A liver tumor segmentation method based on feature selection and residual fusion

Abstract

The present invention relates to a liver tumor segmentation method based on feature selection and residual fusion. The present invention first optimizes the bottleneck features and jump links of UNet through the attention mechanism, and proposes a bottleneck feature selection module based on the global feature compression operation, namely, a global feature selection module and a neighboring feature selection module. Then, the feature of the spatial attention module is recalibrated using jump links, and finally, the semantic mismatch problem of the front and back spatial features is solved by the spatial feature residual fusion module, so that the features can be expressed efficiently while maintaining low complexity. The present invention makes the result of 2D liver tumor segmentation more accurate, performs component ablation tests on the LiTS public dataset and compares them with the current method, and achieves average Dice scores of 96.2% and 68.4% in the liver and liver tumor segmentation tasks, respectively, which is 0.8% higher than the average Dice score of the current best 2D liver tumor segmentation model.

Description

Technical Field

[0001] The invention relates to a liver tumor segmentation method based on feature selection and residual fusion, and belongs to the technical field of computer vision. Background Art

[0002] Liver cancer is one of the most common cancers with the highest mortality rate in the world. Computed tomography (CT) is a commonly used tumor diagnosis method in clinical practice. It can help doctors formulate corresponding treatment plans by measuring the shape, location and other information of the tumor. Manual segmentation of CT images has problems such as low efficiency and the influence of doctors' subjective experience. Therefore, an efficient automatic segmentation method is urgently needed in clinical practice. Liver treatment can benefit from accurate and fast automatic segmentation methods. However, due to the low contrast of soft tissue in CT images, the high variability of the shape and position of liver tumors, the unclear boundaries of liver tumor regions, and the relatively small size of most tumors, automatic segmentation of liver tumors is a difficult task in practice.

[0003] In order to solve the problem of automatic segmentation of liver tumors, many solutions based on deep learning have been proposed in the past few years. These methods can generally be divided into three categories: (1) 2D models, such as Multiple U-Nets (Chlebus et al., 2018), which use residual links in the contraction path and expansion path of each layer to increase training efficiency and add dropout with a probability of 0.5 before the convolution of the expansion path to prevent network overfitting. The model has achieved leading results in the case of 2D input; FED-Net (Chen et al., 2019) designed a novel feature fusion method based on the attention mechanism, which can embed semantic information into low-level features, and replaced the contraction path and expansion path of the traditional UNet with residual convolution blocks and dense upsampling convolution respectively. Convolution was introduced in the skip link part to improve the semantic mismatch problem before and after, but failed to solve the feature redundancy problem. (2) 2.5D models, such as Res-UNet (Han, 2017), use two U-shaped structures with long and short jump connections. The first network is used for rough liver segmentation, and the second network focuses on detailed liver segmentation. The second network can be trained to further segment the liver and liver tumors. The model receives five adjacent slices to provide contextual information for the model, and uses multi-segment cascade processing to gradually refine the model, resulting in model redundancy and difficulty in deployment. LW-HCN (Zhang et al., 2019) replaces the bottom convolution of the 3D encoder with 2D convolution to reduce the complexity of the model, and designs deep and spatiotemporal convolution. The separation (DSTS) operation processes 3D convolution to reduce the complexity of the model; the nonlinear enhancement and graph cut method (Liao et al., 2019) integrates the image and boundary information into the graph cut energy function by tissue contrast enhancement, and finally uses three-dimensional morphology for post-processing to achieve high-precision segmentation, focusing on the preprocessing and post-processing processes; combined with the imaging omics method (Liu et al., 2020), it first passes through a cascaded 2D segmentation model, and then the results pass through the imaging omics classification model and 3D model respectively to reduce false positives and refine the segmentation results. This method relies on post-processing to eliminate false positives to improve the segmentation results, and the intermediate process is not effective. (3) 3D models, such as H-DenseUNet (Li et al., 2018), which uses a 2D-DenseUNet model to extract intra-slice features and performs hierarchical aggregation based on an automatic context algorithm, i.e., a mixed feature learning system of intra-slice and inter-slice features. The segmentation effect is excellent, but the number of 3D model parameters reaches tens of millions, the model is large and difficult to train; CDNN (Yuan, 2017) proposed a hierarchical deep fully convolutional-deconvolutional neural network model. The first layer uses the entire CT voxel to roughly segment the liver area, the second layer performs fine liver segmentation, and the third layer performs supplementary input for liver tumor segmentation. Summary of the invention

[0004] The present invention relates to a liver tumor segmentation method based on feature selection and residual fusion, which makes the liver tumor segmentation result more accurate.

