A method and system for improving the segmentation accuracy of smaller categories of brain structures in whole-brain structure segmentation.

Abstract

This invention provides a method and system for improving the segmentation accuracy of smaller categories of brain structures in whole-brain structure segmentation. The method involves inputting acquired 3D brain images into a trained FCN brain classification network and a trained MAD brain classification network for segmentation, respectively. The segmentation results from the two networks are then fused to improve the segmentation accuracy of smaller categories of brain structures in whole-brain structure segmentation based on the FCN brain classification network. This invention considers the balance between whole-brain data input segmentation and GPU memory usage during FCN network training for whole-brain segmentation. FCN networks typically perform convolutional downsampling on the input images, leading to information loss and affecting the recognition of smaller categories of brain structures. Therefore, this invention proposes a method for reclassifying smaller categories of brain structures. Finally, the results are merged with the original FCN network results to achieve high-precision segmentation of each brain structure.

Description

Technical Field

[0001] This invention relates to medical image processing technology, specifically to a method and system for improving the segmentation accuracy of smaller categories of brain structures in whole-brain structure segmentation. Background Technology

[0002] Magnetic resonance imaging (MRI) of the brain provides a rapid, non-invasive, safe, and painless method for acquiring high-contrast anatomical information of the human brain, and is therefore widely used in various neuroimaging studies and medical practices. Neuroanatomical segmentation is a fundamental task in MRI analysis. In MRI analysis, brain images are classified into voxels by assigning a semantic label to each voxel representing a specific neuroanatomical structure, thus causing voxels belonging to the same category to exhibit similar features. As a fundamental tool allowing quantitative measurements of neuroanatomical volume, thickness, and shape from MRI, brain segmentation is of great significance for clinical applications, including brain development monitoring, brain morphology analysis, deep brain stimulation, and preoperative assessment. Typically, structural brain segmentation focuses on one or more brain structures. In contrast, whole-brain segmentation targets more than dozens of regions, making it more challenging. Detailed whole-brain segmentation is particularly useful in sophisticated quantitative brain studies, such as structural brain network analysis and brain connectivity analysis.

[0003] In practice, manual segmentation of brain MRI according to some brain labeling protocols is a gold standard, but it is difficult to obtain from individuals with expert-level medical knowledge, and reproducibility is poor due to operator variability. Therefore, a precise, fully automated whole-brain segmentation method is highly desirable and has attracted the interest of the research community for a long time.

[0004] Deep learning methods using convolutional neural networks (CNNs) have demonstrated state-of-the-art performance in computer vision tasks. In recent years, an increasing number of CNNs / FCNs have been proposed for whole-brain segmentation because they can simultaneously learn task-specific features and classifiers. Most existing research focuses on automatic brain segmentation methods based on patches, slices, or subvolumes. A significant drawback of these methods is that they only utilize partial information from each data instance during training and inference. Therefore, they are prone to underfitting the data, given that only a subset of available intensity information is provided for training. Furthermore, a handcrafted fusion strategy is needed to aggregate partial predictions. The aforementioned methods are based on partial volumes primarily due to the difficulty of training 3D neural networks and the large volume of brain MRI data. More specifically, on typical graphics processing unit (GPU) platforms, CNNs / FCNs become too shallow to be usable if trained using the entire whole-brain volume without pruning, and generally exhibit inferior performance compared to multi-map methods. Moreover, most proposed CNNs / FCNs have been validated on brain segmentation tasks with only a few brain structures. Only a few deep learning-based methods have been evaluated for their whole-brain segmentation performance on datasets with dozens of anatomical labels.

[0005] Full-volume training schemes using complete whole-brain MRI data as input have attractive characteristics. First, no usable information in the input data is lost during model training and inference. Second, the full-volume approach eliminates the need to design complex segmentation and aggregation schemes for model training and inference. Third, due to the single typical neural network forward pass, the inference time of the input (excluding preprocessing) is approximately milliseconds. Fourth, FCN allows for inputs without hard shape constraints, enabling more seamless integration of advanced FCN techniques. Currently, the VoxHRNet method is one of the high-performance whole-brain segmentation methods using FCN convolutional neural networks (Yeshu Li, Jonathan Cui, Yilun Sheng, Xiao Liang, Jingdong Wang, Eric I.-Chao Chang, Yan Xu, Whole brain segmentation with full volume neural network, Computerized Medical Imaging and Graphics, 2021, 93:101991).

