Brain Glioma Image Segmentation Method Based on Deep Learning

Abstract

The present invention discloses a method for brain glioma image segmentation based on deep learning, which relates to the technical field of image processing. The method includes the following steps: obtaining multi-modal MRI images and preprocessing them to obtain a multi-channel input image tensor; inputting the multi-channel input image tensor into a feature extraction module to obtain a first loss; establishing an improved UNet model, and training the improved UNet model with the multi-channel input image tensor to obtain a second loss, and adjusting the parameters of the improved UNet model by combining the first loss and the second loss to obtain a trained model; the improved UNet model is based on the UNet model, an improved hypergraph convolution module is introduced at the deepest layer of the UNet model, a multi-scale dynamic weighted feature pyramid module is introduced in the skip connection stage, and a dual attention mechanism is introduced in the feature recovery stage; taking the brain glioma image to be segmented and segmenting it using the trained model. The method of the present invention greatly improves the segmentation accuracy and precision.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and particularly to a method for segmenting glioma images based on deep learning. Background Art

[0002] Medical image segmentation technology has been increasingly widely used in clinical auxiliary diagnosis, and has played a key role especially in the detection and localization of brain tumors (such as gliomas). Glioma is one of the most common malignant tumors in the nervous system, with characteristics such as strong invasiveness, high heterogeneity, and blurred boundaries, seriously affecting the survival rate and quality of life of patients. Magnetic Resonance Imaging (MRI) has become an important means for the diagnosis and monitoring of brain tumors. Traditional methods mainly rely on empirical judgment. However, in primary medical institutions, there is a lack of professional doctor resources, and there are three major pain points in the imaging diagnosis of gliomas, namely, high misdiagnosis rate, long average diagnosis time, and low patient survival rate.

[0003] The current mainstream methods for segmenting glioma images mainly rely on architectures based on Deep Convolutional Neural Networks (DCNN), and typical representatives include UNet, VNet, SegNet, etc. These methods generally adopt an Encoder-Decoder structure, which can extract image features and generate segmentation masks, achieving good segmentation performance. The structure represented by UNet has been widely used in medical image segmentation tasks. It mainly extracts semantic features from the input image layer by layer, and fuses shallow features in the decoding stage to achieve more refined boundary recognition. However, these methods still have the following key problems in practical applications: 1. Limited receptive field and insufficient long-range dependence modeling ability; 2. Weak multi-modal feature fusion ability and poor adaptability to clinical images; 3. Limited multi-scale expression ability, weak structural layer perception; 4. Weak ability to focus on key regions and rough boundary segmentation; 5. Lack of graph relationship modeling at the structural level. Summary of the Invention

[0004] In order to solve at least one of the above problems, this application provides a method for segmenting glioma images based on deep learning. By improving the UNet model, it solves the problems of insufficient receptive field, weak multi-modal fusion ability, poor adaptability to clinical images, limited multi-scale expression ability, and weak ability to focus on key regions in the existing segmentation methods.

[0005] In order to achieve the above object, this application provides a method for segmenting glioma images based on deep learning, including the following steps:

[0006] S1. Obtain multi-modal MRI images and preprocess them to obtain a multi-channel input image tensor;

[0007] S2. Input the multi-channel input image tensor into the feature extraction module for training to obtain the first loss L NGF ; The feature extraction module includes a Gaussian mixture model, dynamically adjusts the neighborhood range of data points in the Gaussian mixture model through kernel density estimation, provides initial parameters for the Gaussian mixture model through K-means++, and optimizes the Gaussian mixture model through sparse constraints;

[0008] S3. Establish an improved UNet model, and use the multi-channel input image tensor to train the improved UNet model to obtain the second loss L net , and obtain the total loss L total = L net + γ·L NGF , and adjust the parameters of the improved UNet model based on the total loss to obtain the trained model; The improved UNet model is based on the UNet model, introduces an improved hypergraph convolution module in the deepest layer of the UNet model, introduces a multi-scale dynamic weighted feature pyramid module in the skip connection stage, and introduces a dual attention mechanism in the feature recovery stage;

[0009] S4. Take the glioma image to be segmented and segment it using the trained model in S3.

