Coding and decoding polyp image segmentation algorithm based on space channel fusion MobileVit and hole re-parameterization technology

Abstract

The invention relates to the field of image data processing methods, in particular to a coding and decoding polyp image segmentation algorithm based on a space channel fusion MobileVit and cavity re-parameterization technology, which comprises the following steps: acquiring an original medical image; sCM and VRep are used for the preprocessed medical image at the same time to extract spatial features and deep semantic features valuable for polyp segmentation, and a chained spatial convolution attention module (CSCA) is used to capture spatial polyp features and enhance channel polyp features. Then, Pixel Shuffle is combined with GroupConv to carry out up-sampling on the bottom layer features, and the integrity of polyp image segmentation is ensured. And finally, fusing the features obtained by deep up-sampling with the features extracted by VRep, and splicing the shallow up-sampling features with the CSCA enhanced features. The method has high generalization and robustness for polyps of different types of data sets.

Description

Technical Field

[0001] The present invention relates to the field of image data processing methods, and in particular to a polyp segmentation algorithm based on re-parameterization and convolutional block attention. Background Art

[0002] In the field of medical image analysis, polyp image segmentation is of great significance as a key technology for early detection and diagnosis of cancer. Accurate segmentation of polyps is crucial for the prevention and screening of rectal cancer. Through early detection by imaging doctors, malignant polyps can be treated in time; for high-risk polyps, medical image segmentation not only helps preoperative planning, but also provides a basis for doctors to choose the most appropriate surgical method and treatment plan. In addition, by comparing the segmentation results before and after treatment, the patient's recovery can be effectively evaluated.

[0003] In early screening, accurately judging the status of colorectal polyps has an important impact on the overall health of patients. As one of the precursor lesions of rectal cancer, the changes of colorectal polyps are closely related to rectal cancer. Area expansion, structural abnormalities, surface serrations, local red edges and lymph node enlargement may all be signs of malignant transformation. Similarly, in the early stages of other cancers such as gastric cancer and endometrial cancer, the status of polyps also provides important clues for clinical diagnosis. However, with the increasing attention paid to health in modern society, medical institutions have accumulated a large amount of medical imaging data, especially the surge in the number of polyp images. In contrast, the resources of imaging doctors are relatively limited, and relying on manual observation can no longer meet the needs of modern medicine for polyp identification.

[0004] With the development of artificial intelligence technology, more and more studies are applying it to medical image analysis, and auxiliary diagnosis has gradually become an important means. Artificial intelligence-assisted diagnosis can not only reduce errors caused by factors such as doctor fatigue, but also effectively reduce the workload of doctors and help to efficiently diagnose large-scale diseases. In this context, deep learning, as a cutting-edge technology in the field of artificial intelligence, has shown significant progress in medical image analysis. It has high adaptability, powerful large-scale data processing capabilities, deep abstract feature learning capabilities, and excellent generalization capabilities, making it an ideal tool for solving difficult problems such as diversity, complexity, and individual differences in polyp image segmentation.

[0005] Although deep learning has shown many advantages in the field of polyp segmentation, it still faces some challenges. First, the similarity within polyp images is high, especially in terms of color, contour and background, which poses a challenge to the boundary accuracy of the segmentation model. Second, the differences between images are large, which come from different devices, lighting environments and image types, which limits the generalization ability of the model. To address these problems, researchers have proposed a variety of methods, such as introducing attention mechanisms and densely connected dilated convolution pooling pyramid modules into the DoubleUNet-based segmentation algorithm to improve feature extraction capabilities. However, such methods are prone to overfitting, contour misjudgment and decreased generalization performance when dealing with high-density convolution models. In order to improve generalization and robustness, researchers also introduced multi-scale semantic embedding modules and residual modules, and expanded the receptive field through dilated convolution. However, these methods still have certain limitations in extracting detailed features and fail to give full play to the advantages of dilated convolution.

[0006] In the field of medical image analysis, polyp image segmentation faces many challenges, including high intra-class similarity, significant inter-class differences, difficulty in capturing global context information, and loss of edge details. To address these issues, researchers have proposed a variety of innovative solutions.

[0007] As a lightweight visual transformation model, MobileViT combines the advantages of CNN and Transformer and has strong global and local feature capture capabilities. In polyp image segmentation, MobileViT uses its compact structure and global context modeling capabilities to improve segmentation accuracy and computational efficiency. Its main advantages are that it can handle multi-scale features, is highly adaptable and friendly to small devices, and is suitable for mobile devices. However, due to its lightweight model design, MobileViT may not be perfect in capturing fine-grained features and processing boundary details when processing complex and diverse medical images, and has certain limitations.