[0005] The technical solution of the present invention is: a liver tumor segmentation method based on feature selection and residual fusion, the method specifically includes the following:

[0006] Step 1: The bottleneck feature part uses the global feature selection module (FS) to process liver CT images, and the neighbor feature selection module (NFS) to process liver tumor CT images. Both modules use the global feature compression operation (GFS) to compress the feature map;

[0007] Step 2: The skip link part uses the Spatial Attention Module (SAM) and the Spatial Feature Residual Fusion Module (SFRF) to extract effective information in the spatial dimension and solve the semantic gap problem.

[0008] Step 3: The output segmentation map uses an expansion path to increase the feature map dimension and reduce the number of feature channels. The expansion layer uses a 2x upsampling bilinear interpolation method to generate an extended feature map. The extended feature map passes through a convolution unit and is spliced ​​with the high-resolution feature map of the same layer in the contraction path.

[0009] Furthermore, the global feature selection module in Step 1 specifically includes the following:

[0010] The global feature compression (GFS) operation is used. This operation compresses the feature map and integrates the spatial features through matrix operations to improve the network's expressiveness and generalization, as shown in the calculation formula: Where f GFS Represents the global feature compression operation, input X∈R C×HW , output f GFS (X)∈R C×1 , C is the number of channels of the feature map, H and W represent the height and width of the feature map respectively. represents the transpose of the 1*1 convolution matrix, which aims to compress the number of feature map channels from C to 1. The global feature selection module (FS) is shown in the calculation formula f FS (X) = σ(W E (ReLU(W S f GFS (X)))); Y FS (X) = F S (f FS (X))·X. Where f FSis the global feature selection operation, f FS (X)∈R C×1 .W S ∈R M×C Represents a 1*1 convolution matrix, which is used to compress the number of channels. M is the number of channels after compression. E ∈R C×M represents a 1*1 convolution matrix, which is used to restore the number of channels to the original dimension C. σ is the Sigmoid function, which aims to output the activation value of each channel in the range of [0, 1]. S is the spatial dimension expansion function, which is used to expand f FS (X) The spatial dimension is the same as input X, Y FS ∈R C×HW It is the output of the global feature selection module (FS);

[0011] Furthermore, the neighboring feature selection module in Step 1 specifically includes the following:

[0012] The neighbor feature selection module (NFS) uses the above-mentioned GFS operation when compressing the feature map. The difference between NFS and the global feature selection module (FS) is that the NFS module associates n adjacent feature channels, which is conducive to more subtle feature expression and can capture cross-channel interactions. It is suitable for processing liver tumor CT images, while the FS module compresses all channels to fuse global features, which is suitable for data such as liver CT images that have global position characteristics. The n-neighbor association of NFS can be achieved through a one-dimensional convolution operation, as shown in the calculation formula: Among them, the size of the convolution kernel is n, where the size of the kernel n represents the coverage of the interaction of this group of channels, |t| odd represents the odd number closest to t, C represents the number of feature channels, and the values ​​of γ and β are 2 and 1 respectively. The calculation formula of the neighbor feature selection module (NFS) is: NFS (X, n) = σ(C1D n (f GFS (X))); Y NFS (X, n) = F S (f NFS (X, n))·X. Where f NFS Represents the neighboring feature selection operation, with the input being the feature map X∈R C×HW And one-dimensional convolution kernel size n, output f NFS (X,n)∈R C×1 Keep consistent with input. C1D n represents a one-dimensional convolution with a convolution kernel size of n, σ is a sigmoid function, and its output activation value performs feature selection on the channel dimension of the feature map X. The neighbor feature selection module (NFS) outputs Y NFS ∈R C×HW.

[0013] Furthermore, the spatial attention module in Step 2 specifically includes the following:

[0014] The spatial attention model uses 1*1 convolution to compress the channel dimension to 1, without changing the size of the feature map, to ensure the consistency of spatial dimension features, as shown in the calculation formula: SA (X) = σ(W K X); Y SA (X) = F C (f SA (X))·X. Where f SA represents the spatial attention operation, output f SA (X)∈R 1×HW , W K ∈R 1×C is a 1*1 convolution matrix used to compress the channel dimension to 1, σ is the Sigmoid function, and outputs the activation value of the spatial dimension feature. The output Y of the spatial attention module SA ∈R C×HW , F C It is a channel dimension expansion function, which can be used to expand the channel dimension of the function input to C.

