A brain region segmentation method based on deep learning

Abstract

The application relates to a brain region segmentation method based on deep learning, and relates to brain region segmentation. The steps are as follows: 1) obtaining the result segmented by a Destrieux brain atlas as a training label for model training and evaluation, and obtaining a training set of 122 brain region labels after merging and reducing the labels of part of left and right brain regions; 2) designing a deep learning network model with a multi-scale and multi-branch attention module based on an Unet network; 3) constructing a loss function of the network; 4) solving the optimal parameters of the deep learning network by using the training set obtained in step 1); and 5) inputting brain data to be segmented into the trained network for segmentation. By introducing a multi-scale segmentation attention module and a multi-branch cross attention module, the receptive field of different scales is increased, so that deep features are obtained, accurate segmentation of more brain regions of different scales is realized, the network structure can segment the whole brain into 161 brain regions, and the network structure has the characteristics of fast segmentation speed and high segmentation precision.

Description

Technical Field

[0001] This invention relates to brain region segmentation, and more particularly to a deep learning-based brain region segmentation method with multi-scale, multi-branch attention modules based on the Unet network. Background Technology

[0002] Brain region segmentation is a non-invasive quantitative analysis method for neuroimaging, which can be used to explore brain structure and predict neurological diseases. To date, there has been much research on brain region segmentation. The most accurate brain region segmentation method is manual segmentation by professional brain anatomists, considered the "gold standard," but this is extremely costly and difficult to complete in a short time for large-scale data. Therefore, quantitative brain analysis based on magnetic resonance imaging requires automated algorithms to achieve fast and accurate brain region segmentation.

[0003] With the increasing maturity of deep learning technology, deep convolutional neural networks (CNNs) have been applied to brain region segmentation. Based on the dataset type, networks can be categorized into 2D, 2.5D, and 3D networks. 2D CNNs cannot effectively utilize the spatial relationships between upper and lower slices of brain images, making it difficult to achieve very high brain region segmentation accuracy. 3D medical data is large in volume and limited by the current computational memory constraints of GPUs, making 3D networks difficult to train. Therefore, advanced 3D CNNs cannot be directly used for brain region segmentation. Instead, methods such as downsampling, patching, and block division are needed to reduce the computational load, but this also leads to a loss of segmentation accuracy and discontinuous segmented regions.

[0004] Based on the above analysis, 2.5D convolutional neural networks are more suitable for brain region segmentation. Currently available 2.5D networks include SD-Net, QuickNAT, FastSurferCNN, and MRF-Net. SD-Net (Khan N, Ullah A, Haq IU, et al. SD-Net: Understanding overcrowded scenes in real-time via an efficient dilated convolutional neural network [J]. Journal of Real-Time ImageProcessing, 2021, 18: 1729-1743.) uses a fully convolutional neural network segmentation model that can divide the brain into 27 regions. QuickNAT (Roy AG, Conjeti S, Navab N, et al. QuickNAT: A fully convolutional network for quick and accurate segmentation of neuroanatomy[J]. NeuroImage, 2019, 186: 713-727.) integrates information from three views of the brain (horizontal, coronal, and sagittal planes) on top of a convolutional neural network segmentation model, making brain region segmentation results more accurate. To segment more brain regions, FastSurferCNN (Henschel L, Conjeti S, Estrada S, et al. Fastsurfer-afast and accurate deep learning based neuroimaging pipeline [J]. NeuroImage, 2020, 219: 117012.) replaces ordinary dense blocks in QuickNAT with competitive dense blocks and skip connections with competitive connections, reducing the computational cost of the network and enabling the brain to be divided into 95 brain regions. MRF-Net (Wei C, Yang Y, Guo X, et al. MRF-Net: A multi-branch residual fusion network for fast and accurate whole-brain MRI segmentation [J]. Frontiers in Neuroscience, 2022, 16:940381.) is an improvement on FastSurferCNN, which can achieve fast segmentation of 136 brain regions.Currently, there are still problems such as insufficient precision in brain region segmentation, long segmentation time, and low clinical applicability. Summary of the Invention

[0005] The purpose of this invention is to address the problems of existing deep learning brain region segmentation algorithms, such as limited number of brain regions segmented, insufficient segmentation accuracy, slow segmentation speed, and limited clinical application value. It aims to provide a deep learning-based brain region segmentation method that offers faster segmentation speed and higher segmentation quality. The proposed deep learning-based brain region segmentation method, based on the Unet network and incorporating multi-scale, multi-branch attention modules, has significant advantages, greatly improving the number of brain regions segmented, segmentation accuracy, and speed, and showing great potential for disease prediction.

