Colon polyp segmentation system based on multi-scale feature fusion

Abstract

The invention discloses a colon polyp segmentation system based on multi-scale feature fusion, which comprises an original image acquisition module, a preprocessing module, an SwinFF-Net network module and a result output module, and is characterized in that the SwinFF-Net network module comprises an encoder, a jump connection module and a decoder; the system enhances the capture of detail information in the image by fusing the features of different scales, especially for the conditions that the boundary is fuzzy and the diseased region is difficult to distinguish. The multi-scale feature fusion can help the model to extract effective features from different levels of image information, so that the distinction degree between the colon polyp and the background is improved, and more accurate and more complete polyp segmentation is realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and particularly relates to a colon polyp segmentation system based on multi-scale feature fusion. Background Art

[0002] Colorectal cancer is a highly prevalent gastrointestinal malignant tumor with a relatively high prevalence rate. Therefore, accurate diagnosis of polyps is of great significance for the prevention and treatment of colorectal cancer. Therefore, accurate and efficient medical image segmentation based on deep learning can not only identify the characteristics of the lesion area, but also provide a reliable basis for clinical diagnosis and disease research.

[0003] In recent years, with the development of deep learning, neural networks based on deep learning have been gradually widely applied in the field of medical images. U-Net is used to address the challenges of medical image segmentation. This network adopts an encoder-decoder structure and realizes effective feature learning through skip connections, and performs excellently in many medical image segmentation tasks, quickly becoming the mainstream baseline algorithm (SOTA) in the field of medical image segmentation. With the wide application of U-Net, many improved models based on the U-shaped architecture have been successively proposed and achieved excellent results in practical tasks. For example, UNet++, Attention U-Net, DenseUNet. Although convolutional neural networks perform excellently in medical image segmentation tasks, however, most networks ignore the multi-scale feature fusion of images, resulting in problems such as mis-segmentation and missed-segmentation in the segmentation of polyp boundaries.

[0004] In TransFuse, convolution and Transformer are used as a dual-branch structure to extract feature information with different receptive fields in parallel. Through this parallel processing method, the convolution branch can efficiently capture local edge and detail features, while the Transformer branch is good at modeling global context information. Subsequently, a specific fusion module is used to deeply fuse the features of the dual branches, thereby achieving the full integration of multi-scale features and significantly improving the segmentation accuracy of colon polyps. DeepLabv3+ performs outstandingly in multi-scale feature fusion. By using dilated convolutions with multiple different dilation rates, features of the image can be extracted at different scales. This enables the model to better understand different scales of objects, thereby improving the segmentation effect. Although the above methods have achieved certain effects in colon polyp image segmentation, the similarity between polyps and the background in colonoscopy images results in low contrast, which in turn causes problems such as inaccurate segmentation, blurred segmentation boundaries, and incomplete segmentation. Summary of the Invention

[0005] In the context of colon polyp image segmentation, due to the similarity between polyps and the background in colonoscopy images, the image contrast is low, which makes the segmentation process still face problems such as low accuracy, blurred segmentation boundaries, and incomplete segmentation. To address these challenges, the present invention proposes a colon polyp segmentation method based on multi-scale feature fusion. This method enhances the capture of detailed information in the image by fusing features at different scales, especially for cases where the boundaries are blurred and the diseased areas are difficult to distinguish. Multi-scale feature fusion can help the model extract effective features from image information at different levels, thereby improving the distinguishability between colon polyps and the background, and thus achieving more accurate and complete polyp segmentation.