[0015] Furthermore, the spatial feature residual fusion module SFRF in Step 2 specifically includes:

[0016] The spatial feature residual fusion module (SFRF) is used to solve the semantic gap problem caused by skip links. The spatial feature residual fusion module (SFRF) receives the relatively shallow features (k layers) and the relatively deep features (k+1 layers) of the model as input, where k is an integer in [1, 4]. As shown in the calculation formula: Y SFRF (X k , X k+1 )=X k +F up (W F X k+1 ). Where X k ∈R C×HW Output features from the k-layer contraction path, Output features from the k+1 layer SFRF, only when k=4, X k+1 is the output of the bottleneck feature selection module, Indicates that the size of H and W are halved. F ∈R C×2C It is a 1*1 convolution operation, the purpose is to compress the number of channels from 2C to C, F up It is a 2x upsampling bilinear interpolation operation. The feature map can be expanded by bilinear interpolation to make the output Y SFRF (X k, X k+1 )∈R C×HW With X k Maintain consistency across feature dimensions.

[0017] Furthermore, the Step 3 also includes the following:

[0018] In view of the imbalance of positive and negative samples in the data set, a weighted combination of binary cross entropy (BCE) and dice loss (DL) is adopted. Since dice loss (DL) may cause drastic changes in gradients and thus affect back propagation and make training difficult, we appropriately reduce the weight of DL. The loss function is shown in the calculation formula: Where y represents the true segmentation map, is the segmentation map predicted by the model, ω is the weight of the two losses, which is set to 0.3 in the present invention, and ∈ is a smoothing term set to avoid the denominator being 0, which is set to 1.0 in the present invention.

[0019] The beneficial effects of the present invention are as follows: the present invention makes the result of 2D liver tumor segmentation more accurate, performs component ablation tests on the LiTS public dataset and compares the results with the current method, and achieves average Dice scores of 96.2% and 68.4% in the liver and liver tumor segmentation tasks, respectively. This result is comparable to the effect of some 2.5D and 3D models, and is 0.8% higher than the average Dice score of the current best 2D liver tumor segmentation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is the traditional UNet structure diagram;

[0021] Figure 2 Schematic diagram of the overall model and key component structure; (A.Overall Architecture; B.FeatureSelection Module; C.Global Feature Squeeze Module; D.Neighbor Feature SelectionModule; E.Spatial Attention Module; F.Spatial Feature Residual Fusion Module);

[0022] Figure 3 The figure is a comparison of the segmentation results of the present invention. DETAILED DESCRIPTION

[0023] Example 1, a liver tumor segmentation method based on feature selection and residual fusion, the specific steps of the method are as follows:

[0024] Step 1: The bottleneck feature part uses the global feature selection module (FS) to process liver CT images, and the neighbor feature selection module (NFS) to process liver tumor CT images. Both modules use the global feature compression operation (GFS) to compress the feature map;

[0025] Step 2: The skip link part uses the Spatial Attention Module (SAM) and the Spatial Feature Residual Fusion Module (SFRF) to extract effective information in the spatial dimension and solve the semantic gap problem.

[0026] Step 3: The output segmentation map uses an expansion path to increase the feature map dimension and reduce the number of feature channels. The expansion layer uses a 2x upsampling bilinear interpolation method to generate an extended feature map. The extended feature map passes through a convolution unit and is spliced ​​with the high-resolution feature map of the same layer in the contraction path.

[0027] Furthermore, the global feature selection module in Step 1 specifically includes the following:

[0028] The global feature compression (GFS) operation is used. This operation compresses the feature map and integrates the spatial features through matrix operations to improve the network's expressiveness and generalization, as shown in the calculation formula: Where f GFS Represents the global feature compression operation, input X∈R C×HW , output f GFS (X)∈R C×1 , C is the number of channels of the feature map, H and W represent the height and width of the feature map respectively. represents the transpose of the 1*1 convolution matrix, which aims to compress the number of feature map channels from C to 1. The global feature selection module (FS) is shown in the calculation formula f FS (X) = σ(W E (ReLU(W S f GFS (X)))); Y FS (X) = F S (f FS (X))·X. Where f FS is the global feature selection operation, f FS (X)∈R C×1 .W S ∈R M×CRepresents a 1*1 convolution matrix, which is used to compress the number of channels. M is the number of channels after compression. E ∈R C×M represents a 1*1 convolution matrix, which is used to restore the number of channels to the original dimension C. σ is the Sigmoid function, which aims to output the activation value of each channel in the range of [0, 1]. S is the spatial dimension expansion function, which is used to expand f FS (X) The spatial dimension is the same as input X, Y FS ∈R C×HW It is the output of the global feature selection module (FS);