[0006] To achieve the above objectives, the present invention includes the following steps:

[0007] 1) Obtain the results of Destrieux brain map segmentation as training labels for model training and evaluation, and after merging some left and right brain regions to reduce the number of labels, obtain a training set of 122 brain region labels.

[0008] 2) Design a deep learning network model based on the Unet network with multi-scale and multi-branch attention modules;

[0009] 3) Construct the network's loss function;

[0010] 4) Using the training set obtained in step 1), solve for the optimal parameters of the deep learning network model with multi-scale and multi-branch attention modules based on the Unet network.

[0011] 5) Input the T1-weighted MRI brain data to be segmented into the trained network for segmentation.

[0012] In step 1), the specific method for obtaining the results of Destrieux brain map segmentation as training labels for model training and evaluation, and for merging some left and right brain regions to reduce the number of labels, to obtain a training set of 122 brain region labels, is as follows:

[0013] T1-weighted MRI image data were obtained. , Indicates the first T1-weighted MRI image data, Represents the real number field. These represent the length, width, and number of layers of the image in the data, respectively; using... The brain was segmented, resulting in a sample of 187 brain regions. , Indicates the first Layer image segmentation results; then, cortical regions that are in contact between the left and right hemispheres of the brain are merged into one brain region, and then all other regions are combined, and several less frequently used brain regions are removed, resulting in a result with 122 brain regions. , Indicates the first The result of reducing brain regions using layered images; finally, by , The training set is created and stored in an HDF5 file as input data for the network.

[0014] In step 2), the deep learning network model based on the Unet network with multi-scale, multi-branch attention modules is based on the Unet network and introduces two attention modules: multi-scale segmentation attention module (CSA) and multi-branch cross attention module (MCAM) to increase the receptive field at different scales. Seven modules with different compositions are connected through Maxout activation operations and skip paths to reduce the number of network training parameters. Specifically:

[0015] A. The multi-scale segmentation attention module (CSA) includes the following structure:

[0016] Input data After passing through a convolutional layer (Conv), the layers are divided into two groups, namely... , The size of the convolutional kernel of the convolutional layer is . and Each is obtained after passing through a convolutional layer. , The size of the convolutional kernel of the convolutional layer is . , After addition, the result is obtained by passing through a convolutional layer. The size of the convolutional kernel of the convolutional layer is . and The sum is obtained by passing through a global pooling layer, a convolutional layer, a batch normalization (BN) function, and a rectified linear unit (ReLU) function. After passing through a convolutional layer and a softmax activation function, we obtain... , The size of the convolutional kernel of the convolutional layer is . and Add, and After addition, it passes through a convolutional layer, and then... The output after addition is the size of the convolutional kernel of the convolutional layer. .

[0017] B. The Multi-Branch Cross Attention Module (DCSA) includes the following structure:

[0018] The Multi-Branch Cross Attention (DCSA) module consists of three CSA modules and a Conv Kernel. The size of the Conv Kernel can be k1×k1, k2×k2, or k3×k3.

[0019] C. The contention-dense block (CDB) includes the following structure:

[0020] The Competitive Dense Block (CDB) contains three cascaded encoder modules. Each encoder module consists of a Parametric Rectified Linear Unit (PReLU), a convolutional layer, a batch normalization (BN) layer, and two Maxout activation functions. The kernel size of the convolutional layer is [missing value]. .

[0021] D. The initialization of the contention-intensive block CDB1 includes the following structure:

[0022] Initializing the contention-dense block CDB1 involves replacing the PReLU of the first encoder module with the BN layer, based on the contention-dense block CDB.

[0023] E. Multi-branch intersection attention competition dense block MCDB includes the following structure:

[0024] The multi-branch cross-attention competitive dense block (MCDB) consists of three CDB modules and a convolutional kernel (Conv Kernel). The size of the convolutional kernel (Conv Kernel) can be one of three sizes: k1×k1, k2×k2, or k3×k3.

