Complex plexus segmentation method based on mixed attention mechanism
By using a deep learning model with a hybrid attention mechanism to enhance the boundaries of MRI images, the problems of insufficient accuracy and consistency in choroid plexus segmentation were solved, efficient and accurate choroid plexus automatic segmentation was achieved, and the efficiency of diagnosis and monitoring of neurological diseases was improved.
Patent Information
- Application Number
- CN202510723770.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-01
- Publication Date
- 2025-10-17
AI Technical Summary
Existing choroid plexus segmentation methods have limitations in accuracy and generalization ability. Traditional manual segmentation is time-consuming and easily affected by human factors. Existing automatic segmentation tools cannot meet the high requirements in medical image analysis.
A deep learning model with hybrid attention mechanism is used to perform boundary-enhanced choroid plexus segmentation on MRI images. It includes four dual-path proxy attention blocks, boundary extraction module, dynamic enhancement feature fusion block and residual block. The model parameters are optimized through adaptive moment estimation to achieve accurate choroid plexus segmentation.
It significantly improves the accuracy and consistency of choroid plexus segmentation, saves manpower and time costs, and improves the accuracy and efficiency of early detection and disease assessment of neurological diseases.
Smart Images
Figure CN120807904A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to a 3D segmentation method of a magnetic resonance imaging (MRI), in particular to an MRI choroid plexus segmentation method based on a hybrid attention mechanism, which is suitable for brain image analysis and belongs to the technical field of medical image segmentation and computer-aided medical treatment. BACKGROUND
[0002] The choroid plexus is an important structure in the brain ventricle, participates in the generation of cerebrospinal fluid, and plays an important role in the regulation of neuroinflammation and neurodegenerative diseases (such as multiple sclerosis, Alzheimer's disease, Parkinson's disease, etc.). The volume change of the choroid plexus is closely related to the occurrence and progression of these diseases, therefore, accurate choroid plexus segmentation has important clinical value for early diagnosis, disease assessment and efficacy monitoring. However, due to the complex morphology, fuzzy boundary, small volume of the choroid plexus, and low contrast with the surrounding brain tissue, automatic segmentation still faces significant challenges. Although the traditional manual segmentation method is regarded as the gold standard, it is time-consuming, subjective and easily affected by human factors, which limits its popularization in large-scale clinical applications. Although the existing automatic segmentation tools can improve efficiency, there are still certain limitations in accuracy and generalization ability, which cannot meet the high requirements of medical image analysis. SUMMARY
[0003] The application aims to solve the problems in the prior art, provide a choroid plexus segmentation method for boundary enhancement of MRI under a hybrid attention mechanism, save a large amount of manpower and time cost, improve the accuracy and consistency of choroid plexus segmentation, make the automatic segmentation of the choroid plexus region more accurate and reliable, and provide more effective technical support for early detection, disease assessment and efficacy monitoring of nervous system diseases (such as multiple sclerosis), and ultimately improve the efficiency and accuracy of clinical diagnosis.
[0004] To achieve the above-mentioned purpose, the application realizes the following technical scheme:
[0005] The choroid plexus segmentation method for boundary enhancement of MRI under a hybrid attention mechanism comprises the following steps:
[0006] S1, collecting a sufficient number of 3D brain MRI images;
[0007] S2, preprocessing the collected 3D MRI images, uniformly adjusting them to a fixed size, and expanding the data set by using translation, rotation and other methods, and simultaneously standardizing the image intensity value.
[0008] S3, constructing a choroid plexus segmentation model for boundary enhancement of MRI under a hybrid attention mechanism.
[0009] S4, input the pre-processed training set in S2 into the choroid plexus segmentation model under the mixed attention mechanism constructed in S3 for boundary enhancement of MRI for training, use the dice loss function for back propagation to update the network parameters, and optimize the model through adaptive moment estimation (Adam), to obtain the trained choroid plexus segmentation model under the mixed attention mechanism for boundary enhancement of MRI.
[0010] S5, input the brain MRI image to be segmented and predicted into the trained choroid plexus segmentation model under the mixed attention mechanism for boundary enhancement of MRI obtained in step 4 for segmentation, to obtain the segmentation prediction result of the MRI image of the data.
[0011] Preferably, in step S3, the choroid plexus segmentation model under the mixed attention mechanism for boundary enhancement of MRI includes four double-path proxy attention blocks (DPAA Block), a boundary extraction module (BEM), four dynamic enhancement feature fusion blocks (DEFF), and multiple residual blocks (Res Block).