[0029] Furthermore, the neighboring feature selection module in Step 1 specifically includes the following:

[0030] The neighbor feature selection module (NFS) uses the above-mentioned GFS operation when compressing the feature map. The difference between NFS and the global feature selection module (FS) is that the NFS module associates n adjacent feature channels, which is conducive to more subtle feature expression and can capture cross-channel interactions. It is suitable for processing liver tumor CT images, while the FS module compresses all channels to fuse global features, which is suitable for data such as liver CT images that have global position characteristics. The n-neighbor association of NFS can be achieved through a one-dimensional convolution operation, as shown in the calculation formula: Among them, the size of the convolution kernel is n, where the size of the kernel n represents the coverage of the interaction of this group of channels, |t| odd represents the odd number closest to t, C represents the number of feature channels, and the values ​​of γ and β are 2 and 1 respectively. The calculation formula of the neighbor feature selection module (NFS) is: NFS (X, n) = σ(C1D n (f GFS (X))); Y NFS (X, n) = F S (f NFS (X, n))·X. Where f NFS Represents the neighboring feature selection operation, with the input being the feature map X∈R C×HW And one-dimensional convolution kernel size n, output f NFS (X,n)∈R C×1 Keep consistent with input. C1D n represents a one-dimensional convolution with a convolution kernel size of n, σ is a sigmoid function, and its output activation value performs feature selection on the channel dimension of the feature map X. The neighbor feature selection module (NFS) outputs Y NFS ∈R C×HW .

[0031] Furthermore, the spatial attention module in Step 2 specifically includes the following:

[0032] The spatial attention model uses 1*1 convolution to compress the channel dimension to 1, without changing the size of the feature map, to ensure the consistency of spatial dimension features, as shown in the calculation formula: SA (X) = σ(W K X); Y SA (X) = F C (f SA (X))·X. Where f SA represents the spatial attention operation, output f SA (X)∈R 1×HW , W K ∈R 1×C is a 1*1 convolution matrix used to compress the channel dimension to 1, σ is the Sigmoid function, and outputs the activation value of the spatial dimension feature. The output Y of the spatial attention module SA ∈R C×HW , F C It is a channel dimension expansion function, which can be used to expand the channel dimension of the function input to C.

[0033] Furthermore, the spatial feature residual fusion module SFRF in Step 2 specifically includes:

[0034] The spatial feature residual fusion module (SFRF) is used to solve the semantic gap problem caused by skip links. The spatial feature residual fusion module (SFRF) receives the relatively shallow features (k layers) and the relatively deep features (k+1 layers) of the model as input, where k is an integer in [1, 4]. As shown in the calculation formula: Y SFRF (X k , X k+1 )=X k +F up (W F X k+1 ). Where X k ∈R C×HW Output features from the k-layer contraction path, Output features from the k+1 layer SFRF, only when k=4, X k+1 is the output of the bottleneck feature selection module, Indicates that the size of H and W are halved. F ∈R C×2C It is a 1*1 convolution operation, the purpose is to compress the number of channels from 2C to C, F up It is a 2x upsampling bilinear interpolation operation. The feature map can be expanded by bilinear interpolation to make the output Y SFRF (X k , X k+1 )∈R C×HW With X k Maintain consistency across feature dimensions.

[0035] Furthermore, the Step 3 also includes the following:

[0036] In view of the imbalance of positive and negative samples in the data set, a weighted combination of binary cross entropy (BCE) and dice loss (DL) is adopted. Since dice loss (DL) may cause drastic changes in gradients and thus affect back propagation and make training difficult, we appropriately reduce the weight of DL. The loss function is shown in the calculation formula: Where y represents the true segmentation map, is the segmentation map predicted by the model, ω is the weight of the two losses, which is set to 0.3 in the present invention, and ∈ is a smoothing term set to avoid the denominator being 0, which is set to 1.0 in the present invention.

[0037] Furthermore, in order to verify the effect of the present invention, the model trained by the above steps was input into the training data set of MICCAI 2017LiTS Challenge to test the model, which contains 131 contrast-enhanced abdominal CT scans from 7 clinical institutions. The environment configuration of the test is CPU: Intel(R) Silver; GPU: NVIDIA TITAN XP (12G); memory: 128G; operating system: Ubuntu16.04.