[0010] The beneficial effects of the present invention are as follows: Compared with the prior art, the present invention has stronger long-range dependence modeling ability, multi-modal collaborative expression ability, multi-scale feature expression and fusion ability, ability to construct a full-structure hierarchical semantic modeling system, and key region attention ability. Based on the traditional UNet model, the present invention introduces a feature extraction module, an improved hypergraph convolution module, a multi-scale dynamic weighted feature pyramid module, and a dual attention mechanism into the model, greatly improving the segmentation accuracy and precision, and solving the problems of insufficient receptive field, weak multi-modal fusion ability, poor adaptability to clinical images, limited multi-scale expression ability, and weak key region attention ability in the prior art segmentation methods. Specific embodiments

[0011] The following will clearly describe the technical solutions of the present application in conjunction with the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, other embodiments obtained by those of ordinary skill in the art without creative efforts all fall within the protection scope of the present application.

[0012] A method for segmenting glioma images based on deep learning includes the following steps:

[0013] S1. Obtain multi-modal MRI images and preprocess them to obtain a multi-channel input image tensor;

[0014] In this step, the preprocessing includes the following sub-steps: Take multi-modal MRI images, where the multi-modal includes T1, T1ce, T2, FLAIR, etc., adjust them to a unified input format, such as a 256×256 format, and splice them on the channel dimension to form a multi-channel input image tensor \(X_0\in\mathbb{R}\) 256×256×C , where \(C\) represents the number of channels.

[0015] S2. Input the multi-channel input image tensor into the feature extraction module for training and obtain the first loss \(L\) NGF ; The feature extraction module includes a Gaussian mixture model, dynamically adjusts the neighborhood range of data points in the Gaussian mixture model through kernel density estimation, provides initial parameters for the Gaussian mixture model through K-means++, and optimizes the Gaussian mixture model through sparse constraints;

[0016] In this step, it is a further processing operation on the MRI image. It mainly uses the feature extraction module to perform global feature extraction on the multi-channel input image tensor obtained in S1. During the extraction process, the loss \(L\) is obtained according to the loss function of the Gaussian mixture model. NGF .

[0017] The feature extraction module in this step uses the Gaussian mixture model (GMM) as the main extraction tool, and then improves it with multiple modules based on the characteristics of GMM.

[0018] In GMM, the neighborhood range of each data point is fixed. However, for data, the data distribution densities in different regions are different. Therefore, the traditional GMM has certain limitations. For this reason, in this embodiment, through kernel density estimation, the neighborhood range is dynamically determined for GMM, which can better adapt to the characteristics of uneven data distribution, thereby improving the model's fitting ability to data.

[0019] At the same time, for GMM, its performance highly depends on the selection of initial parameters. The traditional method for selecting initial parameters usually randomly assigns means and variances. Such a method is prone to causing the algorithm to fall into a local optimal solution, thereby affecting the final clustering effect. To improve the quality of the initial distribution, in this step, the K-means++ algorithm is used to provide the initial mean and covariance matrix, and then the initial clustering centers are intelligently selected, so that the center points of each class are far away from each other, thereby providing more reasonable initial means and covariances for GMM.

[0020] During the training process of the GMM, there is a high probability of generating redundant Gaussian distributions, resulting in overfitting of the model or reduced computational efficiency. To solve this problem, this step introduces a sparse constraint with L1 regularization, which can automatically reduce the weights of unnecessary Gaussian components and make the model more compact.