[0006] Despite these advantages, the full-scale training framework presents three major challenges. First, the large volume of brain MRI data makes deploying a standard 3D FCN architecture on current GPUs prohibitively expensive. Second, the FCN architecture itself needs to effectively segment fine-grained brain structures and compress limited GPU resources. Because of the need to balance segmentation of whole-brain MRI data input with GPU memory usage, FCN networks typically perform convolutional downsampling on the input images, leading to information loss, particularly affecting the recognition of small-category brain regions. Therefore, addressing the limitations of existing whole-brain data input and segmenting brain functional regions will further improve the overall segmentation accuracy of images. Summary of the Invention

[0007] To address the problems of existing technologies, this invention proposes a method that selects poorly segmented brain regions from the original FCN network training results as objects to be segmented by another neural network. By extracting image features through a feature encoder-decoder structure, the network can focus on different features based on different visual domains. By using a dilated convolutional module in the last layer of the feature encoder-decoder network to expand the receptive field and capture multi-scale contextual information, the model can better extract semantic features. Encoding and decoding attention modules increase the weight of useful features, allowing the network to focus more on useful features and suppress background. Combining these strategies with the neural network, a MAD model is proposed, which improves the accuracy of segmenting poorly segmented brain regions in the original FCN network. Finally, the results of the original FCN training and the MAD model training are fused to achieve accurate segmentation of each brain region.

[0008] The specific technical solution adopted in this invention is as follows:

[0009] A method to improve the segmentation accuracy of smaller categories of brain structures in whole-brain structure segmentation, specifically:

[0010] The acquired 3D brain images are input into a trained FCN brain classification network and a trained MAD brain classification network for segmentation, respectively. The segmentation results of the two networks are then fused to improve the segmentation accuracy of smaller brain structure categories in the whole-brain structure segmentation based on the FCN brain classification network. The trained MAD brain classification network is obtained by training a first training dataset, where each sample in the first training dataset is a 3D brain image. The label images of the 3D brain images only indicate the categories whose segmentation accuracy in the trained FCN brain classification network is below a threshold.

[0011] A MAD brain classification network is constructed, which uses U-net as the backbone network. The feature encoder of U-net consists of multiple sequentially connected encoding convolutional modules, a dilated convolutional module, a convolutional block, and an encoding attention mechanism module. Each encoding convolutional module includes multiple sequentially connected cascaded convolutional blocks and an encoding attention mechanism module. The multiple encoding convolutional modules are downsampled sequentially. The feature decoder of U-net includes multiple sequentially connected decoding convolutional modules. Each decoding convolutional module includes a sequentially connected decoding attention mechanism module AG (Attention Gate) and multiple cascaded convolutional blocks.

[0012] Each sample in the first training dataset is used as the input to the MAD brain classification network. The training aims to minimize the loss between the segmentation result output by the MAD brain classification network and the label image, thus obtaining the trained MAD brain classification network.

[0013] Furthermore, dilated convolution is used in the lower layer of the U-net backbone network to extract rich semantic information. The encoding attention mechanism module is CBAM, which adds weights to the convolutional results of this layer to capture channel and global information. The decoding attention mechanism module AG fuses and weights the skip-connection and lower-level up-sampling results to guide the network to focus on areas that need intensive processing. The original image size and number of channels are restored at the top layer of the feature decoder, and then the result is output after passing through a fully connected layer.

[0014] Furthermore, the three-dimensional brain image is a three-dimensional brain image acquired using MRI.

[0015] Furthermore, the training dataset includes acquired 3D brain images and 3D brain images augmented based on the acquired 3D brain images. Data augmentation includes adding Gaussian noise, random distortion, and small-scale random displacement.

[0016] Furthermore, the labeled image of the three-dimensional brain image is obtained through the following method:

[0017] The second training dataset used to train the FCN brain classification network is input into the trained FCN brain classification network. The similarity between the output of the trained FCN brain classification network and each category in the label image of the second training dataset is calculated as the segmentation accuracy. Based on the set similarity threshold, the category pixels in the label image of the second training dataset with similarity higher than the similarity threshold are set to 0 and used as the label image corresponding to the first training dataset.