[0025] F. Dense block RD includes the following structure:

[0026] The dense block RD consists of two CBR modules (convolutional layer, batch normalization layer, and ReLU), one convolutional kernel (Conv Kernel), one batch normalization layer, one Maxout activation function, and one ReLU. The size of the convolutional kernel in the convolutional layer is... The size of the convolution kernel (Conv Kernel) can be k1×k1, k2×k2, or k3×k3.

[0027] G. The multi-branch crossover attention dense block DCDB includes the following structure:

[0028] The Multi-Branch Cross-Attention Dense Block (DCDB) consists of three Maxout activation functions, three RD modules, and a Conv Kernel. The size of the Conv Kernel can be k1×k1, k2×k2, or k3×k3. The Conv Kernel sizes in the three RD modules are k1×k1, k2×k2, and k3×k3, respectively.

[0029] In step 3), the loss function definition consists of two parts: the Dice similarity coefficient (DSC) and the cross-entropy. DSC is one of the most commonly used metrics for evaluating the performance of deep learning-based medical segmentation; its essence is to measure the overlap between two samples, which can be expressed mathematically as:

[0030]

[0031] Where T is the ground truth of the training labels, and P is the binary segmentation image predicted by the network. In the actual calculation process, It is the sum of all pixel values ​​after the dot product of T and P. and Then, these represent the sum of all pixel values ​​in T and P, respectively.

[0032] Cross-entropy is used to measure the similarity between two distributions. Cross-entropy can be defined as:

[0033]

[0034] in, It predicts the probability of each brain region in the segmentation feature map, and normalizes the probability values ​​into a probability distribution vector, where the sum of all probabilities is 1. It predicts the probability of each brain region in the label feature map and normalizes the probability values ​​into a probability distribution vector, with the sum of all probabilities being 1. This represents the probability distribution of the segmentation feature map. This represents the probability distribution of the label feature map.

[0035] Network loss function It is a composite loss function that is the sum of DSC and cross-entropy values:

[0036]

[0037] in, This represents the set of parameters within the overall network, where T is the ground truth of the training labels and P is the binary segmentation image of the network's prediction results. This represents the probability distribution of the segmentation feature map. This represents the probability distribution of the label feature map.

[0038] In step 4), the optimal parameters of the deep learning network model with multi-scale, multi-branch attention modules based on the Unet network are obtained by using the Adam optimizer, which performs well in deep learning. The network is trained using the training set generated in step 1), and the loss function in step 3) is minimized. Obtain the optimal parameter set , This represents the set of network parameters that result in optimal network segmentation.

[0039] In step 5), the T1-weighted MRI brain data to be segmented is input into the trained network for segmentation. The network segmentation process can be represented as follows:

[0040]

[0041] in, This represents the process of segmenting the overall network. This represents the input data to be segmented. This represents the optimal set of parameters in the trained network.

[0042] Compared with the prior art, the present invention has the following outstanding technical effects:

[0043] This invention introduces a multi-scale segmentation attention module and a multi-branch cross-attention module to increase receptive fields of different scales in order to obtain depth features and achieve accurate segmentation of more brain regions of different sizes. It can segment the whole brain into a network structure of 161 brain regions. This invention has greatly improved the number of brain regions segmented, the segmentation accuracy and speed. It has the characteristics of fast segmentation speed and high segmentation accuracy, which is helpful for the early detection of diseases. Attached Figure Description

[0044] Figure 1 It is a network structure of a deep learning network model with multi-scale and multi-branch attention modules, based on the Unet network.

[0045] Figure 2 This describes the structure of each module in the network model. Among them, (a) is the competition dense block CDB; (b) is the initialization competition dense block CDB1; (c) is the dense block RD; (d) is the multi-branch cross-attention competition dense block MCDB; (e) is the multi-branch cross-attention dense block DCDB; (f) is the multi-branch cross-attention module DCSA; and (g) is the multi-scale segmentation attention module CSA.

[0046] Figure 3These are the brain data to be segmented and the segmentation results. Among them, (a) is the sagittal brain image; (b) is the horizontal brain image; (c) is the coronal brain image; (d) is the sagittal brain segmentation result; (e) is the horizontal brain segmentation result; and (f) is the coronal brain segmentation result. Detailed Implementation

[0047] The following embodiments, in conjunction with the accompanying drawings, will further illustrate the present invention. The described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0048] This invention utilizes MRI image data from five publicly available datasets for training and testing. 140 datasets from the Autism Brain Imaging Data Exchange II (ABIDE II), the Alzheimer's Disease Neuroimaging Initiative (ADNI), the UCLA Neuropsychopoeia Consortium LA5c Study (LA5c), and the Open Access Imaging Series 2 (OASIS-2) were used for network training, while 10 datasets from Mindboggle-101 were used for testing. For model training, three independent models were trained separately for the coronal, axial, and sagittal planes. The final segmentation was generated by aggregating the probability maps of each model with corresponding weights of 0.4, 0.4, and 0.2. Specific embodiments are given below.