[0012] In the encoder stage, each stage includes a DPAA Block, which is responsible for extracting and processing image features. The output feature vectors of the first and fourth layers of the encoder are input into the BEM to generate boundary-enhanced features.
[0013] The decoder stage utilizes up-sampling techniques and skip connections with the encoder stage, combined with DEFF modules to further enhance feature representation. In each layer of the decoder, the decoder's feature vector and the boundary-enhanced features from the encoder are input into the DEFF module to achieve dynamic fusion and enhancement of the features. The enhanced features are then up-sampled and passed to the upper decoder module, and finally output the segmentation prediction result. In addition, the residual blocks (Res Block) in the model are used to optimize the feature extraction process, ensuring that the model can effectively learn complex image features.
[0014] Preferably, in the choroid plexus segmentation model with boundary enhancement of MRI under the mixed attention mechanism, the design of the encoding stage adopts a multi-stage structure. The processing flow of each stage includes a convolution downsampling layer and a mixed attention mechanism module (DPAA) in turn. The output of each stage will be used as the input of the next stage. In the first stage, the input features are first processed by a 3x3x3 convolution downsampling layer, then enhanced by the DPAA module, and finally normalized by the Layer Norm layer. From the second stage, the processing flow is slightly different: each stage first normalizes the features by the Layer Norm layer, then performs 2x2x2 convolution downsampling, then connects the DPAA module, and finally extracts deep features.
[0015] Preferably, in the choroid plexus segmentation model with boundary enhancement of MRI under the mixed attention mechanism, the design of the encoding stage adopts a multi-stage structure. The processing flow of each stage includes a convolution downsampling layer and a mixed attention mechanism module (DPAA) in turn. The output of each stage will be used as the input of the next stage. In the first stage, the input features are first processed by a 3x3x3 convolution downsampling layer, then enhanced by the DPAA module, and finally normalized by the Layer Norm layer. From the second stage, the processing flow is slightly different: each stage first normalizes the features by the Layer Norm layer, then performs 2x2x2 convolution downsampling, then connects the DPAA module, and finally extracts deep features.
[0016] Preferably, in the choroid plexus segmentation model with boundary enhancement of MRI under the mixed attention mechanism, the DPAA module is composed of three parallel paths, namely the spatial attention path, the proxy attention path, and the local feature enhancement module (LFA).
[0017] The spatial attention module efficiently captures global spatial information through the agent mechanism. First, the input features are transformed through a series of linear transformations to generate query vectors (Q), spatial key vectors (K S ), and spatial value vectors (V S ). Then, the query vector Q is pooled to generate the agent vector A. Next, the similarity between the agent vector A and the spatial key vector K S is calculated, and the attention weight matrix is obtained by Softmax normalization. The weight matrix is multiplied with the spatial value vector V S to generate the agent feature Agent v . Finally, by calculating the Softmax operation again, the query vector Q, the agent vector A, and the agent feature Agent v are combined to obtain the final spatial attention map
[0018] The channel attention module models the channel information by sharing an agent vector A. First, a linear transformation is performed on the input to generate a channel key vector K C and a channel value vector V C . Then, the channel key vector K C is reduced in dimension using convolution to obtain a reduced K' C . Then, the similarity between the agent vector A and the reduced K' C is calculated, and a normalized spatial attention matrix is obtained through a Softmax operation. Then, the matrix is multiplied with the channel value vector Vc to obtain the channel feature
[0019] The local feature enhancement module (LFA) inputs the input feature into a 7x7x7 depth separable convolution layer, and then inputs the output of the convolution into a 1x1x1 point convolution layer. Finally, the output feature is obtained through a ReLU activation function
[0020] Finally, the outputs of the three paths are combined to obtain the output of the DPAA module, wherein and are concatenated along the channel dimension, and LFA(x) is added to the concatenated result to obtain the final output feature.