[0006] The technical solution of the present invention is as follows:

[0007] A colon polyp segmentation system based on multi-scale feature fusion, comprising: an original image acquisition module, a preprocessing module, a SwinFF-Net network module, and a result output module. The SwinFF-Net network module includes an encoder, a skip connection module, and a decoder;

[0008] The encoder includes four feature extraction blocks. The first feature extraction block consists of two Swin Transformer Blocks; the second feature extraction block consists of 8 Swin Transformer Blocks, the third feature extraction block consists of 4 Swin Transformer Blocks, and the fourth feature extraction block consists of 2 Swin Transformer Blocks;

[0009] The skip connection module includes: four multi-scale feature fusion modules, a feature enhancement aggregation module, and four local feature optimization modules; the four multi-scale feature fusion modules: respectively receive the features from the four feature extraction blocks,

[0010] The decoder uses convolution operations and bilinear interpolation as the backbone network in the decoding stage, including four decoder units, which respectively receive the features from the four local feature optimization modules. The first decoder unit receives the output from the first LFR, performs upsampling, and then sends it into the first convolutional block, then into the second upsampling part, and then concatenates it with the output of the second LFR, and then sends it into the second convolutional block, then into the third upsampling part, and then concatenates it with the output of the third LFR, and then sends it into the third convolutional block, then into the fourth upsampling part, and then concatenates it with the output of the fourth LFR, and then sends it into the fourth convolutional block, and finally obtains the final segmentation result through a 1×1 convolution and a sigmoid function;

[0011] The result output module gives the segmentation result of the diseased area of the colon polyp according to the network training result.

[0012] The described system, a multi-scale feature fusion module, extracts multi-scale features from different receptive fields to enhance the network's perception of target boundaries and size changes.

[0013] The described system, a feature enhancement aggregation module, performs deep interaction and saliency enhancement on the multi-scale features, thereby further improving the expression ability of edge features and the segmentation effect of foreground regions.

[0014] The described system, a local feature optimization module, precisely optimizes and enhances the features of local regions, improving the network's ability to locate polyp regions, especially polyp targets with blurred boundaries and large size variations; the local feature optimization module fully excavates and focuses on the significant features of the target region through the combined operation of spatial attention and channel attention mechanisms; the specific calculation formula is as follows:

[0015] Output=σ(MLP(GAP(F))+MLP(GMP(F)))

[0016] where GAP(F) and GMP(F) respectively represent global average pooling and max pooling operations performed in the spatial dimension. Through the optimization of the LFR module, the model performs excellently in processing segmentation tasks with blurred boundaries and different target sizes, and is especially suitable for polyp segmentation in colonoscopy images with low foreground-background contrast.

[0017] The described system, the Swin Transformer Block sequentially includes a first layer normalization (LN) unit, a window multi-head self-attention (W-MSA) unit, a second layer normalization (LN) unit, a first multi-layer perceptron (MLP) unit, a third layer normalization (LN) unit, a shifted window multi-head self-attention (SW-MSA) unit, a fourth layer normalization (LN) unit, and a second multi-layer perceptron (MLP) unit.

[0018] For the described system, first, the input feature map is sent to the first layer normalization (LN) unit for layer normalization (LN) processing, and then enters the window multi-head self-attention (W-MSA) for window multi-head self-attention (W-MSA) calculation; then, the features are processed through the second layer normalization (LN) unit and the first multi-layer perceptron (MLP) unit, and then normalized through the fourth layer normalization (LN) unit, and then enter the shifted window multi-head self-attention (SW-MSA) unit for shifted window multi-head self-attention (SW-MSA) calculation, and then output after being processed by the fourth layer normalization (LN) unit and the second multi-layer perceptron (MLP) unit.

[0019] In the described system, the Swin Transformer Block adopts layer normalization (LN) as a preprocessing strategy to ensure the stability of features; in the input stage, the features are first subjected to layer normalization and then enter the window multi-head self-attention (W-MSA) module; in the W-MSA, the input features are linearly transformed to generate queries (Q), keys (K), and values (V), and the attention weights are calculated through scaled dot-product attention:

[0020]

[0021] where d is the scaling factor, and the output of the W-MSA is added to the original input through a residual connection to obtain Y:

[0022] Y = X + W-MSA(LN(X))

[0023] Subsequently, Y undergoes layer normalization again and is input into the multi-layer perceptron (MLP). The MLP consists of two fully connected layers and the GELU activation function, and its calculation formula is:

[0024] MLP(Y) = W2(σ(W1Y + b1)) + b2

[0025] Finally, the output of the MLP is fused with Y through a residual connection to obtain the output of the current Swin Transformer:

