Network suitable for medical image segmentation and implementation method
By using ResNet-50, Enhanced CBAM module and EnNon-local module in the medical image segmentation model, the problem of lack of global information and high cost of Transformer computing is solved, and more efficient medical image segmentation performance is achieved.
Patent Information
- Application Number
- CN202411228910.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-03
- Publication Date
- 2024-11-22
AI Technical Summary
The existing medical image segmentation model lacks global information in the CNN part, and the calculation of Transformer is large, which limits the performance of the model.
ResNet-50 is used as an encoder, combined with the Enhanced CBAM module and the EnNon-local module, extract global information and encode long-distance dependency features, reducing the computing burden of Transformer.
It effectively increases the receptive field of the model, improves the ability to capture and transmit long-distance information, reduces computing consumption, and provides richer features for medical image segmentation.
Smart Images

Figure CN119006828A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical image processing, and in particular to a network suitable for medical image segmentation and an implementation method thereof. Background Art
[0002] Convolutional neural networks (CNNs) have been widely used in the field of medical imaging, and CNN-based architectures have become mainstream in medical image segmentation. Thanks to the inductive bias of convolutional neural networks, CNNs can effectively capture local information and have characteristics such as translation invariance, and have become the basis of most models. However, due to the lack of long-distance information, CNNs can only encode features in a local range, resulting in limited model performance. In order to alleviate the lack of long-distance information in CNN models, Transformers used in the field of natural language have been introduced into the field of medical image segmentation, and have shown powerful long-distance information encoding capabilities. However, Transformers require a large number of parameters and computing resources, and require large-scale and diverse data to ensure that the model can learn accurate and comprehensive information, which limits the model's application in complex and specific medical images.
[0003] At present, the encoder structure of the hybrid model can be divided into two types, such as Figure 1 As shown in (a)(b). (a) means that CNN first performs local encoding and then uses Transformer to model long-distance dependencies, such as TransUNet, TransNorm and other models. (b) means that CNN and Transformer jointly encode local features, such as HiFormer, TransCeption and other models. When the CNN part of the existing model extracts local features, it lacks global information and has a relatively limited receptive field. At the same time, compared with the Transformer, it needs to stack multiple layers to encode long-distance information, which brings huge computational consumption. It is urgent to provide a new network and implementation method suitable for medical image segmentation to solve the above problems. Summary of the invention
[0004] The technical problem to be solved by the present invention is to provide a network and implementation method suitable for medical image segmentation. The proposed encoder makes up for the lack of global information in the CNN part of the existing model, and uses the self-attention mechanism to encode long-distance information, which reduces the computational burden brought by the Transformer mechanism to a certain extent.
[0005] In order to solve the above technical problems, a technical solution adopted by the present invention is: providing a network suitable for medical image segmentation, including an encoder and a decoder;
[0006] The encoder is based on ResNet-50, and encodes a given medical image through four serial stages. After each stage, the feature resolution is halved. Each stage consists of a single Enhanced CBAM module and a continuous ResidualCNN layer to extract local features with global information, and then encode the local features into features with long-distance dependencies through an EnNon-local module.
[0007] The decoder uses the final output features of the encoder as the initial input of the decoder after passing through a 3×3 convolution block, and concatenates the high-resolution features from the encoder with the high-dimensional features of the decoder through jump connections, and restores the target resolution step by step through cascade upsamplers.
[0008] In a preferred embodiment of the present invention, the Enhanced CBAM module includes a biased channel attention module, a biased spatial attention module, and a residual connection layer. First, the input feature is compressed into a biased feature B through a 1×1 convolution block, and then the input feature and the biased feature B are input into the biased channel attention module to reconstruct the channel dimension feature, and then the channel dimension reconstructed feature and the biased feature are input into the biased spatial attention module to reconstruct the spatial dimension feature. Finally, the spatial dimension reconstructed feature is residually connected with the input feature to output the global feature.
[0009] Furthermore, in the encoder, for the encoding feature x at a certain stage ei (i∈[1~4]), the calculation formula of the output feature f of the EnhancedCBAM module is:
[0010] f = SA (CA (x ei ,B),B)+x ei (1)
[0011] Where CA represents biased channel attention, SA represents biased spatial attention, and B represents biased features.