[0008] Other deep learning-based methods, such as ResUNet++, introduce residual blocks, attention mechanisms and other modules to enhance the model's ability to perceive polyps. BDG-Net improves the segmentation accuracy of polyps of different sizes through a multi-scale interaction strategy. Another method expands the convolution receptive field through the Atrous spatial pyramid module to improve the performance of the model at the same size. Enhancing the backbone network and replacing the convolutional network of HarDNet-DFUS also improves the robustness of the model.

[0009] In addition, the attention mechanism has been widely used in polyp segmentation. For example, the network based on CASCadedAttentionDecoder improves the segmentation performance by improving the ability to capture local background information. TGANet focuses on the representation learning of polyps of different sizes, and the ConvNext backbone network enhances the ability to capture contextual information by solving the semantic gap problem between the encoder and decoder.

[0010] Transformer-based models, such as TransNetR and SSFormer, have also achieved good performance in this area. TransNetR further improves the generalization ability of the model based on ResUNet++, while SSFormer enhances the model's grasp of global context and detail information by combining the pyramid Transformer encoder with the progressive decoder. Summary of the invention

[0011] The present invention aims to provide a coding and decoding polyp image segmentation algorithm based on MobileVit of spatial channel fusion and hole reparameterization technology to solve the problem of unbalanced segmentation accuracy of the existing algorithms in the same category polyp segmentation accuracy and multi-category segmentation accuracy.

[0012] The polyp segmentation algorithm based on reparameterization and convolutional block attention in this scheme includes the following steps:

[0013] S1: Get the original image;

[0014] S2: Scale and fill the polyp images of different sizes in the original image with gray edges, perform data enhancement using operations such as rotation and cropping, and finally perform one-hot encoding;

[0015] S3: The spatial channel fusion MobileVit (SCM) module is used to extract the robust features of different receptive fields of the processed medical image, and the SCM module is used to downsample the medical image, with a total of five partial features;

[0016] S4: In the first two partial features captured by SCM, the chained convolutional attention module (CSCA) is used to enhance the spatial channel level features of medical images;

[0017] S5: Among the last three partial features captured by SCM, the void reparameterization module (VRep) is used to deeply extract medical image features;

[0018] S6: Use PixelShuffle upsampling combined with group convolution module to upsample the features that pass through the bottleneck structure, and then fuse the feature content obtained by VRep deep feature extraction module and CSCA global channel feature module;

[0019] S7: Save the weight parameters of the dilated reparameterized convolutional block attention module network, and reparameterize the saved weight parameters using the dilated reparameterization method;

[0020] S8: Modify the downsampling of the void reparameterization module to the void reparameterization prediction module, and use this model to segment the original polyp medical images without labels;

[0021] S9: Calculate mIou, mDice, mPre and mRec of the segmented image and its label content;

[0022] The beneficial effects of this program are:

[0023] The proposed algorithm model uses the SCM algorithm to achieve lightweight high-performance feature extraction, and captures basic polyp medical image feature information in two dimensions of spatial channels. At the same time, shallow CSCA is used to enhance the polyp edge features and polyp spatial features for global and inductive bias feature enhancement. The deep hole reparameterization structure is used to perform multi-scale feature fusion on the deep feature part extracted by the model to enhance the model's ability to extract polyp feature information at different scales. Finally, the group volume combined with the pixelShuffle upsampling module is used to achieve pixel-level segmentation image restoration. The proposed algorithm model can always obtain high mDice scores in all tests, so it represents high performance within the evaluation range of this indicator. The improvement in performance is the result of many factors, including: the basic polyp feature extraction of the convolution and spatial channel chain attention in the SCM module, while improving the problem of feature extraction omission when using convolution and attention alone; the powerful feature extraction ability and excellent receptive field of the hole re-parameterized convolution module make it an excellent encoder feature enhancement module; similarly, the CSCA module in the deep skip connection uses the inductive bias to enhance the ability of complete polyp extraction from the spatial dimension, and performs global feature analysis and evaluation of the target features from the channel dimension. The reasons for the improvement in the performance of the algorithm model can be obtained through different combinations and separate tests. In this work, an extremely important discovery is that using the SCM module as the main downsampling method can significantly improve the extraction efficiency of the model features, and the hole re-parameterized network structure as a model enhancement effect is very significant. In the image segmentation task, the algorithm model can more accurately extract the features that the decoder needs to segment. The problem of the imbalance between the same-category polyp segmentation accuracy and the multi-category segmentation accuracy of the polyp medical image segmentation algorithm is solved, and the problem of excessive consumption of polyp medical image models, training time and equipment costs, etc. is solved.