[0049] The embodiments of the present invention include the following steps:

[0050] The results of Destrieux brain atlas segmentation were used as training labels for model training and evaluation. After merging some left and right brain regions to reduce the number of labels, a training set of 122 brain region labels was obtained.

[0051] Step 1: Acquire T1-weighted MRI image data , Indicates the first T1-weighted MRI image data, Represent the real number field; then use the Destrieux mind map partitioning method to... The brain was segmented, resulting in a sample of 187 brain regions. , Indicates the first Layer image segmentation results, segmentation labels in this embodiment The data was obtained using FreeSurfer software; then, cortical regions that are in contact between the left and right hemispheres of the brain were merged into one brain region. After combining all other regions and removing several less frequently used regions, a result of 122 brain regions was obtained. , Indicates the first The result of reducing brain regions using layered images; finally, by , The training set is created and stored in an HDF5 file as input data for the network.

[0052] Step 2: Deep Learning Network Model with Multi-Scale, Multi-Branch Attention Modules Based on Unet. Building upon the Unet network, two attention modules are introduced: Multi-Scale Segmentation Attention (CSA) and Multi-Branch Cross-Attention (MCAM) to increase receptive fields at different scales. Seven modules with distinct compositions are connected through Maxout activation and skip paths to reduce the number of network training parameters (see...). Figure 1 ).

[0053] A. The multi-scale segmentation attention module (CSA) includes the following structure:

[0054] Multi-scale segmentation attention module CSA, such as Figure 2 As shown in Figure (g), the Multi-Scale Segmentation Attention Module (CSA) helps the network focus on features at different scales of the image; input data After passing through a convolutional layer (Conv), the layers are divided into two groups, namely... , The kernel size of the convolutional layer is 1×1. and Each is obtained after passing through a convolutional layer. , The kernel size of the convolutional layer is 3×3. , After addition, the result is obtained by passing through a convolutional layer. The kernel size of the convolutional layer is 3×3. and The sum is obtained by passing through a global pooling layer, a convolutional layer, a batch normalization (BN) function, and a rectified linear unit (ReLU) function. After passing through a convolutional layer and a softmax activation function, we obtain... , The kernel size of the convolutional layer is 1×1. and Add, and After addition, it passes through a convolutional layer, and then... The output result after addition has a 1×1 kernel size for the convolutional layer.

[0055] B. The Multi-Branch Cross Attention Module (DCSA) includes the following structure:

[0056] Multi-branch cross attention module DCSA, such as Figure 2 As shown in Figure (f), the Multi-Branch Cross-Attention (DCSA) module consists of three Cross-Attention (CSA) modules and a Conv Kernel. The size of the Conv Kernel can be k1×k1, k2×k2, or k3×k3. In this embodiment, k1=3, k2=5, and k3=7. The DCSA module is used to enhance the interaction and combination between features through multiple CSA modules. Each CSA module is responsible for processing features at different scales. Through the cross-attention mechanism, the DCSA module can effectively fuse features from different branches.

[0057] C. The contention-dense block (CDB) includes the following structure:

[0058] Contested dense block CDB such as Figure 2 As shown in Figure (a), the Competitive Dense Block (CDB) is used to make the network focus more on key features through a competitive mechanism, thereby improving the model's performance and robustness. The CDB module contains three cascaded encoder modules. Each encoder module consists of a Parametric Rectified Linear Unit (PReLU), a convolutional layer, and a batch normalization (BN) layer, as well as two Maxout activation functions. The PReLU activation function is a modified linear activation function that can solve the "death ReLU" problem that may occur during training with the traditional ReLU function. The batch normalization layer helps stabilize the network training process by normalizing the input of each layer. Each encoder module has the same structure, but the convolutional kernel size varies. In this embodiment, the first two encoder modules have 5×5 convolutional kernels, and the third encoder module has 1×1 convolutional kernels. Larger convolutional kernels can capture a wider range of features, while smaller kernels can capture more detailed information. In this way, the Competitive Dense Block (CDB) can perform a more comprehensive analysis of the input data.