[0012] Furthermore, the bias channel attention module receives the input feature x ei And bias feature B, first add the two input features element by element, perform maximum pooling and average pooling to obtain parameters, then reconstruct the initially obtained parameters through shared MLP, and the sum of the reconstructed parameters is normalized by Sigmoid to obtain the final channel weight Weight c , the final output feature x c By channel weight Weight c With input feature x ei The specific mathematical model is obtained by multiplying each element:
[0013] Weight c=Sigmoid(M(Mp(x ei +B))+M(Ap(x ei +B))) (2)
[0014]
[0015] Where M represents a multi-layer perceptron, Mp represents maximum pooling, and Ap represents average pooling.
[0016] Furthermore, the biased spatial attention module receives the output x from the biased channel attention module c and bias feature B as input, x c The features obtained by maximum pooling and average pooling are concatenated with the bias features, and the obtained mixed features are convolved and normalized with Sigmoid to obtain the spatial weight Weight s , input feature x c and spatial weight s Perform element-by-element multiplication to obtain the reconstructed output feature x cs , the specific mathematical model is:
[0017] Weight s =Sigmoid(Conv(Concat(Mp(x c ),Ap(x c ),B))) (4)
[0018]
[0019] Concat represents a concatenation operation.
[0020] In a preferred embodiment of the present invention, the EnNon-local module includes four 1×1 convolution blocks and a SoftMax normalization layer. For local features that require long-distance information encoding, they are first passed through three 1×1 convolution blocks and reshaped to obtain Q, K, and V; according to the classic self-attention mechanism, the result of matrix multiplication of Q and K is processed by SoftMax normalization, and matrix multiplication is performed again with V to obtain global features after reshaping transformation; create learnable spatial features, add the learnable spatial features to the global features element by element, and finally pass through a 1×1 convolution layer and add the initial input features element by element to obtain output features with long-distance dependencies.
[0021] In a preferred embodiment of the present invention, the cascade upsampler consists of a transposed convolutional layer with a stride of 2, a 3×3 convolutional layer with a stride of 1, and a ReLU layer.
[0022] In order to solve the above technical problems, the second technical solution adopted by the present invention is: to provide an implementation method suitable for medical image segmentation, which is implemented by using a network suitable for medical image segmentation as described in any of the above items, and includes the following steps:
[0023] Taking ResNet-50 as the main encoder, a given medical image is first encoded through four serial stages. After each stage, the feature resolution is halved. Each stage consists of a single Enhanced CBAM module and a continuous ResidualCNN layer to extract local features with global information. The local features are then encoded into features with long-distance dependencies through the EnNon-local module.
[0024] The final output features of the encoder are passed through a 3×3 convolution block as the initial input of the decoder, and the high-resolution features from the encoder are concatenated with the high-dimensional features of the decoder through a skip connection, and are gradually restored to the target resolution through a cascade upsampler;
[0025] Finally, a 1×1 convolutional segmentation head is used to decode the features finally decoded by the decoder to obtain a segmentation mask.
[0026] To solve the above technical problems, the third technical solution adopted by the present invention is: to provide an implementation device suitable for medical image segmentation, including a memory and a processor, the memory is used to store at least one program, and the processor is used to load the at least one program to execute any of the methods described above.
[0027] In order to solve the above technical problems, the fourth technical solution adopted by the present invention is: to provide a storage medium, in which processor-executable instructions are stored, and the processor-executable instructions are used to execute any of the methods described above when executed by the processor.