[0024] Furthermore, in S2, the polyp images of different sizes in the original image are scaled and filled with gray edges, and data enhancement is performed using operations such as rotation and cropping, and finally a One-hot encoding operation is performed.

[0025] The beneficial effect is that in the deep learning polyp segmentation task, the unified size and one-hot encoding processing methods have significant advantages. First, adjusting the original image to a uniform size helps to simplify the model design, enabling the neural network to process input images of consistent size, thereby reducing computational complexity and improving training and inference efficiency. Secondly, one-hot encoding provides an efficient label representation method in multi-category problems. By representing the label of each pixel in vector form, one-hot encoding can be easily compared directly with the network output, and then calculate the loss and perform back propagation. This processing method not only adapts to the deep learning framework, but also simplifies the task processing process, allowing the network to be more easily extended to complex multi-category problems. The combination of uniform size and one-hot encoding helps to improve the performance and adaptability of the model.

[0026] Furthermore, in S3, a spatial channel fusion MobileVit (SCM) module is used to extract robust features of different receptive fields of the processed medical image, and the SCM module is used to downsample the medical image, resulting in a total of five partial features.

[0027] The beneficial effect is that each layer in the downsampling block in the algorithm takes the unique output of the previous layer as the current input and transmits this output information to the next layer. It is very difficult to achieve effective feature extraction using a single convolution or attention mechanism, which often requires a very deep network structure, but the SCM module that uses attention and convolution fusion can effectively solve this problem. The spatial channel fusion MobileViT block is similar to convolution, but uses a deeper global processing to replace the local processing in convolution, that is, replaces the matrix multiplication operation in convolution. At the same time, the spatial attention mechanism and channel attention mechanism are used in this local processing to make up for the multi-channel global feature information that traditional convolution cannot take into account. The global processing capability here also becomes the reason for the lightweight model, and it does not need to use too many parameters for learning. At the same time, under the action of the residual structure, the model retains more feature information, making the model more robust, improving the accuracy of the model at the level of local CNN and global ViT features, and effectively utilizing the strong correlation computing power of the inductive bias of convolution and attention.

[0028] Further, in S4, in the first two partial features captured by SCM, a chain convolutional attention module (CSCA) is used to enhance the spatial channel level features of the medical image. CSCA is a convolutional attention module for enhancing the performance of convolutional neural networks. It combines channel self-attention and spatial convolutional attention mechanisms to improve the representation ability of features by learning to effectively allocate channel and spatial attention weights; channel self-attention adaptively adjusts the feature response on the channel dimension from a global level to better capture the relationship between different channels in the world, and spatial attention is used to adaptively adjust the feature response on the spatial dimension to better capture the relationship between different positions; the channel attention part vectorizes the features, captures the global correlation of each feature channel, so as to determine the importance of each channel in the spatial position of another channel, and then uses a global average pooling operation to reset the unnecessary feature map channel weights to 0. Finally, these channels are shrunk to adapt to the original feature map shape. The CSCA derivation process here can be summarized as follows:

[0029]

[0030] F′=M c (F)·F

[0031] M s (F′)=σ(Conv([AvgPool(F′),MaxPool(F′)])The beneficial effect is that CSCA enhances the network's ability to express local information by adaptively adjusting the responses of different channels and different spatial positions in the feature map, extracts richer and more useful features, and suppresses the responses of unimportant positions. This solves the problem of weak global expression of redundant channel information and the problem of information occlusion and imbalance in spatial positions. Therefore, the background with a high similarity to the polyp is suppressed, and the edge segmentation ability of the polyp image is enhanced, thereby forming an accurate polyp segmentation solution.