[0026] Z = Y + MLP(LN(Y))

[0027] After the standard Swin Transformer, a sliding window multi-head self-attention (SW-MSA) mechanism is introduced to enhance the information interaction between different windows; first, Z is layer-normalized and then input into the SW-MSA. This module uses a calculation method similar to that of the W-MSA, but the window position is translated before calculating the attention to break the limitation of the fixed window; its calculation method is as follows:

[0028] Y′ = Z + SW-MSA(LN(Z))

[0029] Subsequently, Y′ enters the MLP for feature transformation:

[0030] Z′ = Y′ + MLP(LN(Y′))

[0031] Finally, Z′ is used as the output of the Swin Transformer, which fuses local and global feature information.

[0032] In the described system, dilated convolutions with dilation rates of 1, 2, 3, and 4 are respectively performed in the multi-scale feature fusion module. The receptive field of the convolution kernel is enlarged through sparse sampling, enabling the network to capture local features and global semantic information. The parallel design of the dilated convolutions ensures feature diversity, capturing both detailed boundary information and global context semantics. The concatenation operation concatenates the outputs of these four dilation rates along the channel dimension, thus combining multi-scale features together to form a high-dimensional feature representation containing rich context information. Finally, the concatenated features are compressed in the channel dimension through a 1×1 convolution.

[0033] Output = Conv 1×1 (Concat(r d＝1 ,r d＝2 ,r d＝3 ,r d＝4 ))

[0034] Among them, Concat() represents concatenation along the channel dimension, and Conv 1×1 () represents the 1×1 convolution operation. r d＝1 ,r d＝2 ,r d＝3 ,r d＝4 are the output features of the dilated convolutions with dilation rates of 1, 2, 3, and 4 respectively.

[0035] In the described system, in the result output module, three commonly used evaluation metrics are adopted: Dice coefficient, mean intersection over union (mIoU), and accuracy (ACC). The calculation formulas are respectively:

[0036]

[0037] Among them, TP, TN, FP, and FN respectively represent true positive, true negative, false positive, and false negative.

[0038] This method enhances the capture of detailed information in the image by fusing features at different scales, especially for cases where the boundaries are blurred and the diseased areas are difficult to distinguish. Multi-scale feature fusion can help the model extract effective features from image information at different levels, thereby improving the distinguishability between colon polyps and the background, and thus achieving more accurate and complete polyp segmentation. Brief Description of the Drawings

[0039] Figure 1 It is a structural diagram of the network (SwinFF-Net) of the colon polyp segmentation method based on multi-scale feature fusion of the present invention.

[0040] Figure 2This is the structural diagram of the Multi Scale Feature Fusion (MSFF) module proposed by the present invention.

[0041] Figure 3 This is the structural diagram of the Feature Enhancement and Aggregation Module (FEAM) proposed by the present invention.

[0042] Figure 4 This is the structural diagram of the Local Feature Refinement (LFR) proposed by the present invention.

[0043] Figure 5 This is the internal structural diagram of the Swin Transformer. Detailed implementation manners

[0044] The present invention will be described in detail below in conjunction with specific embodiments.

[0045] Refer to Figures 1-5 , the colon polyp segmentation system based on multi-scale feature fusion includes:

[0046] S1: Original image acquisition module;

[0047] S2: Preprocessing module, which preprocesses the original image. For polyp images with different sizes, it first adjusts them to a unified size through scaling and padding operations.

[0048] S3: SwinFF-Net network module, constructs the designed SwinFF-Net and conducts network training. As Figure 1 shown, the SwinFF-Net network module includes an encoder, a skip connection module, and a decoder;