[0021] In this step, the operation of determining the neighborhood range for the GMM through kernel density estimation is as follows:

[0022] First, flatten the gray values of all pixels in the MRI image and arrange them in a one-dimensional vector according to the spatial position to form an image gray vector. , for the image gray vector calculate the local density of each data point x i : , where is the kernel density estimate value of the i-th pixel; is the Gaussian kernel function; h is the bandwidth parameter; N is the total number of pixels in the image; ‖x i -x j ‖ represents the Euclidean distance of the gray values between pixel i and j.

[0023] Subsequently, calculate the dynamic neighborhood radius through the above kernel density estimate values: , where r i represents the dynamic neighborhood radius of the i-th pixel; r base is the base radius; α represents the adjustment factor; is the maximum value of all pixel kernel density estimate values.

[0024] The dynamic neighborhood radius obtained through the above calculation can dynamically adjust the neighborhood range of the GMM.

[0025] In this step, the operation of providing the initial mean and covariance matrix through the K-means++ algorithm is as follows:

[0026] 1. Selection of the initial center: Randomly select a center, and the subsequent centers are selected according to the probability , where P(x i ) is the probability of selecting pixel i as the next clustering center, D(x i ) is the distance from pixel i to the current nearest clustering center, is the sum of the squares of the distances from all pixels to their nearest clustering centers.

[0027] 2. Perform K-means++ clustering operation on all pixels to obtain multiple clusters.

[0028] 3. Initialize the parameters of GMM: According to the clustering results of K-means++, initialize the mean, covariance, and mixing weights of GMM: , π k is the mixing weight of the k-th Gaussian component; N k is the number of pixels in the k-th cluster C k ; N is the total number of pixels in the image; μ k is the central pixel point of the k-th cluster, which is essentially the statistical mean of all pixel features in this cluster and represents the typical features of this cluster; is the covariance matrix of the k-th Gaussian component, C k is the set of pixels corresponding to the k-th cluster.

[0029] Through the above operations, not only can the convergence speed of the GMM algorithm be accelerated, but also the robustness and clustering accuracy of the model can be significantly improved.

[0030] In this step, the operation of optimizing GMM through sparse constraints is as follows:

[0031] Regard each pixel point as an independent sample, construct a grayscale vector containing multiple samples, assume that the image grayscale distribution can be generated by mixing K Gaussian distributions, and use the constrained EM algorithm to estimate the parameters of each distribution:

[0032] The E step is to estimate the posterior probability: , where γ ik is the probability that the sample x i belongs to the k-th Gaussian component; is the probability density function of the Gaussian distribution.

[0033] The M step (update parameters): Introduce the L1 regularization term to sparsify the diagonal elements of the covariance matrix:

[0034] ;

[0035] ;

[0036] where λ is the L1 regularization coefficient; diag(I) is the diagonal matrix of the identity matrix, which is used for the sparse constraint of the covariance matrix.

[0037] Under this operation, the response coefficients of all pixel points under each Gaussian component are organized into a multi-channel response map, which is used as a supplementary input feature for the downstream segmentation network, so as to guide the model to focus on the internal structure distribution of the image, enhance the modeling ability of the image tissue type distribution and the gray-scale offset of the lesion area, and effectively improve the distinguishability of the tumor boundary in segmentation.

[0038] S3. Establish an improved UNet model, and train the improved UNet model using the multi-channel input image tensor to obtain the second loss L net , and obtain the total loss L based on the first loss and the second loss total =L net +γ·L NGF , and adjust the parameters of the improved UNet model based on the total loss, thereby obtaining the trained model; the improved UNet model is based on the UNet model, an improved hypergraph convolution module is introduced in the deepest layer of the UNet model, a multi-scale dynamic weighted feature pyramid module is introduced in the skip connection stage, and a dual attention mechanism is introduced in the feature recovery stage

[0039] In this step, based on the UNet model as the basic model, an improved hypergraph convolution module, a multi-scale dynamic weighted feature pyramid module, and a dual attention mechanism are inserted at different positions thereof to further enhance the performance of the model