[0032] Further, in S5, the void reparameterization module (VRep) is used to perform deep feature extraction on the medical image features in the last three partial features captured by the SCM. The use of the VRep module to perform deep feature extraction on the medical image features is mainly reflected in the use of the following during training: Figure 2 The multi-branch structure in (a) includes 5×5 hole convolution connection BatchNormal, 3×3 convolution connection BatchNormal, 1×1 convolution connection BatchNormal and a separate BatchNormal branch. When performing inference operations, the following is used: Figure 2(b) The single topology structure of the RepBlock. When ignoring the activation layer, the equivalent process of reparameterization of the RepBlock during prediction is as follows: Figure 2 As shown in (b), first Figure 2 Each branch structure in (a) is converted into a 3×3 convolution with a bias vector. After adding the multi-branch information, a 3×3 convolution carrying all bias vectors is obtained. That is, when the model makes predictions, the information redundancy generated by the multi-branch topology in the training network can be eliminated. In this way, the multi-branch structure is decoupled into a single structure, but the effect of the multi-branch structure is still retained, and only its parameters are integrated. The feature fusion formula is:

[0033] M (2) =b n (M (1) *w (5) ,μ (5) ,σ (5) ,γ (5) ,β (5) )

[0034] =b n (M (1) *w (3) ,μ (3) ,σ (3) ,γ (3) ,β (3) )

[0035] =b n (M (1) *w (1) ,μ (1) ,σ (1) ,γ (1) ,β (1) )

[0036] =b n (M (1) ,μ (0) ,σ (0) ,γ (0) ,β (0) )

[0037] Therefore, the main body of RepBlock in this algorithm can be regarded as only one type of operator: 3×3conv followed by Relu. The design of the dilated reparameterized convolution layer is inspired by the techniques of parameter sharing and matrix multiplication. Through the dilated reparameterized convolution layer, the traditional multi-loop calculation is converted into an efficient matrix multiplication operation, which significantly improves the calculation speed. While maintaining accuracy, the dilated reparameterized convolution layer obtains a similar calculation speed to ordinary convolution, and at the same time has a wider receptive field, making the entire algorithm efficient and superior in the inference stage. Therefore, it can play a very important role in improving the accuracy of polyp image segmentation.

[0038] Figure 2 The derivation process of the re-parameter block in (c) is shown in Formula 1. The projection vector is calculated for the input feature map after the ordinary convolution operation. The projection vector calculation formula is as follows:

[0039]

[0040] Among them, O(:,:,c) represents the cth channel of the output feature map, and P(c,k) represents the weight of the 1×1 convolution kernel in channel c and output channel k. After obtaining the projection vector, it is necessary to merge the feature map and the projection vector, and finally reorganize the channels to reorganize the new feature map into a feature map with three parts: the first part is the first c_out channels of the original output feature map, the second part is the c_out channels of the original output feature map, and the third part is the last c_out channels of the projection vector. Combined into the final output feature map.

[0041] The beneficial effect is: this design replaces the standard convolution layer with a discretized convolution layer, and adopts a single-branch linear operation, using different parameters to affine transform all feature maps in the input space. Its specific manifestation in polyp image segmentation is that a larger receptive field gives the model better feature extraction capabilities. The design of this module aims to enhance the model's ability to extract spatial and semantic information from the polyp area, while increasing the model's receptive field, so that the network has higher segmentation accuracy and generalization capabilities.

[0042] Furthermore, in S6, PixelShuffle upsampling is combined with a group convolution module to upsample the features passing through the bottleneck structure, and then the features are fused with the feature content obtained by the VRep deep feature extraction module and the CSCA global channel feature module.

[0043] The beneficial effect is that PixelShuffle upsampling combined with group convolution module has significant advantages in deep learning tasks. PixelShuffle is an efficient upsampling method that generates high-resolution feature maps by rearranging the channel dimensions of low-resolution feature maps, avoiding interpolation operations in traditional upsampling methods, and better retaining details and texture information. In contrast, group convolution can effectively reduce the number of parameters and the amount of calculation by grouping convolutions, thereby improving computational efficiency and reducing the risk of overfitting. When PixelShuffle is used in combination with group convolution, this combination can further optimize the utilization of computing resources while ensuring the quality of upsampling. The high-resolution feature maps generated by PixelShuffle contain more detailed information, while group convolution can efficiently process these feature maps through grouping operations to achieve more complex feature extraction. This combination can not only improve the performance of the model, but also reduce the computational complexity during training and inference, and is suitable for resource-constrained application scenarios. In addition, the combination of PixelShuffle and group convolution can achieve better image reconstruction effects while keeping the model lightweight. This method is particularly suitable for tasks that require high-resolution output, such as super-resolution and image segmentation, and achieves excellent performance through fine feature processing and efficient calculation. Overall, the combination of PixelShuffle and group convolution takes advantage of the advantages of both and achieves efficient, accurate, and resource-friendly image processing in polyp segmentation.