[0038] This model uses UNet as the base network (Baseline), and the model components include global feature selection module (FS) or neighbor feature selection module (NFS); spatial attention module (SAM); spatial feature residual fusion module (SFRF). The prediction results are measured using the internationally commonly used average Dice score, which reflects the similarity between the prediction results and the true results. The calculation formula of the average Dice score is shown below. Where y ij and They represent the true segmentation map and the model predicted segmentation map of the j-th slice of the i-th CT respectively, and N represents the total number of all CT slices.

[0039] The test results of the model components are shown in Table 1. After adding the bottleneck feature selection (FS / NFS) operation, the model is significantly improved compared with the basic network. The average Dice score of the liver segmentation prediction results is above 0.95, which is about 37% better than the wrong prediction of the basic network, and the liver tumor segmentation prediction score is above 0.65. The basic network has improved the effect after adding the spatial attention module (SAM) and the spatial feature residual fusion module (SFRF) at the jump link. The effect of adding the SFRF module alone is limited. The reason for the analysis is that the SFRF module receives the original UNet features without the k-layer spatial attention mechanism (SAM) and the k+1-layer bottleneck feature selection (FS / NFS), resulting in redundant features repeatedly interfering with the model and making training difficult. After adding all components, the FS module and the NFS module achieve the highest average Dice scores in the liver segmentation and liver tumor segmentation tasks, respectively. The experimental results demonstrate the effectiveness of the model components. As shown in Table 2, an average Dice score of 68.4% was achieved in the liver and liver tumor segmentation task, which is 0.8% higher than the average Dice score of 67.6% of the current best 2D liver tumor segmentation model.

[0040] Table 1 Ablation test results of key components of the model

[0041] Model Liver Tumor Baseline 0.921 0.633 Baseline+FS 0.952 0.650 Baseline+NFS(5) 0.951 0.655 Baseline+SAM 0.945 0.646 Baseline+SFRF 0.931 0.639 Baseline+SAM+SFRF 0.952 0.653 Baseline+FS+SAM 0.957 0.661 Baseline+FS+SFRF 0.959 0.674 Baseline+NFS(5)+SAM 0.955 0.667 Baseline+NFS(5)+SFRF 0.956 0.676 Baseline+FS+SAM+SFRF 0.962 0.682 Baseline+NFS(5)+SAM+SFRF 0.959 0.684

[0042] Table 2 Comparison of the present invention with other mainstream segmentation methods

[0043] Methods Liver Tumor Attention Unet (Oktay et al., 2018) 0.956 0.676 MS-DG (Ashish et al., 2021) 0.960 0.657 MsAUNet (Soham et al., 2020) 0.962 0.673 The present invention 0.962 0.684

[0044] like Figure 3 The figure shows the comparison of segmentation results, where column a is three original CT images, column b is the real labels, column c is the result labels of the benchmark model UNet, column d is the improved results after the FS module is added to the present invention, and column e is the improved results after the NFS module is added to the present invention. It can be seen from the figure that the images in columns d and e are very close to the original labels, which is better than the benchmark model in column c.

