A decoding method for medical image segmentation with cross-layer feature refinement

Abstract

The present application relates to medical image processing, and relates to a decoding method for medical image segmentation with cross-layer feature refinement, the present application improves the TransUnet decoder, adopts a step-by-step upsampling reconstruction strategy in the decoding stage: cross-layer semantic fusion is carried out on the up-sampling feature and the corresponding jump connection feature, the jump connection feature is filtered through space-channel gating to suppress redundant information; the fusion feature is subjected to axial-channel double attention enhancement, the alignment and strengthening of structure details and boundary information are realized; through a multi-scale enhancement module, local texture and edge structure features under different receptive fields are extracted, restored to the target resolution and output the segmentation result. The present application improves the cross-layer feature fusion quality, enhances the segmentation performance of small structures and boundary regions, is suitable for binary or multi-class medical image segmentation tasks, and can provide more reliable segmentation results for clinical image analysis, and provide support for auxiliary diagnosis and treatment plan making.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, and more specifically to a decoding method for medical image segmentation with cross-layer feature refinement. Background Technology

[0002] Medical image segmentation can accurately delineate the contours of lesions and organs, serving as a crucial foundation for computer-aided diagnosis, personalized treatment planning, and efficacy evaluation. However, traditional segmentation methods heavily rely on extensive manual annotation and expert experience, which is time-consuming, labor-intensive, and susceptible to inter-observer variability, making it difficult to guarantee the stability and consistency of results. In recent years, deep learning-based automatic segmentation technology has rapidly developed, exhibiting significant advantages in efficiency, accuracy, and repeatability compared to manual delineation, and has become a major research direction in this field.

[0003] Medical image segmentation methods have evolved from early local modeling approaches based on convolutional neural networks, represented by U-Net and its improved structures, to hybrid encoder frameworks that introduce Transformers and new paradigms that utilize state-space models to achieve efficient long-range dependency modeling. However, existing methods generally lack cross-layer semantic alignment and redundant noise suppression mechanisms in skip connection processing, making fine structures and boundary information susceptible to interference. Therefore, it is necessary to optimize cross-layer feature fusion before fusing with the decoder and combine multi-scale and attention enhancement to improve the reliability and accuracy of segmentation. Summary of the Invention

[0004] The technical problem to be solved by this invention is that existing medical image segmentation models lack effective semantic alignment and redundancy suppression mechanisms during cross-layer feature fusion.

[0005] To achieve the above-mentioned objective, the present invention provides a decoding method for medical image segmentation with cross-layer feature refinement, comprising the following steps: S1 Dataset Construction: Obtain the original medical image dataset and its corresponding segmentation annotations to form a medical image segmentation dataset. For the dataset Preprocessing was performed, and the dataset was divided into training sets according to a set ratio. and test set ; S2 Network Model Construction: A medical image segmentation network model based on TransUnet is constructed. The network model includes an encoder, an improved decoder, and skip connections. The encoder is a hybrid structure of convolutional neural network (CNN) and Transformer, used to extract multi-scale semantic features.

[0006] The improved decoder comprises four decoding layers. The first three layers sequentially include a spatial-channel gating module, an axial-channel dual attention module, and a multi-scale enhancement module. The fourth layer sequentially includes an axial-channel dual attention module and a multi-scale enhancement module. An upsampling module is positioned between adjacent decoding layers. The spatial-channel gating module generates gating weights based on the output features of the deeper upsampling modules and uses these weights to perform semantic filtering on the skip connection features of the corresponding layer. These weights are then fused with the upsampled output features to obtain fused features. The axial-channel dual attention module is used to align and enhance the structural details and boundary information of the fused features. The multi-scale enhancement module extracts local texture and edge structure information under different receptive fields, outputs enhanced features, and uses them as input features for the next decoding layer. The upsampling module refines the enhanced features and achieves channel alignment with the skip connection features of the previous layer through channel transformation.

[0007] The skip connection is used to pass the feature map of the corresponding layer in the encoder to the corresponding layer in the decoder for feature fusion.

[0008] S3 Model Training: Using the Training Set A TransUnet-based medical image segmentation network model was trained using a joint loss function consisting of cross-entropy loss and Dice loss, and the model parameters were iteratively updated through backpropagation.

[0009] S4 Model Termination and Weight Saving: Repeat step 3) until the preset stopping condition is met, and obtain the trained model weights. The stopping condition is reaching a preset number of training rounds.