[0021] Preferably, in the plexus chorioideus segmentation model for boundary enhancement of MRI under the hybrid attention mechanism, the BEM module first inputs the features of the first stage (Stage 1) and the fourth stage (Stage 4) of the encoder into a 3D Sobel operator to extract edge information. The processed feature is multiplied element by element with the original input feature to generate a new feature map. Then, the channel number is adjusted through a 1x1x1 convolution, and the convolution output is standardized and nonlinearly activated. Next, the activated feature is added element by element to the original input feature to obtain the added feature. The feature is input into the first 3x3x3 convolution unit, and is processed through standardization and an activation function. Finally, the processed feature is input into the second 3x3x3 convolution unit, and is processed again through standardization and an activation function to obtain the output feature of the BEM module.
[0022] Preferably, in the plexus chorioideus segmentation model for boundary enhancement of MRI under the hybrid attention mechanism, the decoder stage is composed of four layers of structure. The input of the first layer decoder is the output of the last layer of the encoder, which is processed through a Res Block and then enters a 2x2x2 transpose convolution for upsampling. As Figure 5As shown, in the Res Block structure, the output after upsampling is input into a 3x3x3 convolution layer. The output of the convolution layer is input into an instance normalization (IN) layer for normalization, and then input into a Leaky ReLU (LR) activation function for nonlinear transformation. Then, the output of the convolution layer after a second 3x3x3 convolution is input into an instance normalization layer for normalization, and finally input into a Leaky ReLU activation function to obtain the final output of the layer. This process is repeated in subsequent decoder layers to finally obtain the segmentation map.
[0023] Preferably, in the choroid plexus segmentation model for boundary enhancement of MRI under the mixed attention mechanism, the DEFF module comprises two paths, one along the channel dimension for fusion and the other along the spatial dimension for fusion.
[0024] The channel dimension path first splices the feature maps along the channel dimension, and performs a 3x3x3 depth separable convolution on the spliced feature maps. The output of the convolution unit is connected to the input of the depth direction pooling layer, and the outputs of the depth direction and height direction and width direction pooling layers are jointly connected to the input of the splicing layer. The output of the splicing layer is connected to the input of the first 1x1x1 convolution unit, the output of the first 1x1x1 convolution unit is connected to the input of the batch normalization layer, and the output of the batch normalization layer is connected to the input of the nonlinear activation function. Then, the activated feature maps are segmented, and the segmented features are respectively connected to the Sigmoid activation function to obtain the attention weights in the depth, height and width directions. The attention weights in the three directions are multiplied element by element with the original spliced feature maps to obtain the channel fused feature maps.
[0025] The spatial dimension path inputs the boundary feature maps and the decoder feature maps into respective 1x1x1 convolution units, adds the outputs of the two convolution units, and connects them to the Sigmoid activation function to obtain the spatial attention weight.
[0026] Finally, the channel fused feature maps are multiplied element by element with the spatial attention weight to obtain the final enhanced feature maps.
[0027] Preferably, in step S4, the model parameters are repeatedly iterated and optimized to improve the accuracy and robustness of the model in the choroid plexus region.
[0028] Compared with the prior art, the present application has the following advantages:
[0029] In order to fully extract the information of the choroid plexus in the MRI image, the 3D MRI image is modeled and learned in the application, and effective feature extraction, three-dimensional boundary enhancement and efficient information fusion of the image are realized. Through the deep learning model, the application can accurately segment and predict the choroid plexus in the 3D MRI image, thereby improving the automatic segmentation accuracy of the choroid plexus region. The application not only significantly saves a large amount of manpower and time cost, but also improves the accuracy and consistency of early detection, disease assessment and efficacy monitoring of nervous system diseases (such as multiple sclerosis, etc.), so that patients can obtain diagnosis results more quickly and accurately, and finally improve the diagnosis and treatment efficiency. BRIEF DESCRIPTION OF DRAWINGS
[0030] Figure 1 is a choroid plexus segmentation method for MRI under the mixed attention mechanism of the application;
[0031] Figure 2 is a deep learning model structure diagram constructed by the application;
[0032] Figure 3 is a schematic diagram of the DPAA module;
[0033] Figure 4 is a schematic diagram of the BEM module;
[0034] Figure 5 is a schematic diagram of the Res block;
[0035] Figure 6 is a schematic diagram of the DEFF module; DETAILED DESCRIPTION
[0036] The application will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the application and not to limit the scope of the application. In addition, it should be understood that those skilled in the art can make various modifications or changes to the application after reading the content taught by the application, and these equivalent forms also fall within the scope defined by the application.