[0018] Furthermore, during training, a mixed-precision training method was employed to reduce GPU memory usage. In testing, the outputs of the two models were merged to improve the accuracy of the FCN network in segmenting small categories of brain regions, while ensuring that the original brain region segmentation structure remained unaffected.

[0019] A whole-brain structure segmentation system based on the above method includes:

[0020] The data acquisition module is used to acquire the three-dimensional brain image to be segmented;

[0021] The first segmentation module is used to input the acquired three-dimensional brain image into a trained FCN brain classification network for segmentation to obtain the first segmentation result;

[0022] The second segmentation module is used to input the acquired three-dimensional brain image into a trained MAD brain classification network for segmentation to obtain the second segmentation result;

[0023] The fusion module is used to fuse the first segmentation result and the second segmentation result to obtain the final segmentation result of the N-type brain region structure.

[0024] The beneficial effects of this invention are:

[0025] This invention addresses the balance between whole-brain data input segmentation and GPU memory usage during FCN network training for whole-brain segmentation. FCN networks typically perform convolutional downsampling on input images, leading to information loss and impacting the recognition of minor brain region structures. Therefore, this invention proposes a reclassification method for minor brain region structures. Specifically, this invention uses a separate neural network model structure to improve the suboptimal segmentation of minor brain region structures when using the entire 3D MRI image as input for brain functional region segmentation. First, the original FCN network is trained. Based on the network's training output, Dice coefficients for brain region structures are set, and structures below a threshold are selected as labels for the MAD model's training data. The MAD model only needs to classify these structures. Finally, the results are merged with those of the original FCN network to achieve high-precision segmentation of each brain region structure. Attached Figure Description

[0026] Figure 1 This is a framework diagram of a method for improving the performance of FCN networks in whole-brain segmentation for small-category brain regions;

[0027] Figure 2 This is a structural diagram of the MAD model used in this invention;

[0028] Figure 3 This is a structural diagram of the CBAM module in the MAD model;

[0029] Figure 4 Here is the structure diagram of the Channel Attention Model in the CBAM module;

[0030] Figure 5 The structure diagram of the Spatial Attention Model in the CBAM module.

[0031] Figure 6 This is a structural diagram of the AG module in the MAD model;

[0032] Figure 7 This invention provides a method for improving the performance of FCN networks in whole-brain segmentation, which is not good for small categories of brain region structures. It also compares the performance of VoxHRNet and manual segmentation methods in one embodiment. Detailed Implementation

[0033] The technical solutions of the embodiments of this application will be described in more detail below with reference to the above-mentioned accompanying drawings and embodiments.

[0034] This invention provides a method to improve the segmentation accuracy of smaller categories of brain structures in whole-brain structure segmentation. The method involves inputting acquired 3D brain images into a trained FCN brain classification network and a trained MAD brain classification network for segmentation, and then fusing the segmentation results from the two networks to improve the segmentation accuracy of smaller categories of brain structures in whole-brain structure segmentation based on the FCN brain classification network. This invention can be applied to the segmentation of various types of 3D brain images, including MRI-acquired 3D brain images. This embodiment uses MRI as an example. Figure 1 As shown, this method consists of two parts: the construction and training of the FCN brain classification network and the MAD brain classification network, and the actual application of whole-brain structure segmentation. The specific implementation steps are as follows:

[0035] I. Construction and Training of FCN Brain Classification Network and MAD Brain Classification Network

[0036] Step 1: Obtain the second training dataset. Each sample in the second training dataset consists of a 3D MRI brain image and its corresponding label image. Generally, because background data occupies a large portion of the original data, affecting GPU memory usage during model training, background cropping is performed on the directly acquired 3D MRI brain images: the regions occupied by the brain in the coronal, sagittal, and transverse sections of all image data are selected, and the size of these regions is adjusted to a multiple of 16. The background is then removed from the coronal, sagittal, and transverse sections based on the results calculated for each section, resulting in smaller volumes of MRI brain image data. Additionally, to address the issue of uneven intensity within the image tissue caused by uneven magnetic field distribution during MRI brain image scanning, tools such as the N4 offset correction function in ITK are used to perform offset field correction on the MRI brain image data.