[0059] D. The initialization of the contention-intensive block CDB1 includes the following structure:

[0060] Initialize the contention-intensive block CDB1 as follows Figure 2As shown in Figure (b), the initialization of the competitive dense block CDB1 module is based on the competitive dense block CDB module, but the PReLU of the first encoder module is replaced with a BN layer. In the CDB1 module, the structure of the first encoder module changes from PReLU + convolutional layer + BN layer to BN layer + convolutional layer + BN layer. The structures of the second and third encoder modules remain unchanged. Replacing PReLU with a BN layer changes the nonlinear characteristics of the network and also changes the way gradients propagate in the network. The initialization of the competitive dense block CDB1 is used to normalize the original input data, obtain data features, and reduce the number of parameters.

[0061] E. Multi-branch intersection attention competition dense block MCDB includes the following structure:

[0062] Multi-branch intersections require attention to contention-rich dense blocks such as MCDB. Figure 2 As shown in Figure (d), the Multi-Branch Cross-Attention Competitive Dense Block (MCDB) is used to segment the low-level features into three branches. The MCDB consists of three competing dense blocks (CDBs) and a convolutional kernel (ConvKernel). The size of the ConvKernel can be k1×k1, k2×k2, or k3×k3 to enhance the network's understanding of the input data. As data passes through these modules, the network learns features at different spatial resolutions.

[0063] F. Dense block RD includes the following structure:

[0064] Dense block RD such as Figure 2 As shown in Figure (c), the Dense Block RD is used in the Multi-Branch Cross-Attention Dense Block (DCDB), where convolutional kernels of different sizes are used for decoding. The Dense Block RD contains two CBR modules (convolutional layer, BN layer, PReLU), one convolutional kernel, one BN layer, one Maxout activation function, and one PReLU. Specifically, in this embodiment, the Dense Block RD includes: the first CBR module has a 3×3 convolutional kernel, and the second CBR module has a 1×1 convolutional kernel. The convolutional kernels following the two CBR modules can have sizes of k1×k1, k2×k2, and k3×k3. After the convolutional kernels, there is a Batch Normalization (BN) layer and a Maxout activation function, and finally a PReLU activation function. The Maxout and PReLU activation functions improve the nonlinearity and generalization ability of the network.

[0065] G. The multi-branch crossover attention dense block DCDB includes the following structure:

[0066] Multi-branch intersection attention dense block DCDB such as Figure 2As shown in Figure (e), the Multi-Branch Cross-Attention Dense Block (DCDB) is used to fuse multi-scale features and upsampling to obtain three feature maps. These feature maps are then sent to the next convolutional group to achieve the highest segmentation accuracy. The DCDB consists of three Maxout activation functions, three RD modules, and one Conv Kernel. The size of the Conv Kernel can be k1×k1, k2×k2, or k3×k3. The Conv Kernel sizes in the three RD modules are k1×k1, k2×k2, and k3×k3, respectively. By using multiple branches, the network learns features at different levels and finally fuses this information. The DCDB can integrate information from different levels to gain a more comprehensive understanding of the input data.

[0067] Step 3: The loss function definition consists of two parts: the Dice similarity coefficient (DSC) and cross-entropy. The loss function is a tool used when training a deep learning model to measure the difference between the model's predictions and the true values. DSC is one of the most commonly used metrics for evaluating the performance of deep learning-based medical segmentation; its essence is to measure the overlap between two samples, which can be expressed mathematically as:

[0068]

[0069] Where T is the ground truth of the training labels, and P is the binary segmentation image predicted by the network. In the actual calculation process, It is the sum of all pixel values ​​after the dot product of T and P. and Then, DSC represents the sum of all pixel values ​​in T and P, respectively. The value of DSC ranges from 0 to 1. The closer the value is to 1, the higher the overlap between the ground truth value T of the training label and the binary segmentation map P predicted by the network.