[0044] Furthermore, in S8, the pre-trained re-parameterized convolution block attention parameters are used to perform polyp segmentation on the unknown segmented image.

[0045] The beneficial effects are: using the SCM algorithm to achieve lightweight high-performance feature extraction, and capturing basic polyp medical image feature information in two dimensions of spatial channels at the same time, using shallow CSCA to enhance the polyp edge features and polyp spatial features for global and inductive bias feature enhancement, using deep hole reparameterization structure to perform multi-scale feature fusion on the deep feature part extracted by the model, enhancing the model's ability to extract feature information of polyps at different scales, and finally using the group volume combined with the pixelShuffle upsampling module to achieve pixel-level segmentation image restoration. The prediction effect performed in this algorithm has approached or even exceeded the maximum achievable performance, so it further highlights the importance of studying the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic diagram of the overall architecture of an embodiment of the coding and decoding polyp image segmentation algorithm based on MobileVit and hole re-parameterization technology of spatial channel fusion of the present invention;

[0047] Figure 2 It is a hollow re-parameterized network structure diagram of an embodiment of a coding and decoding polyp image segmentation algorithm based on MobileVit of spatial channel fusion and hollow re-parameterization technology of the present invention;

[0048] Figure 2 (a) is a network structure diagram of the hollow re-parameterization module training down-sampling in the embodiment of the coding and decoding polyp image segmentation algorithm based on MobileVit and hollow re-parameterization technology of spatial channel fusion of the present invention;

[0049] Figure 2 (b) is a flow chart of the conversion of the hole re-parameterization module from the training network structure to the test network structure in the embodiment of the coding and decoding polyp image segmentation algorithm based on MobileVit and the hole re-parameterization technology of spatial channel fusion of the present invention;

[0050] Figure 2 (c) is a diagram showing the transformation of the hole re-parameterization module from the training network structure to the test network interpretation structure in the embodiment of the coding and decoding polyp image segmentation algorithm based on the MobileVit and hole re-parameterization technology of spatial channel fusion of the present invention, that is, Figure 2 (b) Matrix explanation diagram;

[0051] Figure 3 It is an architecture diagram of the chained convolution block attention module in the embodiment of the encoding and decoding polyp image segmentation algorithm based on MobileVit with spatial channel fusion and hole re-parameterization technology of the present invention;

[0052] Figure 4 It is a structural diagram of the spatial channel fusion MobileViT in the embodiment of the coding and decoding polyp image segmentation algorithm based on the spatial channel fusion MobileVit and the hole re-parameterization technology of the present invention.

Claims

1. A coding and decoding polyp image segmentation algorithm based on MobileVit with spatial channel fusion and hole reparameterization technology, characterized in that: The following steps are included (assuming the original medical image is F(M×M): S1: Get the original image; S2: Scale and fill the polyp images of different sizes in the original image with gray edges, perform data enhancement using operations such as rotation and cropping, and finally perform one-hot encoding; S3: The spatial channel fusion MobileVit (SCM) module is used to extract the robust features of different receptive fields of the processed medical image, and the SCM module is used to downsample the medical image, with a total of five partial features; S4: In the first two partial features captured by SCM, the chained convolutional attention module (CSCA) is used to enhance the spatial channel level features of medical images; S5: Among the last three partial features captured by SCM, the void reparameterization module (VRep) is used to deeply extract medical image features; S6: Use PixelShuffle upsampling combined with group convolution module to upsample the features that pass through the bottleneck structure, and then fuse the feature content obtained by VRep deep feature extraction module and CSCA global channel feature module; S7: Save the weight parameters of the dilated reparameterized convolutional block attention module network, and reparameterize the saved weight parameters using the dilated reparameterization method; S8: Modify the downsampling of the void reparameterization module to the void reparameterization prediction module, and use this model to segment the original polyp medical images without labels; S9: Calculate mIou, mDice, mPre and mRec of the segmented image and its label content; In step S2, the polyp image and the label image are uniformly scaled, gray-edge filled, and data enhanced using traditional visual processing methods, and then one-hot encoded; In step S3, the SCM module is used to extract the robust features of different receptive fields of the preprocessed medical image, and the SCM module is used to downsample the medical image to obtain the next step features. The spatial channel fusion MobileVit module is mainly reflected in: the spatial channel fusion MobileViT block is similar to convolution, but uses a deeper global processing to replace the local processing in convolution, that is, it replaces the matrix multiplication operation in convolution. At the same time, the spatial attention mechanism and channel attention mechanism are used in this local processing to make up for the multi-channel global feature information that traditional convolution cannot take into account. The global processing capability here also becomes the reason for the lightweight model, and does not require too many parameters for learning. At the same time, under the action of the residual structure, the model retains more feature information, making the model more robust, improving the accuracy of the model at the level of local CNN and global ViT features, and effectively utilizing the convolutional inductive bias and the strong correlation computing power of attention.