[0040] In this embodiment, the adopted UNet model consists of an encoder and a decoder. The size of the input image is 572×572. In its encoder part, after the input image passes through the first convolutional layer (conv 3x3, ReLU), the size becomes 568×568. Then, the image enters the max pooling layer (max pool 2x2), and the size is reduced to 284×284. With each layer of convolution and pooling operations, the size of the image gradually decreases, and the number of channels continuously increases. When reaching the deepest layer, the image size becomes 56×56, and the number of channels of the feature map increases to 1024. In its decoder part, the image gradually restores the spatial resolution through upsampling (up-conv 2x2). After several upsamplings, the size of the image is gradually restored, and the final output size is 388×388. Skip connection (copy and crop) transmits information between the encoder and the decoder to help the decoder utilize the high-resolution features of the earlier layer. Finally, a 1x1 convolutional layer (conv 1x1) is used to generate the final segmentation map

[0041] In the output stage, the model uses a 1×1 convolution to map the finally decoded multi-channel feature map to a prediction map of the number of categories (such as background, tumor, tumor core, etc.) , where K is the number of categories. By taking the category to which each pixel prediction vector belongs with the highest probability, the final segmentation mask can be generated. To ensure the performance of the model, this project uses a comprehensive loss function that combines the loss of the improved UNet model and the loss of the feature extraction module. Its mathematical expression is as follows

[0042] ;

[0043] where, L totalrepresents the total loss; γ is a parameter between the network and the feature extraction module, usually taking values from 0.3 to 0.8. After calculating the total loss, backpropagation is performed on it to control the parameters in the UNet model. For the above loss function, if L NGF increases (the effect of the feature extraction module is poor), the gradient of the total loss with respect to the UNet parameters will contain a stronger "correction signal", prompting UNet to adjust the parameters to adapt to the low-quality features output by the feature extraction module; if L NGF decreases, UNet can focus on optimizing the segmentation head by using the high-quality features provided by the feature extraction module, accelerating convergence. If it is finally basically stable, it means that the model algorithm has converged, and the training can be stopped.

[0044] However, the existing UNet models perform poorly in terms of receptive field, multi-scale expression ability, and the ability to focus on key regions. Therefore, in this embodiment, multiple modules are introduced to improve the UNet model.

[0045] To address the problem that the UNet model performs poorly in terms of receptive field, an improved hypergraph convolution module is introduced in this embodiment. In this embodiment, hyperedges are constructed by introducing feature similarity. Subsequently, multiple hyperedges form a hypergraph structure, and then a hypergraph convolution module is constructed. At the same time, residual connections and Jumping Knowledge mechanisms are introduced in the feature aggregation process of the hypergraph convolution module. The insertion position of the improved hypergraph convolution module is at the deepest layer of the encoder of the UNet model, that is, the bottom of the encoder.

[0046] The method for constructing hyperedges is as follows:

[0047] Calculate the similarity between nodes: , where S ij represents the feature similarity between pixel i and pixel j, ‖·‖2 represents the L2 norm, σ scale represents the scale parameter, f i and f j represent the feature vectors of the i-th pixel and the j-th pixel respectively;

[0048] For any node, select the k i nodes with the highest similarity to it to construct hyperedges; where, , where k base is the basic connection number; β is the adjustment factor; S is the similarity matrix between all pixels, S ij ∈S.

[0049] In the feature aggregation process, the node feature update formula of traditional hypergraph convolution is:

[0050] , where H （l）is the node feature of the current layer; is the hyperedge indicator matrix; D v 、D e are the node matrix and the hyperedge degree matrix respectively; σ act is the non-linear activation function; W （l） represents the learnable weight matrix of the l-th layer. However, such feature update easily makes the hypergraph convolution fall into the over-smoothing problem. To solve this problem, this embodiment introduces a residual connection Jumping Knowledge mechanism to aggregate multi-layer features: , where X out represents the fused output feature; W f represents the fusion weight matrix; L layer is the total number of layers of the hypergraph convolution; Concat(·) is to concatenate features along the channel dimension; X （l） represents the feature map of the l-th layer.