[0045] The specific implementation modes of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A liver tumor segmentation method based on feature selection and residual fusion, characterized by: The method specifically includes the following: Step 1: The bottleneck feature part uses a global feature selection module to process liver CT images, and the neighboring feature selection module to process liver tumor CT images. Both modules use global feature compression operations to compress feature maps; Step 2: The skip link part uses the spatial attention module and the spatial feature residual fusion module to extract effective information in the spatial dimension and solve the semantic gap problem; Step 3: The output segmentation map uses an expansion path to increase the feature map dimension and reduce the number of feature channels. The expansion layer uses a 2x upsampling bilinear interpolation method to generate an extended feature map. The extended feature map passes through a convolution unit and is concatenated with the high-resolution feature map of the same layer of the contraction path. The global feature selection module in Step 1 specifically includes the following: Use the global feature compression GFS operation, which compresses the feature map and integrates the spatial features through matrix operations to improve the network's expressiveness and generalization, as shown in the calculation formula: Where f GFS Represents the global feature compression operation, input X∈R C×HW , output f GFS (X)∈R C×1 , C is the number of channels of the feature map, H and W represent the height and width of the feature map respectively; represents the transpose of the 1*1 convolution matrix, which aims to compress the number of feature map channels from C to 1. The global feature selection module FS is shown in the calculation formula f FS (X) = σ(W E (ReLU(W S f GFS (X)))); Y FS (X) = F S (f FS (X))·X; where f FS is the global feature selection operation, f FS (X)∈R C×1 , W S ∈R M×c Represents a 1*1 convolution matrix, which is used to compress the number of channels. M is the number of channels after compression, and W E ∈R C×M represents a 1*1 convolution matrix, which is used to restore the number of channels to the original dimension C. σ is the Sigmoid function, which aims to output the activation value of each channel in the range of [0,1]. S is the spatial dimension expansion function, which is used to expand f FS (X) The spatial dimension is the same as input X, Y FS ∈R C×HW It is the output of the global feature selection module FS; The neighboring feature selection module in Step 1 specifically includes the following: The neighbor feature selection module NFS uses GFS operation when compressing feature maps. The difference between NFS and the global feature selection module FS is that the NFS module associates n adjacent feature channels, which is conducive to more subtle feature expression and can capture cross-channel interactions. It is suitable for processing liver tumor CT images. The global feature selection module FS compresses all channels to fuse global features, which is suitable for data such as liver CT images that have global position characteristics. The n-neighbor association of the neighbor feature selection module NFS can be achieved through a one-dimensional convolution operation, as shown in the calculation formula: Among them, the size of the convolution kernel is n, where the size of the kernel n represents the coverage of the interaction of this group of channels, |t| odd represents the odd number closest to t, C represents the number of feature channels, and the values ​​of γ and β are 2 and 1 respectively; the calculation formula of the neighboring feature selection module NFS is: NFS (X,n)=σ(C1D n (f GFS (X))); Y NFS (X,n)=F S (f NFS (X,n))·X; where f NFS Represents the neighboring feature selection operation, with the input being the feature map X∈R C×HW And one-dimensional convolution kernel size n, output f NFS (X,n)∈R C×1 Keep consistent with input; C1D n represents a one-dimensional convolution with a convolution kernel size of n, σ is a sigmoid function, and its output activation value performs feature selection on the channel dimension of the feature map X; the neighboring feature selection module NFS outputs Y NFS ∈R C×HW ; The spatial attention module in Step 2 specifically includes the following: The spatial attention model uses 1*1 convolution to compress the channel dimension to 1, without changing the size of the feature map, to ensure the consistency of spatial dimension features, as shown in the calculation formula: SA (X) = σ(W K X); Y SA (X) = F C (f SA (X))·X; where f Sa represents the spatial attention operation, output f sA (X)∈R 1×HW , W K ∈R 1×C is a 1*1 convolution matrix used to compress the channel dimension to 1, σ is the Sigmoid function, and outputs the activation value of the spatial dimension feature; the output Y of the spatial attention module SA ∈R C×HW , F C is a channel dimension expansion function, through which the channel dimension of the function input can be expanded to C; The spatial feature residual fusion module SFRF in Step 2 specifically includes: The spatial feature residual fusion module SFRF is used to solve the semantic gap problem caused by jump links; the spatial feature residual fusion module SFRF receives the relatively shallow features (k layers) and the relatively deep features (k+1 layers) of the model as input, where k is an integer in [1,4]; as shown in the calculation formula: Y SFRF (X k ,X k+1 )=X k +F up (W F X k+1 );where X k ∈R C×HW Output features from the k-layer contraction path, Output features from the k+1 layer SFRF, only when k=4, X k+1 is the output of the bottleneck feature selection module, Indicates that the sizes of H and W are halved respectively; W F ∈R C×2C It is a 1*1 convolution operation, the purpose is to compress the number of channels from 2C to C, F up It is a 2x upsampling bilinear interpolation operation. The feature map can be expanded by bilinear interpolation to make the output Y SFRF (X k ,X k+1 )∈R C×HW With X k Be consistent in feature dimensions; The Step 3 also includes the following: In view of the imbalance of positive and negative samples in the data set, a weighted combination of binary cross entropy and dice loss is adopted. Since dice loss may cause drastic changes in gradients, thus affecting back propagation and making training difficult, the weight of DL is reduced. The loss function is shown in the calculation formula: Where y represents the true segmentation map, is the segmentation map predicted by the model, ω is the weight of the two losses, set to 0.3, ∈ is the smoothing term set to avoid the denominator being 0, and ∈ is set to 1.0.

Images