[0010] S5 Medical Image Segmentation: Input the medical image to be segmented into the training weights... The medical image segmentation network model based on TransUnet outputs a medical image segmentation map.

[0011] Furthermore, step 1) above includes: S1.1 Dataset Preprocessing is performed, which includes one or more of size normalization, cropping, and data augmentation. S.1.2 The dataset The training set was divided into two parts at a ratio of 7:3. and test set .

[0012] Furthermore, the working process of the above-mentioned upsampling module includes: Input features of the current decoding level Upsampling is performed using bilinear interpolation; the upsampled features are then thinned by convolution, and channel alignment with the skip connection features is achieved through channel transformation to obtain the upsampled aligned features. The calculation formula is as follows: ; In the formula, For a convolution with a kernel of 1, for Activation function For batch normalization, For a convolution with a kernel of 3, This indicates an upsampling operation.

[0013] Furthermore, the operation process of the aforementioned space-channel gating module includes: 2.1.1) Obtain the skip connection features of the corresponding layer of the decoder. , jump connection features Alignment features with upsampling Separate depthwise separable convolutions are performed, followed by element-wise addition and fusion. The resulting fused features are then activated by ReLU. The calculation formula is as follows: ; In the formula, This indicates a depthwise separable convolution with a kernel size of 3.

[0014] 2.1.2) Fusion Features Spatial context relationship modeling is performed, and spatial gating coefficients are obtained by normalization using the Sigmoid function. The calculation formula is as follows: ; In the formula, This is the Sigmoid function.

[0015] 2.1.3) Fusion Features Perform global average pooling operation The statistical characteristics of each channel are obtained by compressing spatial information, and the calculation formula is as follows: ; In the formula, For the first One sample, For the first One channel, The height of the feature map, The width of the feature map is used; the global average pooling result is then subjected to dimensionality reduction, dimensionality increase, and Sigmoid normalization to obtain the channel gating coefficients. The calculation formula is as follows: .

[0016] 2.1.4) Spatial gating coefficient With channel gating coefficient Perform element-wise multiplication to generate the final joint gating weights. Gating weights Features of skip connections Element-wise multiplication yields the output features of the space-channel gating module. The calculation formula is as follows: ; ; In the formula, This indicates element-wise multiplication.

[0017] Furthermore, the aforementioned axial-channel dual attention process includes: 2.2.1) Output characteristics of the space-channel gating module Upsampling alignment features with the corresponding decoding level The input features are then fused. Input features Input the spatial attention path and the channel attention path respectively.

[0018] 2.2.2) Spatial attention path will input features The channel is divided into G groups, and global average pooling is performed on the features of each group along the height and width directions to obtain a width-oriented one-dimensional attention map. and high-dimensional attention diagram The calculation formula is as follows: ; ; In the formula, Number of groups The number of channels is represented by the H-axis in the height direction and the W-axis in the width direction. Will and The transformation is performed by spatial dimension concatenation, followed by convolution to learn the coupling relationship between the two directions of information, generating high-axis weights. and width weights After Sigmoid activation, the original features are applied element-wise to the current spatial attention path's input features d, resulting in spatial attention-enhanced features. The calculation formula is as follows: ; In the formula, For high-order weights, For wide-axis weights.

[0019] 2.2.3) Channel attention path for input features Perform global average pooling to obtain the channel description vector, and then transform the vector into a one-dimensional sequence. 1D convolution is used to capture the local correlation between adjacent channels, and the calculation formula is as follows: ; In the formula, These are the convolution kernel weights, the output. , It is 3; Will The channel weights are mapped using the Sigmoid algorithm and then multiplied back into the input features element by element. Channel attention enhancement features were obtained. The calculation formula is as follows: .

[0020] 2.2.4) Enhance spatial attention features and channel attention enhancement features Element-by-element addition and fusion, performing the fusion feature sequentially. convolution, Activation and Convolution, output axial-channel dual attention module features The calculation formula is as follows: .

[0021] Furthermore, the working process of the aforementioned multi-scale enhancement module includes: 2.3.1) Features of Axial-Channel Dual Attention Module As input, channel expansion is performed through convolution to generate intermediate features. The calculation formula is as follows: ; 2.3.2) For intermediate features By concatenating at least two levels of depthwise separable convolutions to form feature responses with different receptive fields, multi-scale features are obtained. The calculation formula is as follows: ; 2.3.3) Convolution of multi-scale features The channel number is compressed back to the input channel number, and after batch normalization, it is compared with the input axial-channel dual attention module features. The multi-scale enhancement module output is obtained by fusing through identity residual connections. The calculation formula is as follows: .