[0051] The above is the structure of the hypergraph convolution network introduced in the UNet model.

[0052] Through the collaborative fusion method of "Gaussian distribution modeling + hypergraph structure modeling", the model can simultaneously model the pixel-level gray distribution law and the regional-level semantic topological structure, so as to deeply depict the evolution relationship between different functional regions (such as necrosis region, enhancement region, tumor margin region) inside glioma, and improve the model's understanding ability of tumor heterogeneity and progression trend.

[0053] To solve the problem that the UNet model performs poorly in multi-scale expression ability, this embodiment introduces a multi-scale dynamic weighted feature pyramid module. Among them, the insertion position of the multi-scale dynamic weighted feature pyramid module is at the skip connection stage of the UNet model. In this embodiment, by learning variable weights, multi-layer features are fused, and the specific operations are as follows:

[0054] For the feature map F of the i-th scale i , use 1×1 convolution + global pooling to obtain the weight vector, and then normalize it through Softmax: , where α i represents the normalized weight of the feature map F i ; w i represents the original weight of the feature map F i ; n is the number of scales of the feature pyramid;

[0055] Perform weighted fusion on the feature map: , where F fusedIt represents the fused feature map. This fusion result is then concatenated or element-wise added to the upsampled features of the corresponding layer in the decoder, replacing the fixed fusion method of direct skip connections in the traditional UNet. This mechanism enables the model to dynamically adjust the contribution ratio of features at different scales: high-level features enhance the semantic consistency of the tumor region, and low-level features retain boundary details, thus significantly improving the segmentation accuracy, especially when dealing with the fuzzy boundaries and multi-scale structures (such as the core area and the edema area) of gliomas. Finally, the fused features are used as the input to the decoder, and the key regions are further refined through the dual attention mechanism to achieve end-to-end accurate segmentation.

[0056] This module realizes the importance modeling of feature information at different scales, enabling the model to have stronger hierarchical fusion ability and significantly improving the ability to depict segmentation boundaries and small target regions. By introducing an improved hypergraph convolution modeling module, based on high-order hyperedges, the structural correlation of multiple similar regions in the image is modeled, effectively expanding the model's receptive field and enhancing the global perception and consistency modeling ability for complex topological structures (such as cross-lobe diffusion and heterogeneous lesions), significantly improving problems such as boundary breaks and structural blurring.

[0057] To solve the problem that the UNet model performs poorly in the ability to focus on key regions, this embodiment introduces a dual attention mechanism, where the insertion position of the dual attention mechanism is in the feature recovery stage of the UNet model.

[0058] In this embodiment, the dual attention mechanism includes two branches: channel attention and spatial attention. By introducing the dual attention mechanism, the expression ability of the model in significant regions can be enhanced. Among them, the dual attention mechanism includes channel attention and spatial attention. Channel attention emphasizes the importance between different channels, and spatial attention perceives the saliency of local regions.

[0059] The extraction methods of channel attention and spatial attention are as follows: 、 , where Mc and Ms represent the channel attention weight and the spatial attention weight respectively; σ act1 and σ act represent the Sigmoid activation function and the non-linear activation function respectively; W1 and W2 are the weight matrices of the fully connected layer; ReLU(·) represents the activation function; f 7×7 represents a 7×7 convolution operation; GAP(F fused ) represents the global average pooling of channel attention for F fused ; AvgPool(F fused ) represents the global average pooling of spatial attention for F fused ; MaxPool(F fused ) represents the global max pooling of F fused .

[0060] The output after comprehensive attention is:

[0061]

[0062] Among them, is the feature map weighted by double attention.

[0063] The channel-spatial double attention mechanism significantly enhances the model's ability to focus on discriminative regions (such as tumor margins, lesion transition zones), suppresses background interference, highlights the responses of key regions, reduces problems such as boundary blurring and mis-segmentation, and improves the clinical usability of the segmentation results.

