtain the segmentation probability map.

8. The method for hippocampal subregion segmentation based on ultra-high field magnetic resonance imaging reconstruction according to claim 1, characterized in that: The loss functions of the joint network model include the loss function of the ultra-high field magnetic resonance image reconstruction network and the loss function of the hippocampal subregion segmentation network; the loss function of the ultra-high field magnetic resonance image reconstruction network includes adversarial loss, feature matching loss, and ROI loss; the loss function of the hippocampal subregion segmentation network is the anatomical weighted cross entropy loss based on anatomical prior knowledge.

9. A hippocampal subregion segmentation system based on ultra-high field magnetic resonance image reconstruction, the system is used to execute any one of the hippocampal subregion segmentation methods based on ultra-high field magnetic resonance image reconstruction according to claims 1 to 8, characterized in that: include: Joint network model training module, test set magnetic resonance image receiving module, test set magnetic resonance image segmentation processing module and segmentation result output module; The joint network model training module includes a training image receiving unit, a training image analysis and processing unit, an ultra-high field magnetic resonance imaging synthesis unit, and a hippocampal subregion segmentation unit; The training image receiving unit is used to receive a training image; The training image analysis and processing unit is used to perform feature analysis on the training image and fuse the feature information of the image; The ultra-high field magnetic resonance image synthesis unit reconstructs the ultra-high field MRI image by using an adversarial learning method according to the fused feature information; The hippocampal subregion segmentation unit is used to perform feature analysis on the training image and the synthesized ultra-high field MRI image, fuse effective feature information, and perform segmentation processing on the fused feature information until the loss functions of the ultra-high field MRI image synthesis unit and the hippocampal subregion segmentation unit in the joint network model converge to a preset threshold, thereby completing the deep learning training; The test set magnetic resonance image receiving module is used to receive magnetic resonance image data information to be segmented; The test set magnetic resonance image segmentation processing module is used to perform segmentation processing on the magnetic resonance image data information to be segmented received by the magnetic resonance image receiving module (2); The segmentation result output module is used to output the hippocampal subregion segmentation result.

10. The hippocampal subregion segmentation system based on ultra-high field magnetic resonance imaging reconstruction according to claim 9, characterized in that: The test set magnetic resonance image segmentation processing module includes an ultra-high field MRI image synthesis unit, a feature decomposition unit, a feature processing unit, a feature fusion unit, and a classification unit; The ultra-high field MRI image synthesis unit is used to perform feature analysis on the test set magnetic resonance image data information, perform multiple convolution and deconvolution processing on the data information under different modalities, obtain feature information of each sub-block, and fuse the feature information to obtain a synthesized ultra-high field MRI image; The feature decomposition unit is used to perform multi-scale decomposition on the test image and the synthesized ultra-high field MRI image data information to obtain multiple sub-blocks under different modes; The feature processing unit is used to perform multiple convolution and deconvolution processes on each sub-block to obtain feature information of each sub-block; The feature fusion unit is used to fuse the feature information of all sub-blocks to obtain fused features; The classification unit is used to classify the fused features through a classifier to obtain a hippocampal subregion segmentation result.

Images