[0022] Furthermore, the above S3) uses a joint loss function consisting of cross-entropy loss and Dice loss to iteratively update the model parameters through backpropagation, including: S3.1 When segmenting medical images into binary categories, the joint loss function is... Binary cross-entropy loss and Dice loss The weighted composition is calculated using the following formula: ; ; ; In the formula, It is the weighting balance coefficient. and Let Dice loss and binary cross-entropy loss represent the losses respectively. This represents the total number of voxels. and The first The true and predicted values ​​of individual elements.

[0023] S3.2 Joint loss function when segmenting medical images into multiple categories Multi-class cross-entropy loss and Dice loss The weighted composition is calculated using the following formula: ; ; In the formula, This represents the cross-entropy loss for multi-class classification. This represents the total number of voxels. For the number of categories, For the first The sample at the th The true label of the class, For the model to the first The sample belongs to the first The predicted probability of a class.

[0024] Compared with existing technologies, the method of the present invention can improve the accuracy and robustness of automatic medical image segmentation, and improve the segmentation performance of fine structures and boundary regions, thereby providing more reliable segmentation results to support clinical diagnosis and treatment decisions. Attached Figure Description

[0025] Figure 1This is a flowchart of the improved TransUNet medical image segmentation method in this invention.

[0026] Figure 2 This is a flowchart illustrating the design of the decoder layer in this invention.

[0027] Figure 3 This is a diagram of the improved TransUNet network structure in this invention.

[0028] Figure 4 This is a structural diagram of the modules in the decoder of this invention. Detailed Implementation

[0029] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be noted that the technical solution and design principle of the present invention will be described in detail below with reference to only one optimized technical solution, but the protection scope of the present invention is not limited thereto.

[0030] The embodiments described above are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essence of the present invention shall fall within the protection scope of the present invention.

[0031] This invention provides a decoding method for medical image segmentation with cross-layer feature refinement, such as... Figure 1 , 3 As shown, it includes the following steps: S1 Dataset Construction: Obtain the original medical image dataset and its corresponding segmentation annotations to form a medical image segmentation dataset. For the dataset Preprocessing was performed, and the dataset was divided into training sets according to a set ratio. and test set .

[0032] In a preferred embodiment of the present invention, S1 includes the following steps: 1.1) For the dataset Preprocessing is performed, including one or more of size normalization, cropping, and data augmentation; in a specific embodiment of the present invention, the dataset... You can choose one or more of the Synapse, Kvasir-Seg, ISIC2018, and ISIC 2017 datasets.

[0033] 1.2) The dataset The training set was divided into two parts at a ratio of 7:3. and test set In a specific embodiment of the present invention, the ratio is 7:3.

[0034] S2 Network Model Construction: A medical image segmentation network model based on TransUnet is constructed. The network model includes an encoder, an improved decoder, and skip connections. The encoder is a hybrid structure of convolutional neural network (CNN) and Transformer, used to extract multi-scale semantic features.

[0035] like Figure 2 , 4 As shown, the improved decoder comprises four decoding layers. The first three decoding layers sequentially include a spatial-channel gating module, an axial-channel dual attention module, and a multi-scale enhancement module. The fourth layer sequentially includes an axial-channel dual attention module and a multi-scale enhancement module. An upsampling module is positioned between adjacent decoding layers. The spatial-channel gating module generates gating weights based on the output features of the deep-level upsampling module, and uses these gating weights to perform semantic filtering on the skip connection features of the corresponding layer. These weights are then fused with the upsampled output features to obtain fused features. The axial-channel dual attention module is used to align and enhance the structural details and boundary information of the fused features. The multi-scale enhancement module is used to extract local texture and edge structure information under different receptive fields, output enhanced features, and use them as input features for the next decoding layer. The upsampling module is used to refine the enhanced features and achieve channel alignment with the skip connection features of the previous layer through channel transformation. The skip connection is used to pass the feature map of the corresponding layer in the encoder to the corresponding layer in the decoder for feature fusion.

[0036] In a preferred embodiment of the present invention, the working process of the upsampling module includes: Input features of the current decoding level Upsampling is performed using bilinear interpolation; the upsampled features are then thinned by convolution, and channel alignment with the skip connection features is achieved through channel transformation to obtain the upsampled aligned features. The calculation formula is as follows:

[0037] In the formula, For a convolution with a kernel of 1, for Activation function For batch normalization, For a convolution with a kernel of 3, This indicates an upsampling operation.