[0037] At present, the automatic medical image analysis method has become a research hotspot, and the deep learning and computer vision technology provides a new possibility for the automatic analysis of medical images and disease diagnosis. The deep learning model with mixed attention mechanism is used in the application, which can accurately extract the information of the choroid plexus region and efficiently segment the choroid plexus in the 3D brain MRI image. The traditional medical image analysis method usually relies on manual annotation, which is tedious and easy to be disturbed by human factors, resulting in poor accuracy and consistency of the segmentation result. The automatic segmentation method of the application uses deep learning to train the MRI image, so that the segmentation result of the choroid plexus is more accurate and reliable.
[0038] The choroid plexus segmentation method of the present application first performs data preprocessing on the 3D MRI image, including image standardization, size adjustment, and data enhancement processing steps. Then, by constructing a deep learning model based on a hybrid attention mechanism, combined with boundary enhancement technology, the segmentation accuracy of the choroid plexus region can be effectively improved. Compared with traditional methods, the automatic segmentation method of the present application not only significantly improves the accuracy of the segmentation result, but also reduces manual intervention and saves a lot of time and labor cost.
[0039] The hybrid attention mechanism model of the present application can automatically analyze and extract choroid plexus features in 3D MRI images and perform fine segmentation. This process significantly improves the accuracy and consistency of choroid plexus segmentation through training and optimization of a large amount of medical image data. This technology is not only suitable for early diagnosis and disease monitoring of neurological diseases such as multiple sclerosis (MS), but also provides reliable auxiliary diagnostic tools for clinicians, improving diagnosis and treatment efficiency.
[0040] Therefore, the choroid plexus segmentation method based on the hybrid attention mechanism provided by the present application has wide medical application value and can provide more efficient and accurate technical support for the diagnosis and treatment of neurological diseases, with significant practical application prospects.
[0041] Embodiment: As shown in the accompanying Figures 1-6 The present application is a choroid plexus segmentation method based on boundary enhancement of MRI under a hybrid attention mechanism. This method aims to use deep learning and computer vision technology to quickly and accurately segment the choroid plexus of brain 3D MRI data collected from the Department of Neurological Radiology of the First Affiliated Hospital of Chongqing Medical University. This method accurately segments the choroid plexus region in medical images, achieving accuracy and efficiency in early diagnosis and disease assessment of neurological diseases such as multiple sclerosis.
[0042] Specifically includes the following steps:
[0043] S1, collect data sets, including 3D MRI images of 75 healthy people and 3D MRI images of 164 patients diagnosed with multiple sclerosis. (This embodiment is based on data collected by the Department of Neurological Radiology of the First Affiliated Hospital of Chongqing Medical University.)
[0044] S2, a series of preprocessing operations are performed on the collected 3D MRI images. First, all images are uniformly adjusted to a size of 128x128x128. Then, the preprocessed image dataset is divided into a training set and a test set according to a ratio of 8:2. In addition, translation, rotation, and other data enhancement techniques are applied to all images, and intensity value standardization processing is performed.
[0045] S3, construct a model for MRI boundary enhancement under mixed attention mechanism, and set the model parameters.
[0046] The model for MRI boundary enhancement under mixed attention mechanism, as shown in Figure 2 , includes four double-path agent attention blocks (DPAA Block), a boundary extraction module (BEM), four dynamic enhancement feature fusion blocks (DEFF), and multiple residual blocks (Res Block).
[0047] 1) The design of the encoder module adopts a multi-stage structure. The processing flow of each stage includes a convolution downsampling layer and a mixed attention mechanism module (DPAA) in turn. The output of each stage will be used as the input of the next stage. In the first stage, the input features are first processed by a 3x3x3 convolution downsampling layer, then enhanced by the DPAA module, and finally normalized by the standardization layer (Layer Norm). From the second stage, the processing flow is slightly different: each stage first normalizes the features by the standardization layer, then performs 2x2x2 convolution downsampling, followed by the DPAA module, and finally extracts deep features.
[0048] 2) DPAA feature extraction module
[0049] Referring to Figure 3 , the DPAA module consists of three parallel paths, which are spatial attention path, agent attention path, and local feature enhancement module (LFA).