[0049] The encoder includes four feature extraction blocks. Specifically, the first feature extraction block consists of two Swin Transformer Blocks, as Figure 1 and 5As shown. The Swin Transformer Block sequentially includes a first layer normalization (LN) unit, a window self-attention (W-MSA) unit, a second layer normalization (LN) unit, a first multi-layer perceptron (MLP) unit, a third layer normalization (LN) unit, a self-attention of sliding window (SW-MSA) unit, a fourth layer normalization (LN) unit, and a second multi-layer perceptron (MLP) unit; First, the input feature map is sent to the first layer normalization (LN) unit for layer normalization (LN) processing, and then enters the window self-attention (W-MSA) for window self-attention (W-MSA) calculation. Then, the features are processed through the second layer normalization (LN) unit and the first multi-layer perceptron (MLP) unit, and then normalized through the fourth layer normalization (LN) unit, and then enter the self-attention of sliding window (SW-MSA) unit for self-attention of sliding window (SW-MSA) calculation, and then output after being processed by the fourth layer normalization (LN) unit and the second multi-layer perceptron (MLP) unit. Next, downsampling is performed through the Patch Merging layer and sent into the second feature extraction block, and the second feature extraction block consists of 8 Swin Transformer Blocks. Then, the feature map is downsampled again through Patch Merging and sent into the third feature extraction block, and the third feature extraction block consists of 4 Swin Transformer Blocks. Subsequently, downsampling operation is performed again and sent into the fourth feature extraction block, and the fourth feature extraction block consists of 2 Swin Transformer Blocks.

[0050] The present invention uses Swin Transformer as the backbone network of the encoder. Considering that there are many similarities between polyps and the background in colonoscopy images, and in addition, the number of polyps varies, it is necessary to build a global feature information model. However, the traditional Vision Transformer has computational redundancy and is difficult to efficiently capture local features. While Swin Transformer can efficiently capture local and global information through the sliding window mechanism and hierarchical feature extraction strategy, significantly reducing the computational complexity and being more suitable for the colonoscopy image segmentation task.

[0051] Swin Transformer adopts a self-attention mechanism with local windows, which can effectively reduce the computational complexity and achieve global information perception through cross-window feature interaction. The feature maps at each stage gradually extract higher-dimensional semantic information through hierarchical downsampling operations. This multi-scale feature representation is crucial for identifying polyp regions with blurred boundaries and different sizes. In addition, in the feature representation of each stage of Swin Transformer, the interaction within and between windows is achieved through the Shifted Window strategy, overcoming the limitations of traditional sliding window methods and enhancing the feature expression ability.

[0052] Each Swin Transformer consists of the following parts:

[0053] Swin Transformer uses layer normalization (LN) as a preprocessing strategy to ensure the stability of features. At the input stage, the features are first layer-normalized and then enter the window multi-head self-attention (W-MSA) module. In W-MSA, the input features are linearly transformed to generate queries (Q), keys (K), and values (V), and the attention weights are calculated through scaled dot-product attention:

[0054]

[0055] where d is the scaling factor. The output of W-MSA is added to the original input through a residual connection to obtain Y:

[0056] Y = X + W-MSA(LN(X))

[0057] Subsequently, Y is layer-normalized again and input into a multi-layer perceptron (MLP), which consists of two fully-connected layers and a GELU activation function. Its calculation formula is:

[0058] MLP(Y) = W2(σ(W1Y + b1)) + b2

[0059] Finally, the output of MLP is fused with Y through a residual connection to obtain the output of the current Swin Transformer:

[0060] Z = Y + MLP(LN(Y))

[0061] After the standard Swin Transformer, a shifted window multi-head self-attention (SW-MSA) mechanism is introduced to enhance the information interaction between different windows. First, Z is layer-normalized and then input into SW-MSA. This module uses a similar calculation method to W-MSA, but the window positions are shifted before calculating the attention to break the limitation of fixed windows. Its calculation method is as follows:

[0062] Y' = Z + SW - MSA(LN(Z))

[0063] Subsequently, Y' enters the MLP for feature transformation:

[0064] Z' = Y' + MLP(LN(Y'))

[0065] Finally, Z' serves as the output of the Swin Transformer, fusing local and global feature information. This design combines local self-attention, a sliding window strategy, and residual connections, making the Swin Transformer more efficient in processing objects with blurred boundaries while enhancing the feature modeling ability and the global information integration ability.

[0066] The described skip connection module includes: a Multi Scale Feature Fusion (MSFF) module, a Feature Enhancement and Aggregation Module (FEAM), and a Local Feature Refinement (LFR).