[0028] The beneficial effects of the present invention are:
[0029] (1) The encoder of the network model described in the present invention uses the designed Enhanced CBAM module to extract global information, guides local feature encoding with global information, captures richer local features, and then uses the self-attention mechanism to encode long-distance dependencies of local features with global information, expands the receptive field of the model, and effectively increases the ability to capture and transmit long-distance information in the model; compared with the huge computational consumption caused by the Transformer needing to stack multiple layers for long-distance information encoding, a single EnNon-local module can model long-distance dependencies, reducing the computational burden;
[0030] (2) The implementation method proposed in the present invention targets the combination of local information and long-distance information, and utilizes the advantages of the encoder structure that can effectively combine the attention and self-attention mechanisms to provide the decoder with richer and more comprehensive features. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 It is a schematic diagram of the combination of the existing hybrid model and the encoder of the present invention;
[0032] Figure 2 This is a network architecture diagram suitable for medical image segmentation in the present invention;
[0033] Figure 3 is a schematic diagram of the structure of the Enhanced CBAM module;
[0034] Figure 4 is a flowchart of the method for implementing the Enhanced CBAM module;
[0035] Figure 5 is a schematic diagram of the structure of the bias channel attention module;
[0036] Figure 6 is a schematic diagram of the structure of the biased spatial attention module;
[0037] Figure 7 is a schematic diagram of the structure of the EnNon-local module;
[0038] Figure 8 is a flow chart of the method for implementing the EnNon-local module;
[0039] Fig. 9 This is a comparison chart of the effects of each network module;
[0040] Fig.10 This is a comparison chart of the segmentation results of each network model for the Synapse dataset. DETAILED DESCRIPTION
[0041] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.
[0042] First, the design idea of the present invention is briefly introduced. Figure 1(c) represents the encoder structure proposed by the present invention. Compared with the existing model, the present invention combines the attention mechanism in the CNN encoding part and replaces the Transformer with the self-attention module. The attention mechanism can effectively increase the transmission and capture capabilities of long-distance information in the model through its unique dynamic weight generation mechanism. The self-attention mechanism allows the model to refer to the information of all other positions when processing information at a certain position, and also has the ability to encode long-distance information. The combination of the attention mechanism and the self-attention mechanism can enrich the features captured by the encoder and lay the foundation for obtaining better medical image segmentation results.
[0043] In order to solve the limitations of CNN itself and alleviate the computational burden brought by Transformer, the Atten-Nonlocal Unet proposed in the present invention combines the Enhanced CBAM attention module and the EnNon-local self-attention module to capture 2D features, aiming to make up for the lack of long-distance information in the CNN model. More specifically, the Enhanced CBAM module is used to capture global information in the encoder to guide local information modeling and alleviate the lack of global information in the early stage of the model; and finally the EnNon-local module is used to query the long-distance dependencies of local features to fully solve the limitations of the CNN model. Compared with hybrid models such as TransUNet, HiFormer, and TransCeption, the proposed encoder makes up for the lack of global information in the CNN part of the existing model, and uses the self-attention mechanism to encode long-distance information, which reduces the computational burden brought by the Transformer mechanism to a certain extent.
[0044] See also Figure 2 The present invention provides a network (Atten-NonlocalUnet model) suitable for medical image segmentation, including an encoder and a decoder. Given an image x with a spatial resolution of H×W and a number of channels C, that is, x∈R H ×W×C , the goal is to predict the corresponding pixel-level label map of size H×W as an example.
[0045] The encoder is based on ResNet-50, but uses convolutional downsampling instead of pooling downsampling. A given medical image is first encoded through four serial stages. After each stage, the feature resolution is halved. Each stage consists of a single Enhanced CBAM module and a continuous Residual CNN layer to extract local features with global information. Where ei represents the i-th stage of the encoder, S ei represents the downsampling multiple of the encoder at stage i, C eirepresents the number of channels in the i-th stage of the encoder. Then the local feature x e4 Encoded into features with long-range dependencies through the EnNon-local module
[0046] The existing CBAM module reorganizes features through a sequential structure of first channel attention module (CAM) and then spatial attention module (SAM), which significantly improves the model's ability to capture key information. However, the inventors found that the two attention modules not only have an impact on the combination method, but also lack a certain connection in the learning of weight parameters. Therefore, the present invention designs an Enhanced CBAM module, such as Figure 3 As shown, the EnhancedCBAM module includes a bias channel attention module, a bias space attention module, and a residual connection layer. The implementation method flowchart is as follows Figure 4 As shown in the figure, the input features are compressed into bias features B through 1×1 convolution blocks, which serve as a bridge between the channel attention module and the spatial attention module. The bias features, as the common bias of the two modules, deepen the deep intrinsic connection between the two in weight parameter learning, so that the Enhanced CBAM module can better reorganize features and have richer information.