[0038] As a preferred embodiment of the present invention, the operation process of the space-channel gating module includes: 2.1.1) Obtain the skip connection features of the corresponding layer of the decoder. , jump connection features Alignment features with upsampling Separate depthwise separable convolutions are performed, followed by element-wise addition and fusion. The resulting fused features are then activated by ReLU. The calculation formula is as follows: ;

[0039] In the formula, This indicates a depthwise separable convolution with a kernel of 3. 2.1.2) Fusion Features Spatial context relationship modeling is performed, and spatial gating coefficients are obtained by normalization using the Sigmoid function. The calculation formula is as follows: ;

[0040] In the formula, For the Sigmoid function; 2.1.3) Fusion Features Perform global average pooling operation The statistical characteristics of each channel are obtained by compressing spatial information, and the calculation formula is as follows: ; In the formula, For the first One sample, For the first One channel, The height of the feature map, The width of the feature map is used; the global average pooling result is then subjected to dimensionality reduction, dimensionality increase, and Sigmoid normalization to obtain the channel gating coefficients. The calculation formula is as follows: ; 2.1.4) Spatial gating coefficient With channel gating coefficient Perform element-wise multiplication to generate the final joint gating weights. Gating weights Features of skip connections Element-wise multiplication yields the output features of the space-channel gating module. The calculation formula is as follows: ;; ; In the formula, This indicates element-wise multiplication.

[0041] As a preferred embodiment of the present invention, the axial-channel dual attention process includes: 2.2.1) Output characteristics of the space-channel gating module Upsampling alignment features with the corresponding decoding level The input features are then fused. Input features Input the spatial attention path and the channel attention path respectively; 2.2.2) Spatial attention path will input features The channel is divided into G groups, and global average pooling is performed on the features of each group along the height and width directions to obtain a width-oriented one-dimensional attention map. and high-dimensional attention diagram The calculation formula is as follows: ; ; In the formula, Number of groups Let G be the number of channels, H be the height axis, and W be the width axis. In a specific embodiment of the present invention, G is a user-defined variable set to 8.

[0042] Will and The transformation is performed by spatial dimension concatenation, followed by convolution to learn the coupling relationship between the two directions of information, generating high-axis weights. and width weights After Sigmoid activation, the original features are applied element-wise to the current spatial attention path's input features d, resulting in spatial attention-enhanced features. The calculation formula is as follows: ; In the formula, For high-order weights, For wide-axis weights.

[0043] 2.2.3) Channel attention path for input features Perform global average pooling to obtain the channel description vector, and then transform the vector into a one-dimensional sequence. 1D convolution is used to capture the local correlation between adjacent channels, and the calculation formula is as follows:

[0044] In the formula, These are the convolution kernel weights, the output. , It is 3; Will The channel weights are mapped using the Sigmoid algorithm and then multiplied back into the input features element by element. Channel attention enhancement features were obtained. The calculation formula is as follows: .

[0045] 2.2.4) Enhance spatial attention features and channel attention enhancement features Element-by-element addition and fusion, performing the fusion feature sequentially. convolution, Activation and Convolution, output axial-channel dual attention module features The calculation formula is as follows: .

[0046] As a preferred embodiment of the present invention, the multi-scale enhancement module operation process includes: 2.3.1) Features of Axial-Channel Dual Attention Module As input, channel expansion is performed through convolution to generate intermediate features. The calculation formula is as follows:

[0047] 2.3.2) For intermediate features By concatenating at least two levels of depthwise separable convolutions to form feature responses with different receptive fields, multi-scale features are obtained. The calculation formula is as follows: ; As a specific embodiment of the present invention, two-level depth-separable convolution is preferred. and Two scales.

[0048] 2.3.3) Convolution of multi-scale features The channel number is compressed back to the input channel number, and after batch normalization, it is compared with the input axial-channel dual attention module features. The multi-scale enhancement module output is obtained by fusing through identity residual connections. The calculation formula is as follows: .

[0049] S3 Model Training: Using the Training Set A TransUnet-based medical image segmentation network model was trained using a joint loss function consisting of cross-entropy loss and Dice loss, and the model parameters were iteratively updated through backpropagation.

[0050] In a preferred embodiment of the present invention, S3) using a joint loss function composed of the cross-entropy loss function and the Dice loss to iteratively update the model parameters through backpropagation includes: S3.1 When segmenting medical images into binary categories, the joint loss function is... Binary cross-entropy loss and Dice loss The weighted composition is calculated using the following formula: ; ; ; In the formula, It is the weighting balance coefficient. and Let Dice loss and binary cross-entropy loss represent the losses respectively. This represents the total number of voxels. and The first The true and predicted values ​​of individual elements.