[0067] The Multi Scale Feature Fusion (MSFF) module aims to extract multi-scale features from different receptive fields to enhance the network's perception of object boundaries and size changes. Its design inspiration comes from the multi-scale changes in foreground and background regions in images. Especially in tasks such as polyp detection, the size and shape of the object vary greatly, and it is difficult for traditional single convolutional kernels to simultaneously focus on the features of large and small objects. As Figure 2 shown, the features from the encoder are input into dilated convolutions with dilation rates of 1, 2, 3, and 4 respectively. The purpose of dilated convolutions is to expand the receptive field of the convolutional kernel through sparse sampling, enabling the network to capture local features and global semantic information. The parallel design of dilated convolutions ensures feature diversity, capturing both detailed boundary information and global context semantics. The concatenation operation concatenates the outputs of these four dilation rates in the channel dimension, thus combining multi-scale features together to form a high-dimensional feature representation containing rich context information. Finally, the concatenated features usually go through a 1×1 convolution for channel compression, aiming to reduce the number of channels, reduce computational overhead, and at the same time fuse the feature information from different receptive fields.

[0068] Output = Conv 1×1 (Concat(r d＝1 ,r d＝2 ,r d＝3 ,r d＝4))

[0069] Among them, Concat() represents concatenation in the channel dimension, and Conv 1×1 () represents a 1×1 convolution operation, r d＝1 ,r d＝2 ,r d＝3 ,r d＝4 are the output features of dilated convolutions with dilation rates of 1, 2, 3, and 4 respectively. This design of multi-scale feature fusion is particularly crucial in dealing with image segmentation tasks with targets of different sizes. Especially in scenarios with blurred boundaries and similar foreground and background, it can effectively perceive the edge information and overall shape of the target, improving the robustness and accuracy of segmentation.

[0070] The Feature Enhancement and Aggregation Module (FEAM) aims to perform deep interaction and saliency enhancement on multi-scale features, thereby further improving the expression ability of edge features and the segmentation effect of foreground regions. The core idea of this module is to screen, enhance, and aggregate multi-scale features to ensure that key information can be fully utilized while suppressing redundant and irrelevant features. As Figure 3 shown, the input of the FEAM module comes from the four output feature maps of the multi-scale feature fusion module (MSFF). Through this design of feature enhancement and aggregation, the FEAM module has stronger selectivity and robustness in the feature expression of the foreground region, making the network perform excellently in dealing with segmentation tasks with blurred boundaries and different target sizes.

[0071] The Local Feature Refinement (LFR) module, as specifically Figure 4 shown, aims to precisely optimize and enhance the features of local regions, improving the network's ability to locate polyp regions, especially polyp targets with blurred boundaries and large size variations. The LFR module fully explores and focuses on the significant features of the target region through the joint operation of the Spatial Attention (SA) and Channel Attention (CA) mechanisms. The specific calculation formula is as follows:

[0072] Output = σ(MLP(GAP(F)) + MLP(GMP(F)))

[0073] where GAP(F) and GMP(F) represent global average pooling and max pooling operations in the spatial dimension respectively. Through the optimization of the LFR module, the model performs excellently in dealing with segmentation tasks with blurred boundaries and different target sizes, especially suitable for polyp segmentation in colonoscopy images with unclear foreground-background contrast.

[0074] Finally, the outputs from the skip connection part are respectively input into the decoder part. The decoder part uses convolutional operations and bilinear interpolation as the backbone network in the decoding stage. Specifically, the first decoder receives the output from the first LFR for upsampling, and then sends it to the first convolutional block, which is specifically composed of a 3×3 convolution and a ReLU activation function. Subsequently, it is sent to the second upsampling part, and then concatenated with the output of the second LFR. Then it is sent to the second convolutional block, which is specifically composed of a 3×3 convolution and a ReLU activation function. Subsequently, it is sent to the third upsampling part, and then concatenated with the output of the third LFR. Then it is sent to the third convolutional block, which is specifically composed of a 3×3 convolution and a ReLU activation function. Subsequently, it is sent to the fourth upsampling part, and then concatenated with the output of the fourth LFR. Then it is sent to the fourth convolutional block, which is specifically composed of a 3×3 convolution and a ReLU activation function. Finally, the final segmentation result is obtained through a 1×1 convolution and a sigmoid function.