[0070] Cross-entropy is used to measure the similarity between two distributions. Cross-entropy can be defined as:

[0071]

[0072] in, It predicts the probability of each brain region in the segmentation feature map, and normalizes the probability values ​​into a probability distribution vector, where the sum of all probabilities is 1. It predicts the probability of each brain region in the label feature map and normalizes the probability values ​​into a probability distribution vector, with the sum of all probabilities being 1. This represents the probability distribution of the segmentation feature map. This represents the probability distribution of the label feature map.

[0073] Network loss function It is a composite loss function that is the sum of DSC and cross-entropy values:

[0074]

[0075] in, This represents the set of parameters within the overall network, where T is the ground truth of the training labels and P is the binary segmentation image of the network's prediction results. This represents the probability distribution of the segmentation feature map. This represents the probability distribution of the label feature map.

[0076] Step 4: Train the optimal parameters of a deep learning network model based on the Unet network, featuring multi-scale, multi-branch attention modules.

[0077] The U-Net architecture was used as the foundation, with the addition of multi-scale and multi-branch attention modules. The Adam optimizer, which performs well in deep learning, was employed. The training iterations were 30 epochs, the input data had 7 channels, the batch size was 8, and the weight decay rate was [missing value]. The initial learning rate was 0.01, and it decreased by 95% every 5 epochs. After training for all epochs, the model with the lowest validation loss was selected as the best model.

[0078] Step 5: Input the T1-weighted MRI brain data to be segmented into the trained network for segmentation. The network segmentation process can be represented as follows:

[0079]

[0080] in, This represents the process of segmenting the overall network. This represents the input data to be segmented. This represents the optimal set of parameters in the trained network.

[0081] In one embodiment, a brain image is input into a trained network to obtain a segmented image, such as... Figure 3 As shown.

[0082] The proposed method was tested on the Mindboggle-101 dataset using FastSurferCNN, MRF-Net, and other methods. The average Dice values ​​and segmentation times are shown in Table 1. Compared to the comparison methods FastSurfer and MRF-Net, the proposed method improved the average Dice by 0.97% and 0.63%, respectively. While maintaining segmentation accuracy, the proposed method reduced segmentation time, averaging 92.21 seconds. This represents a 6.74% and 15.29% reduction in segmentation time compared to FastSurfer and MRF-Net on the same server, respectively. This invention significantly improves upon the number of brain regions segmented, segmentation accuracy, and speed, demonstrating great potential for disease prediction.

[0083] Table 1 Comparison of segmentation results of different segmentation methods

[0084]

[0085] Note: The server's operating system is CentOS Linux 7, and the processor is an Intel Xeon Silver 4210.

[0086] This invention introduces lightweight multi-scale segmentation attention blocks and multi-branch cross-attention modules on the basis of the Unet network structure to improve the segmentation accuracy of brain regions of different sizes. At the same time, it introduces Maxout activation connection dense blocks to reduce the computational load of the network, enabling the segmentation of 161 brain regions with higher segmentation accuracy and faster segmentation speed.