[0047] Specifically, in the encoder, for the encoding feature x at a certain stage ei , if Indicates the number of channels of the current feature C ei , Indicates the spatial resolution of the current feature Right now The output characteristics of the Enhanced CBAM module are The calculation formula is,
[0048] f = SA (CA (x ei ,B),B)+x ei (1)
[0049] Where CA represents biased channel attention and SA represents biased spatial attention.
[0050] More specifically, the biased channel attention module BCAM reconstructs features in the channel direction, and its structure is as follows: Figure 5 As shown. This module receives the input feature x ei and bias features The two input features are added element by element, and the parameters are obtained by performing maximum pooling and average pooling respectively. The initially obtained parameters are reconstructed through the shared MLP, and the sum of the reconstructed parameters is normalized by Sigmoid to obtain the final channel weight. Final output features By channel weight Weight c With input feature x ei The specific mathematical model is obtained by multiplying each element:
[0051] Weight c =Sigmoid(M(Mp(x ei +B))+M(Ap(x ei +B))) (2)
[0052]
[0053] Where M represents a multi-layer perceptron, Mp represents maximum pooling, and Ap represents average pooling.
[0054] The structure of the biased spatial attention module BSAM is as follows Figure 6 As shown, this module receives the output x from the channel attention module c and bias feature B as input. c The features obtained by maximum pooling and average pooling are concatenated with the bias features, and the obtained mixed features are convolved and normalized with Sigmoid to obtain spatial weights. Input feature x c and spatial weight s Perform element-by-element multiplication to obtain the reconstructed output features The specific mathematical model is:
[0055] Weight s =Sigmoid(Conv(Concat(Mp(x c ),Ap(x c ),B))) (4)
[0056]
[0057] Concat represents a concatenation operation.
[0058] The Enhanced CBAM module introduces a convolutional layer between the channel attention module and the spatial attention module to add a common bias. This common bias strengthens the dependence and constraints of the two attention modules on weight learning, thereby further improving the quality of features and enhancing the model's ability to refine features.
[0059] The existing Nonlocal captures long-range dependencies by calculating feature interactions in a global scope, which effectively improves the information capture capability and model performance in visual tasks. However, directly applying the Nonlocal module to medical image segmentation tasks lacks spatial information. Adding a spatial bias model can learn the feature differences and relationships between different positions, and can learn richer feature representations. Therefore, the present invention inserts learnable spatial features into the Nonlocal module to enhance the utilization of spatial position information and improve feature diversity. The EnNon-local module is designed as follows: Figure 7 As shown, it includes four 1×1 convolution blocks and a SoftMax normalization layer. The implementation method flowchart is as follows Figure 8 shown.
[0060] For local features x that require long-distance information encoding e4 , if D represents the number of channels C e4 , h×w represents its spatial resolution That is x e4 ∈R D×h×w . e4 Through three 1×1 convolution blocks and reshape transformation, Q, K, and V are obtained. According to the classic self-attention mechanism, the result of matrix multiplication of Q and K is normalized by SoftMax, and matrix multiplication with V is performed again. After reshaping transformation, the global feature f is obtained. g ∈R d×h×w In order to supplement the spatial information and enrich the diversity of features, the present invention uses the self-learning spatial feature f s ∈R d×h×w and the global feature f g After passing through the convolution layer, it is superimposed with the initial input x e4 Add element by element to get the output features Specifically, the self-learning spatial feature f s is a set of learnable parameters created and updated each time training. The mathematical model of the EnNon-local module is:
[0061]
[0062] f g =Reshape(SoftMax(Matmul(Q,K)),V))∈R d×h×w (7)
[0063]
[0064] Conv1 represents a 1×1 convolution, which is used to convert the number of channels. d represents the number of channels after conversion, Reshape represents the reshaping operation, and Matmul represents matrix multiplication.