[0075] In the decoder stage, to reduce the number of parameters and computational complexity, we use convolutional operations and bilinear interpolation as the backbone network in the decoding stage, and finally obtain the final output result through a 1x1 convolution. Compared with the traditional transposed convolution operation, the strategy of combining convolutional operations with bilinear interpolation can effectively reduce the computational overhead while maintaining the reconstruction ability of the decoder.

[0076] This decoding structure, while ensuring efficient decoding, uses the efficient upsampling strategy of bilinear interpolation to significantly reduce the parameters and computational complexity of the decoder, thus achieving excellent performance in the image segmentation task.

[0077] The loss function adopted in the training of the present invention is defined as follows:

[0078]

[0079] Among them, TP, TN, FP, and FN respectively represent true positive, true negative, false positive, and false negative.

[0080] S4: Result output module, according to the network training result, give the segmentation result of the lesion area of the colon polyp.

[0081] In the said result output module, in order to comprehensively evaluate the overall performance of the proposed SwinFF-Net in medical image segmentation, three commonly used evaluation metrics are adopted: Dice coefficient, mean intersection over union (mIoU), and accuracy (ACC). These metrics provide a comprehensive evaluation of the model performance in the image segmentation task. The specific calculation formulas are respectively:

[0082]

[0083] Among them, TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative respectively.

[0084] The data in CVC-ClinicDB and Kvasir-SEG are input into the network for training and testing. The experimental results of the SwinFF-Net algorithm in this embodiment are shown in Table 1.

[0085] Table 1

[0086]

[0087] It should be understood that those of ordinary skill in the art can make improvements or transformations according to the above description, and all such improvements and transformations should fall within the protection scope of the appended claims of the present invention.

Claims

1. A colon polyp segmentation system based on multi-scale feature fusion, characterized in that, Including: An original image acquisition module, a preprocessing module, a SwinFF-Net network module, and a result output module. The SwinFF-Net network module includes an encoder, a skip connection module, and a decoder; The encoder includes four feature extraction blocks. The first feature extraction block consists of two Swin Transformer Blocks; the second feature extraction block consists of eight Swin Transformer Blocks, the third feature extraction block consists of four Swin Transformer Blocks, and the fourth feature extraction block consists of two Swin Transformer Blocks; The skip connection module includes: four multi-scale feature fusion modules, a feature enhancement aggregation module, and four local feature optimization modules; the four multi-scale feature fusion modules: respectively receive features from the four feature extraction blocks, The decoder uses convolution operations and bilinear interpolation as the backbone network in the decoding stage, including four decoder units, which respectively receive features from the four local feature optimization modules. The first decoder unit receives the output from the first LFR, performs upsampling, and then sends it to the first convolutional block, then sends it to the second upsampling part, and then concatenates it with the output of the second LFR, and then sends it to the second convolutional block, then sends it to the third upsampling part, and then concatenates it with the output of the third LFR, and then sends it to the third convolutional block, then sends it to the fourth upsampling part, and then concatenates it with the output of the fourth LFR, and then sends it to the fourth convolutional block, and finally obtains the final segmentation result through a 1×1 convolution and a sigmoid function; The result output module gives the segmentation result of the lesion area of the colon polyp according to the network training result.

2. The system according to claim 1, wherein The multi-scale feature fusion module extracts multi-scale features from different receptive fields to enhance the network's perception ability of target boundaries and size changes.

3. The system according to claim 1, characterized in that, The feature enhancement aggregation module performs deep interaction and saliency enhancement on the multi-scale features, thereby further improving the expression ability of edge features and the segmentation effect of foreground regions.