[0064] After establishing the above improved UNet model, it is also necessary to train the model using the multi-channel input image tensor obtained by S1. The training process is the same as that of the conventional model and will not be elaborated here.

[0065] S4. Take the glioma image to be segmented and use the trained model of S3 to segment it. In this step, the glioma image processed by S1 can be segmented using the trained improved UNet model.

[0066] To further illustrate the advantages of the embodiments of the present invention, specific examples are given below.

[0067] In this test example, the dataset comes from lgg-mri-segmentation. This dataset contains 7,858.tif files, including 3,929 MRI images with a size of 256×256 and 3,929 corresponding segmentation masks, which are divided into a training set and a validation set in a ratio of 8:2. The training set has a total of 6,286 images, and the validation set has a total of 1,572 images.

[0068] Four metrics, namely Dice coefficient (Dice), intersection over union (IoU), precision, and recall, are used to evaluate the performance of different methods on this dataset.

[0069] At the same time, to illustrate the superiority of the method of the embodiments of the present invention, multiple existing different methods are uniformly tested, and ablation experiments are carried out on the method of the embodiments of the present invention. The final results are shown in Table 1 and Table 2.

[0070] Table 1 Test results of different methods

[0071]

[0072] Table 2 Test results of ablation experiments

[0073]

[0074] As can be seen from Table 1, the method of the embodiment of the present invention is significantly superior to the traditional UNet and its variants in all indicators.

[0075] As can be seen from Table 2, each improvement module in the method of the embodiment of the present invention can greatly improve the effect of the embodiment of the present invention. In particular, the improved Gaussian mixture model shows a relatively key application. This module enhances the multi-scale features of the tumor core, edema area and normal brain tissue through differentiation, and significantly improves the boundary discrimination of glioma sub-regions (such as enhanced tumor, necrotic core). Especially when dealing with the problems of blurred tumor margins and high density heterogeneity in T1 / T2-FLAIR multi-modal MRI, this module effectively enhances the model's ability to identify tiny tumor infiltration foci and low-contrast edema areas by dynamically weighted fusion of semantic features at different levels, providing reliable technical support for the clinical accurate positioning of the tumor range.

[0076] It should be noted that those skilled in the art will easily think of other implementation schemes of the present application after considering the specification and practicing the application disclosed herein. The present application is intended to cover any variations, uses or adaptations of the present application, which follow the general principles of the present application and include the common general knowledge or conventional technical means in the technical field not disclosed in the present application.

[0077] It should be understood that the present application is not limited to the exact structure already described and shown in the drawings, and various modifications and changes can be made without departing from its scope. The true scope is indicated by the present application.

Claims

1. A method for glioma image segmentation based on deep learning, characterized in that, Including: S1. Obtain multi-modal MRI images and preprocess them to obtain a multi-channel input image tensor; S2. Input the multi-channel input image tensor into the feature extraction module for training to obtain the first loss L NGF ; The feature extraction module includes a Gaussian mixture model, dynamically adjusts the neighborhood range of data points in the Gaussian mixture model through kernel density estimation, provides initial parameters for the Gaussian mixture model through K-means++, and optimizes the Gaussian mixture model through sparse constraints; S3. Establish an improved UNet model, and train the improved UNet model using the multi-channel input image tensor to obtain the second loss L net , and obtain the total loss L based on the first loss and the second loss total =L net +γ·L NGF , and adjust the parameters of the improved UNet model based on the total loss, and then obtain the trained model; the improved UNet model is based on the UNet model, an improved hypergraph convolution module is introduced in the deepest layer of the UNet model, a multi-scale dynamic weighted feature pyramid module is introduced in the skip connection stage, and a dual attention mechanism is introduced in the feature recovery stage; S4. Take the glioma image to be segmented and segment it using the trained model in S3.