[0037] Furthermore, as a preferred embodiment, the second training dataset includes, in addition to the acquired MRI brain images, MRI brain images augmented based on the acquired MRI brain images to prevent network overfitting and improve the model's generalization ability. Specifically, data augmentation includes: adding Gaussian noise with a mean of 0 and a variance of 1, performing non-rigid distortion on the images, and randomly shifting the images up, down, left, and right by n pixels (0 <= n <= 8) from their original positions.

[0038] Step 2, Construction and training of FCN brain classification network: Input each sample of the second training dataset obtained in Step 1 into FCN, and train it with the goal of minimizing the loss between the segmentation result output by FCN brain classification network and the label image to obtain the trained FCN brain classification network.

[0039] Step 3: Obtain the first training dataset. In this embodiment, the first training dataset is the same as the aforementioned second training dataset sample, but the labeled images of the first training dataset only label the categories in the trained FCN brain classification network whose segmentation accuracy is lower than the threshold.

[0040] The segmentation accuracy can be evaluated using similarity. In this embodiment, the Dice similarity coefficient is used. The second training dataset is input into the trained FCN brain classification network. The Dice similarity coefficient between the output of the trained FCN brain classification network and each category in the label image of the second training dataset is calculated. A Dice coefficient threshold is set, and the categories with Dice coefficients higher than the threshold for brain region structure segmentation are selected. The pixels of the categories with similarity higher than the similarity threshold in the label image of the second training dataset are set to 0. The categories with similarity lower than the threshold are retained in the label image corresponding to the original data as the label images of the MAD model training data.

[0041] Step 4: Construct the MAD (Multiple Attention Dilated Model) brain classification network: The MAD brain classification network uses U-net as the backbone network. The U-net feature encoder consists of multiple sequentially connected encoding convolutional modules, one dilated convolutional module, one convolutional block, and one encoding attention mechanism module. Each encoding convolutional module includes multiple sequentially connected cascaded convolutional blocks and one encoding attention mechanism module. The U-net feature decoder includes multiple sequentially connected decoding convolutional modules. Each decoding convolutional module includes a sequentially connected decoding attention mechanism module AG (Attention Gate) and multiple cascaded convolutional blocks. The results of skip-connection and low-level up-sampling are fused and weighted to guide the network to focus on areas that need intensive processing. The original image size and number of channels are restored at the top layer of the feature decoder, and then the result is output after passing through a fully connected layer.