[0065] By integrating learnable features into the EnNon-local module, it can better utilize spatial information and capture more diverse features. This enhancement enables it to effectively model long-range dependencies and strengthen the classification ability of the module.
[0066] Combination Figure 2 , the decoder converts the encoder's final output features After a 3×3 convolution block, it is used as the initial input x of the decoder d0 , and the target resolution is restored step by step through the cascade upsamplers. In the decoder design, the skip connection concatenates the high-resolution features from the encoder with the high-dimensional features of the decoder, and the decoder features are restored through the cascade upsamplers. Where di represents the i-th stage of the decoder, S di represents the downsampling multiple of the decoder at stage i, C di Indicates the number of channels in the i-th stage of the decoder.
[0067] Each cascade upsampling block consists of a transposed convolution layer with a stride of 2, a 3×3 convolution layer with a stride of 1, and a ReLU layer. The transposed convolution restores the output of the convolution layer to a specific resolution through reverse operations, and can also restore more complex spatial structures, such as corners, edges, and other specific features, thereby improving the accuracy of the model. Finally, the segmentation head of the 1×1 convolution is used to d4 Decode to get the segmentation mask.
[0068] During decoding, the cascade upsampler receives the high-dimensional features x that need to be upsampled di and the high-resolution features x from the encoder ei The output of the cascade upsampler is the input x to the next stage of the decoder. d(i+1) , its mathematical model is:
[0069]
[0070] Among them, TConv represents transposed convolution and Conv3 represents 3×3 convolution.
[0071] The network model of the present invention selects Dice loss and Cross Entropy loss as the loss of Atten-Nonlocal Unet, both of which have certain weights, specifically,
[0072] L=δL dice +(1-δ)L ce (10)
[0073]
[0074] Where L represents the total loss, L dice represents Dice loss, L ce represents CE loss. Variable y i and Denote the true label and predicted probability respectively. N denotes the total number of pixels and M is the number of categories. In this example, the value of δ is set to 0.5.
[0075] In order to verify the network performance of the present invention, two groups of experiments, module comparison and model comparison, are provided for specific description.
[0076] First, the Synapse multi-organ segmentation dataset was selected as the test object. The Synapse dataset contains 30 abdominal CT scans, a total of 3779 axial contrast-enhanced abdominal clinical CT images. Each CT image consists of 85 to 198 slices with a resolution of 512×512, including 8 abdominal organs (aorta, gallbladder, left kidney, right kidney, liver, pancreas, spleen, and stomach). The entire dataset is divided into 18 training cases (2212 slices) and 12 cases for verification.
[0077] All experiments were trained on a single A40 GPU. The SGD optimizer was used to optimize the network parameters, with an initial learning rate of 0.1, a momentum of 0.9, and a weight decay of 0.0001. In the experiment, the batch size was set to 24, the training round was 150 epochs, and the uniform input size of all data was 224×224. The rating indicators of segmentation accuracy are Dice similarity coefficient (DSC) and 95% Hausdorff distance (HD).
[0078] (1) Module comparison
[0079] To evaluate the effectiveness of the Enhanced CBAM module and the EnNon-local module, five experiments were conducted on the Synapse dataset. These experiments included using each module separately in the model, replacing them with residual CBAM and Non-local modules, and integrating the two modules into the model. The quantitative results of these experiments are summarized in Table 1, and the comparison of the actual labels generated by the network with the predicted probability heatmap is shown in Fig. 9 shown.
[0080] As can be seen from Table 1, when the Enhanced CBAM module and the EnNon-local module are used alone, their performance is better than the original module. Specifically, EnNon-local is better than the Non-local module for the segmentation of the left kidney, right kidney, gallbladder and spleen, and its segmentation accuracy is improved by 2.68%, 0.59%, 0.78% and 3.29%, respectively. In addition, it is reduced by 1.39 on the 95% HD standard, which shows that EnNon-local enhances the ability to capture long-range dependencies. In contrast, the Enhanced CBAM module is better at capturing global information and achieves improved segmentation accuracy on multiple organs. According to the DSC score, it is 0.29%, 0.73%, 3.5%, 1.52% and 0.47% higher than the CBAM module on the aorta, left kidney, right kidney, pancreas and stomach organs, respectively. Comparing the data of the Enhanced CBAM module and the EnNon-local module, the Enhanced CBAM module can improve the segmentation accuracy of small organs such as the gallbladder and pancreas, while the EnNon-local module maintains a high segmentation accuracy on large organs including the liver and stomach, revealing the complementary relationship between the two. The results of the Atten-Nonlocal Unet confirm that the network model of the present invention combines the advantages of two enhanced modules (Enhanced CBAM module and EnNon-local module), ensuring high accuracy in segmenting large organs and improving the segmentation effect of small organs.