2. The encoding and decoding polyp image segmentation algorithm based on MobileVit and cavity re-parameterization technology of spatial channel fusion according to claim 1 is characterized by: In step S4, the shallow medical image features captured by SCM are processed using the CSCA module to further obtain spatial features and semantic features in the features. CSCA is a convolutional attention module used to enhance the performance of convolutional neural networks. It combines channel self-attention and spatial convolutional attention mechanisms to improve the representation ability of features by learning to effectively allocate channel and spatial attention weights; channel self-attention adaptively adjusts the feature response on the channel dimension from a global level to better capture the relationship between different channels in the world, and spatial attention is used to adaptively adjust the feature response on the spatial dimension to better capture the relationship between different positions; the channel attention part vectorizes the features and captures the global correlation of each feature channel to determine the importance of each channel in the spatial position of another channel, and then uses a global average pooling operation to reset the unnecessary feature map channel weights to 0. Finally, these channels are shrunk to adapt to the original feature map shape. The CSCA derivation process here can be summarized as follows: F′=M c (F)·F M s (F′)=σ(Conv([AvgPool(F′),MaxPool(F′)])。 3. The encoding and decoding polyp image segmentation algorithm based on MobileVit and hole re-parameterization technology of spatial channel fusion according to claim 1 is characterized by: In step S5, the VRep module is used to perform deep feature extraction on medical image features, which is mainly reflected in the use of a multi-branch structure as shown in Figure 2(a) during training, including 5×5 dilated convolution connected to BatchNormal, 3×3 convolution connected to BatchNormal, 1×1 convolution connected to BatchNormal, and a separate BatchNormal branch. When performing inference operations, a single topology structure as shown in Figure 2(b) is used. Ignoring the activation layer, the equivalent process of reparameterization of RepBlock during prediction operations is shown in Figure 2(b). First, each branch structure in Figure 2(a) is converted into a 3×3 convolution with a bias vector. After adding the multi-branch information, a 3×3 convolution carrying all bias vectors is obtained. That is, when the model makes predictions, the information redundancy generated by the multi-branch topology in the training network can be eliminated. In this way, the multi-branch structure is decoupled into a single structure, but the effect of the multi-branch structure is still retained, and only its parameters are integrated. The feature fusion formula is: M (2) =b n (M (1) *w (5) ,m (5) ,s (5) ,c (5) ,b (5) ) =b n (M (1) *w (3) ,m (3) ,s (3) ,c (3) ,b (3) ) =b n (M (1) *W (1) ,m (1) ,s (1) ,c (1) ,b (1) ) =b n (M (1) ,m (0) ,s (0) ,c (0) ,b (0) ) Therefore, the main body of RepBlock in this algorithm can be regarded as only one type of operator: 3×3conv followed by Relu. The design of the dilated reparameterized convolution layer is inspired by the techniques of parameter sharing and matrix multiplication. Through the dilated reparameterized convolution layer, the traditional multi-loop calculation is converted into an efficient matrix multiplication operation, which significantly improves the calculation speed. While maintaining accuracy, the dilated reparameterized convolution layer obtains a similar calculation speed to ordinary convolution, and at the same time has a wider receptive field, making the entire algorithm efficient and superior in the inference stage. Therefore, it can play a very important role in improving the accuracy of polyp image segmentation. The derivation process of the re-parameter block in Figure 2(c) is shown in Formula 1. The projection vector is calculated for the input feature map after the ordinary convolution operation. The projection vector calculation formula is as follows: Among them, O(:,:,c) represents the cth channel of the output feature map, and P(c,k) represents the weight of the 1×1 convolution kernel in channel c and output channel k. After obtaining the projection vector, it is necessary to merge the feature map and the projection vector, and finally reorganize the channels to reorganize the new feature map into a feature map with three parts: the first part is the first c_out channels of the original output feature map, the second part is the c_out channels of the original output feature map, and the third part is the last c_out channels of the projection vector. Combined into the final output feature map.

Images