[0050] The spatial attention module efficiently captures global spatial information through the agent mechanism. First, the input features are transformed by a series of linear transformations to generate the query vector (Q), the spatial key vector (K S ), and the spatial value vector (V S ). Then, the query vector Q is pooled to generate the agent vector A. Next, the similarity between the agent vector A and the spatial key vector K S is calculated, and the attention weight matrix is obtained by Softmax normalization. The weight matrix is multiplied with the spatial value vector V S to generate the agent feature Agent v . Finally, by calculating Softmax again, the query vector Q, the agent vector A, and the agent feature Agent v are combined to obtain the final spatial attention map
[0051] The channel attention module models the channel information through a shared agent vector A. First, the input is linearly transformed to generate the channel key vector K Cand channel value vector V C . Then use convolution to the channel key vector K C , and get the reduced dimension K' C . Then, calculate the similarity between the agent vector A and the reduced dimension K' C , and get the normalized spatial attention matrix through the Softmax operation, then use the matrix to multiply the channel value vector V C , and get the channel feature
[0052] The local feature enhancement module (LFA) inputs the input feature into a 7x7x7 depth separable convolution layer, and then inputs the output of the convolution into a 1x1x1 point convolution layer. Finally, the output feature is obtained through the ReLU activation function
[0053] Finally, the outputs of the three paths are combined to obtain the output of the DPAA module, where and are concatenated along the channel dimension, and is element-wise added to the concatenated result to obtain the final output feature.
[0054] 3) BEM boundary enhancement module
[0055] Referring to Figure 4 (A), the BEM module first inputs the features of the first stage (Stage 1) and the fourth stage (Stage
[0056] 4) of the encoder into a 3D Sobel operator to extract edge information. The processed feature is element-wise multiplied with the original input feature to generate a new feature map. Then, the channel number is adjusted through a 1x1x1 convolution, and the convolution output is normalized and nonlinearly activated.
[0057] Next, the activated feature is element-wise added to the original input feature to obtain the added feature. Referring to Figure 4 (B), the feature is input into the first layer of 3x3x3 convolution unit, and is processed through normalization and activation function. Finally, the processed feature is input into the second 3x3x3 convolution unit, and is processed again through normalization and activation function to obtain the output feature of the BEM module.
[0058] 4) The decoder stage consists of four layers of structure. The input of the first layer decoder is the output of the last layer of the encoder, which is processed through the Res Block and then enters the 2x2x2 transpose convolution for up-sampling. As Figure 5As shown, in the Res Block structure, the output after upsampling is input into a 3x3x3 convolution layer. The output of the convolution layer is input into an instance normalization (IN) layer for normalization, and then input into a Leaky ReLU (LR) activation function for nonlinear transformation. Then, the output of the convolution layer after a second 3x3x3 convolution is input into an instance normalization layer for normalization, and finally input into a Leaky ReLU activation function to obtain the final output of the layer. This process is repeated in subsequent decoder layers to finally obtain the segmentation map.
[0059] 5) DEFF feature fusion module
[0060] Referring to Figure 6 The DEFF module includes two paths, one for fusion along the channel dimension and the other for fusion along the spatial dimension.
[0061] The channel path first concatenates the feature maps along the channel dimension, and performs a 3x3x3 depth separable convolution on the concatenated feature maps. The output of the convolution unit is connected to the input of the depth direction pooling layer, and the outputs of the depth direction, height direction, and width direction pooling layers are collectively connected to the input of the concatenation layer. The output of the concatenation layer is connected to the input of the first 1x1x1 convolution unit, the output of the first 1x1x1 convolution unit is connected to the input of the batch normalization layer, and the output of the batch normalization layer is connected to the input of the nonlinear activation function. Then, the activated feature maps are segmented, and the segmented features are respectively connected to the Sigmoid activation function to obtain the attention weights in the depth, height, and width directions. The attention weights in the three directions are multiplied element by element with the original concatenated feature maps to obtain the channel fused feature maps.
[0062] The spatial dimension path inputs the boundary feature map and the decoder feature map into respective 1x1x1 convolution units, adds the outputs of the two convolution units, and connects them to a Sigmoid activation function to obtain the spatial attention weight. Finally, the channel fused feature maps are multiplied element by element with the spatial attention weight to obtain the final enhanced feature maps.
[0063] After the deep learning model is constructed, the initial parameters of the network need to be set. In this embodiment, the network parameters set and the training process include: 300 rounds of training process, the initial learning efficiency is 10-4, the learning rate adopts a cosine annealing strategy, gradually reduces, and an early stopping mechanism is introduced to avoid overfitting.