4. The system according to claim 1, wherein The local feature optimization module precisely optimizes and enhances the features of the local region, improving the network's localization ability for polyp regions, especially for polyp targets with blurred boundaries and large size variations; the local feature optimization module fully excavates and focuses on the significant features of the target region through the joint operation of the spatial attention and channel attention mechanisms; the specific calculation formula is as follows: Output = σ(MLP(GAP(F)) + MLP(GMP(F))) Where, GAP(F) and GMP(F) respectively represent global average pooling and max pooling operations performed in the spatial dimension. Through the optimization of the LFR module, the model performs well in processing segmentation tasks with blurred boundaries and different target sizes, especially suitable for polyp segmentation in colonoscopy images with unclear foreground-background contrast.

5. The system according to claim 1, wherein The Swin Transformer Block sequentially includes a first layer normalization unit, a window self-attention unit, a second layer normalization unit, a first multi-layer perceptron unit, a third layer normalization unit, a self-attention unit of a sliding window, a fourth layer normalization unit, and a second multi-layer perceptron unit.

6. The system according to claim 5, characterized in that, First, the input feature map is sent to the first layer normalization unit for layer normalization processing, and then enters the window self-attention for window self-attention calculation; then, the feature is processed through the second layer normalization unit and the first multi-layer perceptron unit, and then normalized through the fourth layer normalization unit, and then enters the self-attention unit of the sliding window for self-attention calculation of the sliding window, and then is processed through the fourth layer normalization unit and the second multi-layer perceptron unit and output.

7. The system according to claim 1, wherein, The Swin Transformer Block uses layer normalization as a preprocessing strategy to ensure the stability of features; In the input stage, first, the feature is layer-normalized and then enters the window multi-head self-attention module; in the window multi-head self-attention module, the input feature generates query Q, key K, and value V through a linear transformation, and the attention weight is calculated through scaled dot-product attention: where d is the scaling factor, and the output of W-MSA is added to the original input through a residual connection to obtain Y: Y = X + W-MSA(LN(X)) Subsequently, Y undergoes layer normalization again and is input into the multi-layer perceptron (MLP). The MLP consists of two fully connected layers and a GELU activation function, and its calculation formula is: MLP(Y) = W2(σ(W1Y + b1)) + b2 Finally, the output of the MLP is fused with Y through a residual connection to obtain the output of the current Swin Transformer: Z = Y + MLP(LN(Y)) After the standard Swin Transformer, a sliding window multi-head self-attention SW-MSA mechanism is introduced to enhance the information interaction between different windows; first, Z is layer-normalized and then input into SW-MSA. This module uses a calculation method similar to W-MSA, but translates the window position before calculating the attention to break the limitation of the fixed window; its calculation method is as follows: Y' = Z + SW-MSA(LN(Z)) Subsequently, Y' enters the MLP for feature transformation: Z' = Y' + MLP(LN(Y')) Finally, Z' is used as the output of the Swin Transformer, which fuses local and global feature information.

8. The system according to claim 2, wherein Atrous convolutions with dilation rates of 1, 2, 3, and 4 are respectively performed in the multi-scale feature fusion module. The receptive field of the convolutional kernel is enlarged through sparse sampling, enabling the network to capture local features and global semantic information. The parallel design of the atrous convolutions ensures feature diversity, capturing both detailed boundary information and preserving global context semantics. The concatenation operation concatenates the outputs of these four dilation rates along the channel dimension, thus combining multi-scale features together to form a high-dimensional feature representation containing rich context information. Finally, the concatenated features are compressed in channels through a 1×1 convolution; Output=Conv 1×1 (Concat(r d＝1 ,r d＝2 ,r d＝3 ,r d＝4 )) Among them, Concat() represents concatenation in the channel dimension, and Conv 1×1 () represents a 1×1 convolution operation, and r d＝1 , r d＝2 , r d＝3 , r d＝4 are the output features of dilated convolutions with dilation rates of 1, 2, 3, and 4 respectively.

9. The system according to claim 1, wherein In the result output module, three evaluation metrics are adopted: Dice coefficient, mean intersection over union mIoU, and accuracy ACC. The calculation formulas are respectively: where TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative respectively.

Images