[0081] Table 1 Module comparison experiment
[0082]
[0083] Fig. 9 Ground Truth is the real label. Fig. 9 The middle right image shows that the EnNon-local module can effectively identify misclassified or missed segmentation areas, demonstrating its superior positioning and classification capabilities. Fig. 9 Middle left image: Compared with the under-segmentation problem observed in the residual CBAM module, the Enhanced CBAM module is able to refine the organ edges and improve the correct segmentation at the boundaries.
[0084] (2) Model comparison
[0085] To demonstrate the effectiveness of the proposed model, Atten-Nonlocal Unet was compared with four network models, including TransUNet (see "Chen J. Lu Y. Yu Q. et al. Transunet: Transformers Make Strong Encoders For Medical Image Segmentation [J]. Arxiv Preprint Arxiv: 2102.04306, 2021."), TransNorm (see "Azad R. Al-Antary M T. Heidari M. et al. Transnorm: Transformer Provides A Strong Spatial Normalization Mechanism For A Deep Segmentation Model [J]. Ieee Access, 2022, 10: 108205-108215."), and HiFormer (see "Heidari M. Kazerouni A. Soltany M. et al. Hiformer: Hierarchical Multi-Scale Representations Using Transformers For Medical Image Segmentation [C] / / Proceedings Of The Ieee / Cvf Winter Conference On Applications Of ComputerVision.2023:6202-6212.") and TransCeption (for details, see "Azad R.Jia Y.Aghdam E K. et al. Enhancing Medical Image Segmentation With Transception: A Multi-Scale FeatureFusion Approach[J]. Arxiv Preprint Arxiv:2301.10847, 2023.").
[0086] The objective indicators of the experiment on the Synapse dataset are shown in Table 2. Atten-Nonlocal Unet achieved the best performance of 82.76% in the DSC standard and achieved the best results in the aorta, left kidney, right kidney, liver and spleen. Overall, it improved by 5.28% and 0.52% compared with the TransUNet and TransCeption models respectively. At the same time, it ranked second in the HD standard, only 0.69 higher than the HiFormer model. Fig.10 Analysis shows that Atten-Nonlocal Unet can refine the segmentation boundary of the liver organ, and can correctly identify and segment the target area in response to the missing stomach segmentation in other models. Fig.10 GT represents the true label.
[0087] Table 2 Synapse experimental results
[0088]
[0089] If the implementation method for medical image segmentation described in the embodiment of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, server or network device, etc.) to perform all or part of the steps of the image reconstruction method described in the embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard drives, ROM, RAM, magnetic disks or optical disks.
[0090] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A network suitable for medical image segmentation, characterized in that: Including encoder and decoder; The encoder is based on ResNet-50, and encodes a given medical image through four serial stages. After each stage, the feature resolution is halved. Each stage consists of a single Enhanced CBAM module and a continuous Residual CNN layer to extract local features with global information, and then encode the local features into features with long-distance dependencies through an EnNon-local module. The decoder uses the final output features of the encoder as the initial input of the decoder after passing through a 3×3 convolution block, and concatenates the high-resolution features from the encoder with the high-dimensional features of the decoder through jump connections, and restores the target resolution step by step through cascade upsamplers.
2. The network suitable for medical image segmentation according to claim 1, characterized in that: The EnhancedCBAM module includes a biased channel attention module, a biased spatial attention module, and a residual connection layer. First, the input feature is compressed into a biased feature B through a 1×1 convolution block, and then the input feature and the biased feature B are input into the biased channel attention module to reconstruct the channel dimension feature. Then, the channel dimension reconstructed feature and the biased feature are input into the biased spatial attention module to reconstruct the spatial dimension feature. Finally, the spatial dimension reconstructed feature is residually connected with the input feature to output the global feature.