[0064] S4, input the preprocessed training set in S2 into the model of the choroid plexus segmentation method for boundary enhancement of MRI under the mixed attention mechanism constructed in S3 for training, use the dice loss function for back propagation to update the network parameters, and optimize the model through adaptive moment estimation (Adam), to obtain the trained model of the choroid plexus segmentation method for boundary enhancement of MRI under the mixed attention mechanism.
[0065] The formula of the dice loss function is as follows:
[0066]
[0067] Wherein, N represents the number of samples, r n represents the label of sample n, and the positive class is 1 and the negative class is 0. p n represents the probability that sample n is predicted as a positive class.
[0068] S5, input the 3D brain MRI image to be segmented and predicted into the trained model of the choroid plexus segmentation method for boundary enhancement of MRI under the mixed attention mechanism obtained in S4 for segmentation, to obtain the segmentation prediction result of the 3D MRI image.
Claims
1. A choroid plexus segmentation method for MRI with boundary enhancement under a hybrid attention mechanism, characterized by: Including steps: S1. Acquire a sufficient number of 3D brain MRI images; S2. Preprocess the collected 3D MRI images, adjust them to a fixed size, apply data enhancement methods such as translation and rotation, and standardize the image intensity values; S3. Construct a choroid plexus segmentation model for MRI boundary enhancement under a hybrid attention mechanism; S4. Input the preprocessed training set in S2 into the choroid plexus segmentation model for MRI boundary enhancement under the hybrid attention mechanism constructed in S3 for training, use the dice loss function for backpropagation to update the network parameters, and optimize the model through adaptive moment estimation (Adam), so as to obtain the trained choroid plexus segmentation model for MRI boundary enhancement under the hybrid attention mechanism; S5. Send the brain MRI image to be segmented and predicted into the choroid plexus segmentation model that performs boundary enhancement on MRI under the trained hybrid attention mechanism obtained in S4 for segmentation, and obtain the segmentation prediction result of the MRI image of the data.
2. The choroid plexus segmentation model for MRI boundary enhancement under the hybrid attention mechanism according to claim 1, characterized in that: The choroid plexus segmentation model for MRI boundary enhancement based on the hybrid attention mechanism in step S3 includes four dual-path proxy attention blocks (DPAA blocks), a boundary extraction module (BEM), four dynamic enhancement feature fusion blocks (DEFF) and multiple residual blocks (Res blocks): During the encoding phase, each stage includes a DPAA block responsible for extracting and processing image features. The feature vectors output by the first and fourth layers of the encoding phase are input into the BEM to generate boundary-enhanced features. The decoding phase utilizes upsampling techniques and skip connections with the encoding phase, combined with the DEFF module to further enhance feature representation. At each layer in the decoding phase, the decoder's feature vector is fed into the DEFF module along with the boundary-enhanced features from the encoder, enabling dynamic feature fusion and enhancement. The enhanced features are then upsampled and passed to the upper decoder module, ultimately outputting the segmentation prediction. Furthermore, the residual block (Res Block) in the model optimizes the feature extraction process, ensuring that the model can effectively learn complex image features.
3. The choroid plexus segmentation model for MRI boundary enhancement under the hybrid attention mechanism according to claim 2, characterized in that: The encoding stage is designed with a multi-stage structure. The processing flow of each stage includes a convolutional downsampling layer and a hybrid attention mechanism module (DPAA). The output of each stage serves as the input to the next stage. In the first stage, the input features are first processed by a 3×3×3 convolutional downsampling layer, then enhanced by the DPAA module, and finally normalized by the normalization layer (Layer Norm). Starting from the second stage, the processing flow is slightly different: each stage first normalizes the features by the normalization layer, then performs a 2×2×2 convolutional downsampling, and then connects to the DPAA module to finally extract deep features.