[0051] S3.2 Joint loss function when segmenting medical images into multiple categories Multi-class cross-entropy loss and Dice loss The weighted composition is calculated using the following formula: ; ; In the formula, This represents the cross-entropy loss for multi-class classification. This represents the total number of voxels. For the number of categories, For the first The sample at the th The true label of the class, For the model to the first The sample belongs to the first The predicted probability of a class.

[0052] S4 Model Termination and Weight Saving: Repeat step 3) until the preset stopping condition is met, and obtain the trained model weights. The stopping condition is reaching a preset number of training rounds.

[0053] In a specific embodiment of the present invention, the Synapse dataset has 150 rounds, while the Kvasir-Seg, ISIC 2018, and ISIC 2017 datasets have 300 rounds. S5 Medical Image Segmentation: Input the medical image to be segmented into the training weights... The medical image segmentation network model based on TransUnet outputs a medical image segmentation map.

Claims

1. A decoding method for medical image segmentation with cross-layer feature refinement, characterized in that, Includes the following steps: S1 Dataset Construction: Obtain the original medical image dataset and its corresponding segmentation annotations to form a medical image segmentation dataset. For the dataset Preprocessing was performed, and the dataset was divided into training sets according to a set ratio. and test set ; S2 Network Model Construction: A medical image segmentation network model based on TransUnet is constructed. The network model includes an encoder, an improved decoder, and skip connections. The encoder is a hybrid structure of convolutional neural network (CNN) and Transformer, used to extract multi-scale semantic features; The improved decoder comprises four decoding layers. The first three decoding layers are configured with a spatial-channel gating module, an axial-channel dual attention module, and a multi-scale enhancement module, respectively. The fourth layer is configured with an axial-channel dual attention module and a multi-scale enhancement module, respectively. An upsampling module is set between adjacent decoding layers. The spatial-channel gating module generates gating weights based on the output features of the deep-level upsampling modules, and uses the gating weights to perform semantic filtering on the skip connection features of the corresponding layer. Then, it is fused with the upsampling output features to obtain fused features. The axial-channel dual attention is used to align and enhance the structural details and boundary information of the fused features; the multi-scale enhancement module is used to extract local texture and edge structure information under different receptive fields, output enhanced features and use them as input features for the next decoding layer; the upsampling module is used to refine the enhanced features and achieve channel alignment with the skip connection features of the previous layer through channel transformation; the skip connection is used to pass the feature map of the corresponding layer in the encoder to the corresponding layer in the decoder for feature fusion; S3 Model Training: Using the Training Set A TransUnet-based medical image segmentation network model was trained using a joint loss function consisting of cross-entropy loss and Dice loss, and the model parameters were iteratively updated through backpropagation. S4 Model Termination and Weight Saving: Repeat step 3) until the preset stopping condition is met, and obtain the trained model weights. The stopping condition is reaching a preset number of training rounds. S5 Medical Image Segmentation: Input the medical image to be segmented into the training weights... The medical image segmentation network model based on TransUnet outputs a medical image segmentation map.

2. The decoding method for medical image segmentation with cross-layer feature refinement as described in claim 1, characterized in that, Step 1) includes: S1.1 Dataset Preprocessing is performed, which includes one or more of size normalization, cropping, and data augmentation. S.1.2 The dataset The training set was divided into two parts at a ratio of 7:

3. and test set .

3. The decoding method for medical image segmentation with cross-layer feature refinement as described in claim 1, characterized in that, The working process of the upsampling module includes: Input features of the current decoding level Upsampling is performed using bilinear interpolation; the upsampled features are then thinned by convolution, and channel alignment with the skip connection features is achieved through channel transformation to obtain the upsampled aligned features. The calculation formula is as follows: ； In the formula, For a convolution with a kernel of 1, for Activation function For batch normalization, For a convolution with a kernel of 3, This indicates an upsampling operation.