2. The method according to claim 1, characterized in that, In S1, the preprocessing includes the following sub-steps: Take the multi-modal MRI images, adjust them to a unified input format, and splice them in the channel dimension to form a multi-channel input image tensor.

3. The method according to claim 1, characterized in that, In S2, dynamically adjusting the neighborhood range of data points in the Gaussian mixture model through kernel density estimation includes the following steps: Kernel density estimation: , where is the kernel density estimation value of the i-th pixel; is the Gaussian kernel function; h is the bandwidth parameter; N is the total number of pixels in the image; ‖x i -x j ‖ represents the Euclidean distance of the gray values between pixels i and j; Dynamic neighborhood radius: , where r i represents the dynamic neighborhood radius of the i-th pixel; r base is the base radius; α represents the adjustment factor; is the maximum value of the kernel density estimates of all pixels; Dynamically adjust the neighborhood of the Gaussian mixture model according to the calculation result of the dynamic neighborhood radius.

4. The method according to claim 1, wherein In S2, the method of using K-means++ to provide initial parameters for the Gaussian mixture model is as follows: Randomly select a center, and subsequent centers are selected according to the probability , where P(x i ) is the probability of selecting pixel i as the next cluster center; D(x i ) is the distance from pixel i to the current nearest cluster center; is the sum of the squares of the distances from all pixels to their nearest cluster centers; Perform K-means++ clustering operation on all pixels to obtain multiple clusters; Initialize the mean, covariance, and mixing weights of the Gaussian mixture model: ， ， ， where, π k is the mixing weight of the k-th Gaussian component; N k is the number of pixels in the k-th cluster C k ; N is the total number of pixels in the image; μ k is the central pixel point of the k-th cluster; is the covariance matrix of the k-th Gaussian component, and C k is the set of pixels corresponding to the k-th cluster.

5. The method according to claim 1, characterized in that The establishment method of the improved hypergraph convolution module is as follows: Construct hyperedges based on feature similarity: Multiple hyperedges form a hypergraph structure, and then construct a hypergraph convolution module. At the same time, introduce residual connections and Jumping Knowledge mechanisms in the feature aggregation process of the hypergraph convolution module.

6. The method according to claim 5, wherein The construction method of the hyperedges includes the following sub-steps: Calculate the similarity between nodes, , where S ij represents the feature similarity between pixel i and pixel j, and σ scale represents the scale parameter, and f i and f j represent the feature vectors of the i-th pixel and the j-th pixel respectively; ‖·‖2 represents the L2 norm; For any node, select the k i nodes with the highest similarity to it to construct a hyperedge; where , where k base is the basic number of connections; β is the adjustment factor; S is the similarity matrix between all pixels, and S ij ∈ S.

7. The method according to claim 1, characterized in that The establishment method of the multi-scale dynamic weighted feature pyramid module includes the following steps: For the feature map F at the i-th scale i , calculate its weight vector and normalize it: , where α i represents the normalized weight of the feature map F i ; w i represents the original weight of the feature map F i ; n is the number of scales of the feature pyramid Perform weighted fusion on the feature maps: , where F fused represents the fused feature map; Splice or add the fused feature map with the upsampled features of the corresponding layer of the decoder of the UNet model to replace the fixed fusion method of the skip connection of the UNet model.

8. The method according to claim 1, wherein The introduction of the dual attention mechanism includes the following steps: Calculate channel attention and spatial attention: , , where Mc and Ms represent the channel attention weight and the spatial attention weight respectively; σ act1 and σ act represent the Sigmoid activation function and the non-linear activation function respectively; W1 and W2 are the weight matrices of the fully connected layers; ReLU(·) represents the activation function; f 7×7 represents a 7×7 convolution operation; GAP(F fused ) represents the global average pooling of F fused by the channel attention, F fused represents the fused feature map; AvgPool(F fused ) represents the global average pooling of F fused by the spatial attention; MaxPool(F fused ) represents the global max pooling of F fused ; Output after comprehensive attention: , where is the feature map weighted by double attention.

Images