4. The choroid plexus segmentation model for MRI boundary enhancement under the hybrid attention mechanism according to claim 2, characterized in that: In the choroid plexus segmentation model for MRI boundary enhancement under the hybrid attention mechanism, the DPAA module consists of three parallel paths, namely the spatial attention path, the proxy attention path and the local feature enhancement module (LFA). The spatial attention module efficiently captures global spatial information through an agent mechanism. First, the input features undergo a series of linear transformations to generate query vectors (Q), spatial key vectors (K S ) and space-valued vectors (V S ). Subsequently, the query vector Q is pooled to generate a proxy vector A. Next, the proxy vector A and the spatial key vector K are calculated. S The similarity between them is calculated and the attention weight matrix is obtained by Softmax normalization. This weight matrix is combined with the spatial value vector V S Perform matrix multiplication to generate agent features v Finally, by calculating the Softmax operation again, the query vector Q, the agent vector A and the agent feature Agent are combined v , and obtain the final spatial attention map The channel attention module models channel information by sharing the proxy vector A. First, the input is linearly transformed to generate the channel key vector K C and channel value vector V C . Then use convolution to transform the channel key vector K C Perform dimensionality reduction and obtain the K after dimensionality reduction ′ C Then, calculate the proxy vector A and the K after dimensionality reduction ′ C The similarity between them is calculated and the normalized spatial attention matrix is obtained through the Softmax operation, which is then used with the channel value vector V C Perform matrix multiplication to obtain channel features The local feature enhancement module (LFA) passes the input features into the 7×7×7 depth-separable convolution layer, and the convolution output is then input into the 1×1×1 point convolution layer, and finally the output features are obtained through the ReLU activation function. Finally, the outputs of the three paths are combined to obtain the output of the DPAA module, where and The concatenation is performed along the channel dimension, and LFA(x) is added to the concatenated result to obtain the final output feature.
5. The choroid plexus segmentation model for MRI boundary enhancement under the hybrid attention mechanism according to claim 2, characterized in that: The BEM module first feeds the features from the first (Stage 1) and fourth (Stage 4) encoder stages into a 3D Sobel operator to extract edge information. The processed features are element-wise multiplied with the original input features to generate a new feature map. The number of channels is then adjusted through a 1×1×1 convolution, and the convolution output is normalized and nonlinearly activated. Next, the activated features are element-wise added to the original input features to produce the summed features. This feature is fed into the first layer of 3×3×3 convolutional units, where it is normalized and activated. Finally, the processed features are fed into a second layer of 3×3×3 convolutional units, where they are again normalized and activated to produce the output features of the BEM module.
6. The choroid plexus segmentation model for MRI boundary enhancement under the hybrid attention mechanism according to claim 2, characterized in that: The decoder stage consists of a four-layer structure. The input of the first decoder layer is the output of the last layer of the encoder. After being processed by the Res Block, it enters a 2×2×2 transposed convolution for upsampling. As shown in Figure 5, in the Res Block structure, the upsampled output passes through a 3×3×3 convolution layer. The output of the convolution layer enters the instance normalization layer (IN) for normalization, and then passes through the Leaky ReLU (LR) activation function for nonlinear transformation. Then, after a second 3×3×3 convolution, the output of the convolution layer enters the instance normalization layer again for normalization, and finally passes through the Leaky ReLU activation function to obtain the final output of this layer. This process is repeated in subsequent decoder layers to finally obtain the segmentation map.
7. The choroid plexus segmentation model for MRI boundary enhancement under the hybrid attention mechanism according to claim 2, characterized in that: The DEFF module contains two paths, one for fusion along the channel dimension and the other for fusion along the spatial dimension. The channel-dimensional path first concatenates the feature maps along the channel dimension and performs a 3×3×3 depthwise separable convolution on the concatenated feature maps. The output of the convolution unit is connected to the input of the depthwise pooling layer, and the outputs of the depthwise, heightwise, and widthwise pooling layers are connected to the input of the concatenation layer. The output of the concatenation layer is connected to the input of the first 1×1×1 convolution unit, which in turn is connected to the input of the batch normalization layer. The output of the batch normalization layer is connected to the input of the nonlinear activation function. The activated feature map is then segmented, and the segmented features are connected to the Sigmoid activation function to obtain attention weights in the depth, height, and width directions. The attention weights in these three directions are element-wise multiplied with the original concatenated feature map to obtain a channel-fused feature map. The spatial dimension path inputs the boundary feature map and the decoder feature map into their respective 1×1×1 convolution units, adds the outputs of the two convolution units and connects them to the Sigmoid activation function to obtain the spatial attention weight. Finally, the channel-fused feature map is multiplied element-wise with the spatial attention weight to obtain the final enhanced feature map.
8. The choroid plexus segmentation model for MRI boundary enhancement under the hybrid attention mechanism according to claim 1, characterized in that: In step S4, the model parameters are optimized through repeated iterations to improve the accuracy and robustness of the model for the choroid plexus region.