3. The network suitable for medical image segmentation according to claim 2, characterized in that: In the encoder, for the encoded feature x at a certain stage ei (i∈[1~4]), the calculation formula of the output feature f of the Enhanced CBAM module is: f=SA(CA(x ei ,B),B)+x ei (1) Where CA represents biased channel attention, SA represents biased spatial attention, and B represents biased features.
4. The network suitable for medical image segmentation according to claim 3, characterized in that: The bias channel attention module receives the input feature x ei And bias feature B, first add the two input features element by element, perform maximum pooling and average pooling to obtain parameters, then reconstruct the initially obtained parameters through shared MLP, and the sum of the reconstructed parameters is normalized by Sigmoid to obtain the final channel weight Weight c , the final output feature x c By channel weight Weight c With input feature x ei The specific mathematical model is obtained by multiplying each element: Weight c =Sigmoid(M(Mp(x ei +B))+M(Ap(x ei +B))) (2) Where M represents a multi-layer perceptron, Mp represents maximum pooling, and Ap represents average pooling.
5. The network suitable for medical image segmentation according to claim 3, characterized in that: The biased spatial attention module receives the output x from the biased channel attention module c and bias feature B as input, x c The features obtained by maximum pooling and average pooling are concatenated with the bias features, and the obtained mixed features are convolved and normalized with Sigmoid to obtain the spatial weight Weight s , input feature x c and spatial weight s Perform element-by-element multiplication to obtain the reconstructed output feature x cs , the specific mathematical model is: Weight s =Sigmoid(Conv(Concat(Mp(x c ),Ap(x c ),B))) (4) Concat represents a concatenation operation.
6. The network suitable for medical image segmentation according to claim 1, characterized in that: The EnNon-local module includes four 1×1 convolution blocks and a SoftMax normalization layer. For local features that require long-distance information encoding, they are first passed through three 1×1 convolution blocks and reshaped to obtain Q, K, and V; according to the classic self-attention mechanism, the result of matrix multiplication of Q and K is processed by SoftMax normalization, and matrix multiplication with V is performed again, and global features are obtained after reshaping transformation; create learnable spatial features, add the learnable spatial features to the global features element by element, and finally pass through a 1×1 convolution layer and add them element by element to the initial input features to obtain output features with long-distance dependencies.
7. The network suitable for medical image segmentation according to claim 1, characterized in that: The cascade upsampler consists of a transposed convolutional layer with a stride of 2, a 3×3 convolutional layer with a stride of 1, and a ReLU layer.
8. A method for implementing medical image segmentation, implemented using the network for medical image segmentation as claimed in any one of claims 1 to 7, characterized in that: The following steps are involved: Taking ResNet-50 as the main encoder, a given medical image is first encoded through four serial stages. After each stage, the feature resolution is halved. Each stage consists of a single Enhanced CBAM module and a continuous Residual CNN layer to extract local features with global information. The local features are then encoded into features with long-distance dependencies through the EnNon-local module. The final output features of the encoder are passed through a 3×3 convolution block as the initial input of the decoder, and the high-resolution features from the encoder are concatenated with the high-dimensional features of the decoder through a skip connection, and are gradually restored to the target resolution through a cascade upsampler; Finally, a 1×1 convolutional segmentation head is used to decode the features finally decoded by the decoder to obtain a segmentation mask.
9. A device for implementing medical image segmentation, characterized in that: The invention comprises a memory and a processor, wherein the memory is used to store at least one program, and the processor is used to load the at least one program to execute the method according to any one of claims 1 to 7.
10. A storage medium storing instructions executable by a processor, characterized in that: The processor-executable instructions are used to perform the method according to any one of claims 1 to 7 when executed by the processor.
Citation Information
Cited By
Phase evolution analysis method of battery SEI membrane, electronic equipment and storage medium
CN120411103A
Medical image restoration method and device, computer equipment and storage medium
CN120707437A