[0087] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A brain region segmentation method based on deep learning, characterized in that... Includes the following steps: 1) Obtain the segmentation results from the Destrieux brain atlas as training labels for model training and evaluation. After merging some left and right brain regions to reduce the number of labels, a training set of 122 brain region labels is obtained. The specific method is as follows: First, obtain T1-weighted MRI image data. , Indicates the first T1-weighted MRI image data, Represents the real number field. These represent the length, width, and number of layers of the image in the data, respectively; then, the Destrieux brain mapping partitioning method is used to partition the T1-weighted MRI image data. The brain was segmented, resulting in a sample of 187 brain regions. , Indicates the first Layer image segmentation results; then, cortical regions that are in contact between the left and right hemispheres of the brain are merged into one brain region, and then all other regions are combined, and several less frequently used brain regions are removed, resulting in a result with 122 brain regions. , Indicates the first The result of reducing brain regions using layered images; finally, by , The training set is constructed and stored in an HDF5 file as input data for the network. 2) Design a deep learning network model based on the Unet network with multi-scale and multi-branch attention modules. Based on the Unet network, two attention modules are introduced: multi-scale segmentation attention module CSA and multi-branch cross attention module MCAM to increase the receptive field at different scales. The seven modules with different compositions are connected through Maxout activation operation and skip path to reduce the number of network training parameters. A. Multi-scale segmentation attention module (CSA): Input data After passing through a convolutional layer (Conv), the data is divided into two groups, namely... , The size of the convolutional kernel of the convolutional layer is ; and Each is obtained after passing through a convolutional layer. , The size of the convolutional kernel of the convolutional layer is ; , After addition, the result is obtained by passing through a convolutional layer. The size of the convolutional kernel of the convolutional layer is ; and The sum is obtained by passing through a global pooling layer, a convolutional layer, a normalization function, and a linear rectified function. After passing through a convolutional layer and a softmax activation function, we obtain... , The size of the convolutional kernel of the convolutional layer is ; and Add, and After addition, it passes through a convolutional layer, and then... The output after addition is the size of the convolutional kernel of the convolutional layer. ; B. Multi-branch Cross Attention Module (DCSA): It consists of three CSA modules and one Conv Kernel. The size of the Conv Kernel is k1×k1, k2×k2, k3×k3. C. Contested Dense Block (CDB): It contains three cascaded encoder modules. Each encoder module consists of a parameter-corrected linear function, a convolutional layer, a batch normalization (BN) layer, and two Maxout activation functions. The kernel size of the convolutional layer is [missing value]. ; D. Initialize the contention-intensive block CDB1: Initialize the contention-dense block CDB1 by replacing the PReLU of the first encoder module with the BN layer based on the contention-dense block CDB. E. Multi-branch intersection attention to contention-intensive block MCDB: It consists of three CDB modules and one Conv Kernel. The size of the Conv Kernel is k1×k1, k2×k2, k3×k3. F. Dense Block RD: Dense block RD contains two CBR modules, one convolutional kernel (Conv Kernel), one batch normalization (BN) layer, one Maxout activation function, and one ReLU; the size of the convolutional kernel in the convolutional layer is... The size of the convolution kernel (Conv Kernel) can be k1×k1, k2×k2, or k3×k3. G. Multi-branch intersection attention dense block DCDB: It consists of three Maxout activation functions, three RD modules and one Conv Kernel. The size of the Conv Kernel is k1×k1, k2×k2, k3×k3. The sizes of the convolutional kernels in the three RD modules are k1×k1, k2×k2, and k3×k3, respectively. 3) Construct the network's loss function; 4) Using the training set obtained in step 1), solve for the optimal parameters of the deep learning network model with multi-scale and multi-branch attention modules based on the Unet network. 5) Input the T1-weighted MRI brain data to be segmented into the trained network for segmentation.

2. The brain region segmentation method based on deep learning as described in claim 1, characterized in that... In step 3), the loss function definition consists of two parts: the Dice similarity coefficient and the cross-entropy. DSC is one of the most commonly used metrics for evaluating the performance of deep learning-based medical segmentation. Essentially, it measures the overlap between two samples, expressed mathematically as: Where T is the ground truth of the training labels, and P is the binary segmentation map of the network prediction result; in the actual calculation process, It is the sum of all pixel values ​​after the dot product of T and P. and Then, these represent the sum of all pixel values ​​in T and P, respectively; Cross-entropy is used to measure the similarity between two distributions. Cross-entropy is defined as: in, It predicts the probability of each brain region in the segmentation feature map, and normalizes the probability values ​​into a probability distribution vector, where the sum of all probabilities is 1. It predicts the probability of each brain region in the label feature map and normalizes the probability values ​​into a probability distribution vector, with the sum of all probabilities being 1; This represents the probability distribution of the segmentation feature map. Represents the probability distribution of the label feature map; Network loss function It is a composite loss function that is the sum of DSC and cross-entropy values: in, This represents the set of parameters within the overall network, where T is the ground truth of the training labels and P is the binary segmentation image of the network's prediction results. This represents the probability distribution of the segmentation feature map. This represents the probability distribution of the label feature map.

3. The brain region segmentation method based on deep learning as described in claim 1, characterized in that... In step 4), the optimal parameters of the deep learning network model with multi-scale, multi-branch attention modules based on the Unet network are obtained by using the Adam optimizer, which performs well in deep learning. The network is trained using the training set generated in step 1), and the loss function in step 3) is minimized. Obtain the optimal parameter set , This represents the set of network parameters that result in the optimal network segmentation performance.

4. The brain region segmentation method based on deep learning as described in claim 1, characterized in that... In step 5), the T1-weighted MRI brain data to be segmented is input into the trained network for segmentation. The network segmentation process is represented as follows: in, This represents the process of segmenting the overall network. This represents the input data to be segmented. This represents the optimal set of parameters in the trained network.

Images