[0042] Taking four encoding convolutional modules and four decoding convolutional modules as an example, such as Figure 2 As shown, the input MRI brain image data first passes through a U-net sampling convolutional block, which contains 32 convolutional layers with a kernel size of 3*3*3, a Group Normalization (GN) layer, and a ReLU (rectified linear units) layer. Based on this, a feature encoder extracts features at different levels; the feature encoder consists of four encoding convolutional modules. The first encoding convolutional module of the feature encoder consists of two cascaded convolutional blocks and an encoding attention mechanism module. The cascaded convolutional blocks have 32 and 16 kernels respectively, forming 3*3*3 convolutional layers, GN layers, and ReLU layers. The encoding attention mechanism module is a CBAM (Convolution Block Attention Model) attention module, which includes channel attention and spatial attention mechanisms, such as... Figure 3 As shown; the implementation of channel attention is divided into two parts, as follows: Figure 4 As shown, for a single feature layer F input to the channel attention mechanism after passing through a convolutional block, global average pooling and global max pooling are performed respectively to obtain two feature bars, MaxPool and AvgPool, with a length equal to the number of channels in the feature layer. The results of average pooling and max pooling are then processed using two shared fully connected layers, and the two results are added together before passing through a sigmoid activation function. This yields the weights (between 0 and 1) for each channel of the input feature layer. Multiplying these weights by the original input feature layer F gives the output of the channel attention mechanism. The implementation of the spatial attention mechanism is as follows... Figure 5As shown, for the output feature layer F1 of the channel attention mechanism, the maximum value and average value are taken for each feature point's channel. These two results are then stacked, and the number of channels is adjusted using a single convolution with 1 channel. After passing through the sigmoid activation function, the weights (between 0 and 1) for each feature point in the input feature layer are obtained. These weights are then multiplied by the original input feature layer to obtain the final output of the attention module, denoted as C1. In the second, third, and fourth layers of the feature encoder's encoding convolutional modules, the number of kernels in the two concatenated convolutional blocks is twice the number of channels in the final output of the previous feature encoder. The remaining structure is the same as the first layer of the feature encoder's encoding convolutional module, and the final outputs of each layer are denoted as C2, C3, and C4, respectively. The output C4 of the last layer of the feature encoder is input into the dilated convolution module. The dilated convolution module consists of two cascaded convolutional blocks and an attention mechanism. Each convolutional block consists of a convolutional layer with a dilation rate of 2, a kernel number equal to the number of channels in C4, and a kernel size equal to 3*3*3, a GN layer, and a ReLU layer. The output of C4 after inputting into the dilated convolution module is denoted as C5. C5 and C4 are stacked and denoted as C6. At this time, the number of channels in C6 is twice the number of channels in C5 or C4. Subsequently, C6 is input into the bottom convolutional block and the encoding attention mechanism module. At this time, the convolutional module consists of only a single convolutional block with a kernel number equal to the number of channels in C6 divided by 2. The output result is denoted as C7. C7 is upsampled through trilinear interpolation to obtain the output G1. The output C3 from G1 and the feature encoder at the same layer is simultaneously input into the attention module AG of the third-layer feature decoder. The attention module AG is implemented as follows: The inputs C3 and G1 are added, passed through a ReLU layer, then through a 1x1x1 convolutional layer with one kernel and a sigmoid activation function to obtain an attention weight map. This attention weight map is multiplied by C3 to obtain C3', which suppresses irrelevant regions and focuses attention on the target region, thus allowing deep features to supervise shallow features. The output C3' obtained from the attention module AG is stacked with G1, and then passed through two cascaded convolutional blocks in the third layer of the feature decoder. The number of kernels in these two cascaded convolutional blocks is the same as the number of channels in G1, denoted as x1 and x1 / 2. The number of kernels in the two cascaded convolutional blocks in the second and second layers of the feature decoder is the same as the number of channels in the final output of the previous layer, denoted as x2 and x2 / 2. The remaining structure is the same as the third layer structure of the feature decoder. Finally, the network ends with a fully connected layer for output, and the final output 3D image has n channels (where n is the number of categories that need to be classified).

[0043] Step 5, Training the MAD brain classification network model: Build the MAD model as in Step 4, input the preprocessed data into the MAD model as in Steps 2-3, use mixed precision training to reduce GPU memory usage, and set the initial learning rate to 10. -3 The Adam optimizer is used to continuously update the parameter weights of the entire network through backpropagation. Finally, a trained model for brain regions segmented below a threshold is obtained and saved.

[0044] II. Practical Applications of Whole Brain Structural Segmentation:

[0045] The acquired MRI brain images were input into the trained FCN brain classification network and the trained MAD brain classification network for segmentation, respectively. This yielded N-channel probability maps (where N is the total number of brain region structure categories to be segmented) for all brain region categories, and n-channel probability maps (where n is the number of brain region structure categories below a threshold) for brain region structures below the threshold. The channel probability maps for brain region categories below the threshold from the probability maps for all brain region categories were then fused with the probability maps for brain region structures below the threshold to obtain the final segmentation result for the N types of brain region structures.

[0046] Test data were input into the models saved after training the original FCN model and the models saved after training the MAD model, respectively, and the Dice similarity coefficient was used to evaluate the segmentation results of brain region structures.

[0047]

[0048] In the above formula, X represents the segmentation result predicted by the model, and Y represents the artificial segmentation label for the corresponding brain region category. According to step 1, a total of n = 10 brain region structure categories were selected. Here, three test samples were chosen as validation data to evaluate the method proposed in this invention. In the final results, compared with the VoxHRNet segmentation results, the selected brain region structures showed a significant improvement in Dice coefficients: the left orbitofrontal gyrus increased from 0.7473 to 0.7614, an increase of 1.89%; the right orbitofrontal gyrus increased from 0.7912 to 0.7973, an increase of 0.77%; the right orbitolateral gyrus increased from 0.6991 to 0.7190, an increase of 2.85%; the left supramarginal gyrus increased from 0.7812 to 0.8061, an increase of 3.19%; and the left... The accuracy of the angular gyrus improved from 0.7496 to 0.7527, an increase of 0.41%; the left precuneus improved from 0.7679 to 0.7828, an increase of 1.89%; the right precuneus improved from 0.7881 to 0.7906, an increase of 0.32%; the right suboccipital gyrus improved from 0.7980 to 0.8058, an increase of 0.98%; the left cuneus improved from 0.7820 to 0.8075, an increase of 3.26%; and the right cuneus improved from 0.7716 to 0.7916, an increase of 2.59%. The results show that the method proposed in this invention for improving the segmentation accuracy of smaller brain structure categories in whole-brain structure segmentation significantly improves the poor segmentation results of the original FCN network for smaller brain structure categories in MR brain image data.