4. The decoding method for medical image segmentation with cross-layer feature refinement as described in claim 1, characterized in that, The operation of the space-channel gating module includes: 2.1.1) Obtain the skip connection features of the corresponding layer of the decoder. , jump connection features Alignment features with upsampling Separate depthwise separable convolutions are performed, followed by element-wise addition and fusion. The resulting fused features are then activated by ReLU. The calculation formula is as follows: ； In the formula, This indicates a depthwise separable convolution with a kernel of 3. 2.1.2) Fusion Features Spatial context relationship modeling is performed, and spatial gating coefficients are obtained by normalization using the Sigmoid function. The calculation formula is as follows: ； In the formula, For the Sigmoid function; 2.1.3) Fusion Features Perform global average pooling operation The statistical characteristics of each channel are obtained by compressing spatial information, and the calculation formula is as follows: ； In the formula, For the first One sample, For the first One channel, The height of the feature map, The width of the feature map is used; the global average pooling result is then subjected to dimensionality reduction, dimensionality increase, and Sigmoid normalization to obtain the channel gating coefficients. The calculation formula is as follows: ； 2.1.4) Spatial gating coefficient With channel gating coefficient Perform element-wise multiplication to generate the final joint gating weights. Gating weights Features of skip connections Element-wise multiplication yields the output features of the space-channel gating module. The calculation formula is as follows: ； ； In the formula, This indicates element-wise multiplication.

5. The decoding method for cross-layer semantic fusion and multi-scale feature enhancement for medical image segmentation as described in claim 1, characterized in that, The axial-channel dual attention process includes: 2.2.1) Output characteristics of the space-channel gating module Upsampling alignment features with the corresponding decoding level The input features are then fused. Input features Input the spatial attention path and the channel attention path respectively; 2.2.2) Spatial attention path will input features The channel is divided into G groups, and global average pooling is performed on the features of each group along the height and width directions to obtain a width-oriented one-dimensional attention map. and high-dimensional attention diagram The calculation formula is as follows: ； ； In the formula, Number of groups The number of channels is represented by the H-axis in the height direction and the W-axis in the width direction. Will and The transformation is performed by spatial dimension concatenation, followed by convolution to learn the coupling relationship between the two directions of information, generating high-axis weights. and width weights After Sigmoid activation, the original features are applied element-wise to the current spatial attention path's input features d, resulting in spatial attention-enhanced features. The calculation formula is as follows: ； In the formula, For high-order weights, For width-oriented weights; 2.2.3) Channel attention path for input features Perform global average pooling to obtain the channel description vector, and then transform the vector into a one-dimensional sequence. 1D convolution is used to capture the local correlation between adjacent channels, and the calculation formula is as follows: ； In the formula, These are the convolution kernel weights, the output. , It is 3; Will The channel weights are mapped using the Sigmoid algorithm and then multiplied back into the input features element by element. Channel attention enhancement features were obtained. The calculation formula is as follows: ； 2.2.4) Enhance spatial attention features and channel attention enhancement features Element-by-element addition and fusion, performing the fusion feature sequentially. convolution, Activation and Convolution, output axial-channel dual attention module features The calculation formula is as follows: 。 6. The decoding method for medical image segmentation with cross-layer feature refinement as described in claim 1, characterized in that, The multi-scale enhancement module operates as follows: 2.3.1) Features of Axial-Channel Dual Attention Module As input, channel expansion is performed through convolution to generate intermediate features. The calculation formula is as follows: ； 2.3.2) For intermediate features By concatenating at least two levels of depthwise separable convolutions to form feature responses with different receptive fields, multi-scale features are obtained. The calculation formula is as follows: ； 2.3.3) Convolution of multi-scale features The number of channels is compressed back to the number of input channels, and after batch normalization, it is compared with the input axial-channel dual attention module features. The multi-scale enhancement module output is obtained by fusing through identity residual connections. The calculation formula is as follows: 。 7. The decoding method for medical image segmentation with cross-layer feature refinement as described in claim 1, characterized in that, S3) uses a joint loss function consisting of cross-entropy loss and Dice loss to iteratively update model parameters through backpropagation, including: S3.1 When segmenting medical images into binary categories, the joint loss function is... Binary cross-entropy loss and Dice loss The weighted composition is calculated using the following formula: ； ； ； In the formula, It is the weighting balance coefficient. and Let Dice loss and binary cross-entropy loss be represented respectively. This represents the total number of voxels. and The first The true and predicted values ​​of individual elements; S3.2 Joint loss function when segmenting medical images into multiple categories Multi-class cross-entropy loss and Dice loss The weighted composition is calculated using the following formula: ； ； In the formula, This represents the cross-entropy loss for multi-class classification. This represents the total number of voxels. For the number of categories, For the first The sample at the th The true label of the class, For the model to the first The sample belongs to the first The predicted probability of a class.

Images