[0049] Corresponding to the aforementioned methods for improving the segmentation accuracy of smaller categories of brain structures in whole-brain structure segmentation, the present invention also provides embodiments of a whole-brain structure segmentation system.

[0050] The whole-brain structure segmentation system provided in this embodiment of the invention includes:

[0051] The data acquisition module is used to acquire the three-dimensional brain image to be segmented;

[0052] The first segmentation module is used to input the acquired three-dimensional brain image into a trained FCN brain classification network for segmentation to obtain the first segmentation result;

[0053] The second segmentation module is used to input the acquired three-dimensional brain image into a trained MAD brain classification network for segmentation to obtain the second segmentation result;

[0054] The fusion module is used to fuse the first segmentation result and the second segmentation result to obtain the final segmentation result of the N-type brain region structure.

[0055] The embodiments of the whole-brain structure segmentation system of the present invention can be applied to any device with data processing capabilities, such as a computer or other equipment or apparatus.

[0056] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0057] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for improving the segmentation accuracy of smaller classes of brain structures in whole brain structural segmentation, characterized in that, Specifically: The obtained three-dimensional brain image is input into a trained FCN brain classification network and a trained MAD brain classification network for segmentation, and the segmentation results of the two networks are fused to improve the segmentation accuracy of small class brain structures in the whole brain structure segmentation based on the FCN brain classification network; wherein the trained MAD brain classification network is obtained by the following method: Obtain a first training data set, each sample in the first training data set is a three-dimensional brain image, and the label image of the three-dimensional brain image only labels the class whose segmentation accuracy is lower than the threshold in the trained FCN brain classification network; the label image of the three-dimensional brain image is obtained by the following method: The second training data set for training the FCN brain classification network is input into the trained FCN brain classification network, the similarity between the output of the trained FCN brain classification network and each class in the label image of the second training data set is calculated as the segmentation accuracy, and based on the set similarity threshold, the class pixels in the label image of the second training data set with a similarity higher than the similarity threshold are set to 0 as the corresponding label image of the first training data set; The MAD brain classification network is constructed, the MAD brain classification network adopts U-net as the backbone network, the feature encoder of U-net is composed of a plurality of encoding convolution modules connected in turn, a hole convolution module, a convolution block and an encoding attention mechanism module, wherein each encoding convolution module includes a plurality of convolution blocks connected in series and an encoding attention mechanism module; the feature decoder of U-net includes a plurality of decoding convolution modules connected in turn, and each decoding convolution module includes a decoding attention mechanism module and a plurality of convolution blocks connected in series; Each sample in the first training data set is input into the MAD brain classification network as the input, and the loss between the segmentation result output by the MAD brain classification network and the label image is minimized as the target for training, to obtain the trained MAD brain classification network.

2. The method of claim 1, wherein, The three-dimensional brain image is a three-dimensional brain image collected by MRI.

3. The method of claim 1, wherein, The training data set includes the collected three-dimensional brain image and the three-dimensional brain image obtained by data enhancement based on the collected three-dimensional brain image.

4. The method of claim 1, wherein, During training, the method of mixed precision training is used for training.

5. A system for whole brain structure segmentation based on the method of any of claims 1-4, characterized in that, Comprise: The data acquisition module is used for acquiring the three-dimensional brain image to be segmented; The first segmentation module is used for inputting the obtained three-dimensional brain image into a trained FCN brain classification network for segmentation to obtain a first segmentation result; The second segmentation module is used for inputting the obtained three-dimensional brain image into a trained MAD brain classification network for segmentation to obtain a second segmentation result; The fusion module is used for fusing the first segmentation result and the second segmentation result to obtain the final segmentation result of N brain region structures.

Images