Colorectal polyp image segmentation method combining mixed attention and semantic complementation

Abstract

The invention belongs to the field of colorectal image segmentation, and particularly relates to a colorectal polyp image segmentation method combining mixed attention and semantic complementation, which comprises the following steps: constructing and training a colorectal polyp segmentation model, and inputting a to-be-segmented colorectal polyp image into the trained colorectal polyp segmentation model to obtain a polyp segmentation graph; the colorectal polyp segmentation model comprises a PVT feature encoder, an AHA module, an SCDB module, an attention guidance decoder and an MSP module; according to the model provided by the invention, the segmentation accuracy of the polyp edge can be effectively improved, the local features and the global features are mutually supplemented, and the generalization of the network is enhanced.

Description

Technical Field

[0001] The present invention belongs to the field of colorectal image segmentation, and in particular relates to a colorectal polyp image segmentation method combining mixed attention and semantic complementarity. Background Art

[0002] As one of the most common cancers worldwide, colorectal cancer poses a serious threat to human health. The disease mainly affects the colon or rectum, has a high mortality rate, and seriously endangers human health. Colorectal cancer usually originates from the growth of benign polyps in the colorectal mucosa, which may be flat or convex in appearance. Colonoscopy is the gold standard for preventing colorectal cancer. The key is to be able to detect and prevent polyps from becoming cancer early. However, during colonoscopy, polyps are sometimes missed due to the subjectivity of manual examination and reliance on clinician experience. In addition, manual examination is also very time-consuming and inefficient. Therefore, in order to improve the speed and accuracy of diagnosis, it is particularly important to develop an effective automatic segmentation algorithm for colorectal polyp images.

[0003] The key to effectively reducing the mortality rate of colorectal cancer is to detect and remove polyps early, especially before they turn malignant. Doctors will carefully observe every part of the colon during colonoscopy in order to remove early lesions that may turn into cancer in time, which significantly reduces the risk of colorectal cancer. However, the identification of small or flat polyps (such as lesions with a diameter of less than 10 mm) is particularly challenging. The complex topological structure of the colon makes the navigation and treatment process difficult and often requires doctors to have advanced skills. In addition, when removing polyps, it is difficult to continuously track the boundaries of the polyps due to the constant changes in the structure, making complete resection more difficult, which is not only time-consuming, but also easily affected by the doctor's subjective judgment. These characteristics make the colorectal polyp segmentation task a challenging task. The existing polyp segmentation models are mainly divided into colorectal polyp segmentation based on multi-scale feature fusion, colorectal polyp segmentation based on attention mechanism, and colorectal polyp segmentation based on Transformer. However, most of the current methods cannot efficiently combine the features output by Transformer with the features output by CNN. Summary of the invention

[0004] To solve the above problems, the present invention provides a colorectal polyp image segmentation method combining hybrid attention and semantic complementarity, including constructing and training a colorectal polyp segmentation model, inputting the colorectal polyp image to be segmented into the trained colorectal polyp segmentation model to obtain a polyp segmentation map; the colorectal polyp segmentation model includes a PVT feature encoder, an AHA module, an SCDB module, an attention-guided decoder, and an MSP module.

[0005] The training process of the colorectal polyp segmentation model includes the following steps:

[0006] S1. Obtaining an original colorectal polyp image training dataset, which includes multiple groups of images; each group of images includes an original colorectal polyp image and its corresponding polyp segmentation map label;

[0007] S2. Input the original colorectal polyp image into the PVT feature encoder to obtain four features, namely F1, F2, F3, and F4;

[0008] S3. Input features F2, F3, and F4 into the attention-guided decoder to obtain the attention prediction feature P att ;

[0009] S4. Predicting feature P based on attention att And features F1, F2, F3, F4, combined with the residual structure to obtain features F1", F2", F3", F4";

[0010] S5. Input the features F1”, F2”, F3”, and F4” into the AHA module to obtain features A1, A2, and A3;

[0011] S6. Perform 1×1 convolution on feature F4” to obtain feature F4”’, input feature F4”’, A1, A2, and A3 into the SCDB module to obtain features S1, S2, and S3;

[0012] S7. Input the features F4”’, S1, S2, and S3 into the MSP module to obtain the feature maps P1, P2, P3, and P4, calculate the loss through the loss function, and back-propagate the training model parameters until the model parameters converge.

[0013] Beneficial effects of the present invention:

[0014] The present invention adopts the PVT (Pyramid Vision Transformer) feature encoder based on the Transformer mechanism as the backbone network of the model, extracts robust multi-scale feature information, including low-level features F1 and high-level features F2-F4, optimizes feature information in space and channel dimensions through the AHA (Adaptive hybrid attention) module; realizes the mutual complementation of global and local features through the SCDB (Semantic Complementary Dual Branch) module; generates preliminary segmentation through the attention-guided decoder to guide the network to focus on important areas in the image in the early stage; and dynamically allocates weights to the segmentation results of multiple models through the MSP (Multi-scale Prediction) module. In view of the shortcomings of the existing polyp segmentation model in boundary refinement processing and the problem that traditional convolution is limited by the local receptive field, the model proposed by the present invention can effectively improve the segmentation accuracy of the polyp edge, let the local features complement the global features, and enhance the generalization of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of the structure of the colorectal polyp segmentation model of the present invention;

[0016] Figure 2 It is a schematic diagram of the structure of the adaptive hybrid attention AHA unit of the present invention;

[0017] Figure 3 It is a structural schematic diagram of the semantically complementary double-branch SCDB unit of the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] The present invention provides a colorectal polyp image segmentation method combining hybrid attention and semantic complementarity, constructs and trains a colorectal polyp segmentation model, and inputs the colorectal polyp image to be segmented into the trained colorectal polyp segmentation model to obtain a polyp segmentation map; the colorectal polyp segmentation model includes a PVT feature encoder, an AHA module, an SCDB module, an attention-guided decoder, and an MSP module.

[0020] like Figure 1 As shown, the training process of the colorectal polyp segmentation model includes the following steps:

[0021] S1. Obtain an original colorectal polyp image training dataset, which includes multiple groups of images; each group of images includes an original colorectal polyp image and its corresponding polyp segmentation map label.

[0022] Specifically, the present invention uses the Kvasir-SEG dataset as the original colorectal polyp image training dataset. The Kvasir-SEG dataset is a colorectal polyp segmentation dataset containing endoscopic images, with a total of 1,000 groups of images, each of which contains an endoscopic image containing polyps (i.e., the original colorectal polyp image as a sample image) and a true value label mask image corresponding to the endoscopic image (i.e., the polyp segmentation image label), which marks the exact position of the polyp in the endoscopic image, and the image resolution ranges from 332×487 to 1920×1072. According to the requirements of experimental training, when the present invention is implemented, the size of all sample images and true value label mask images in the network training experiment is adjusted to 352×352, of which 800 groups of images are used for model training, and the remaining 200 groups of images are used for testing.

[0023] S2. Input the original colorectal polyp image into the PVT feature encoder to obtain four features, namely F1, F2, F3, and F4.

[0024] Specifically, Figure 1 As shown, in the implementation of the present invention, the PVT (Pyramid Vision Transformer) feature encoder is used as the core network architecture, which is mainly used to extract the global semantic features of the image. After an image is input into the PVT feature encoder, a series of multi-scale features of different levels can be obtained. These features increase in level, the corresponding feature map size gradually decreases, and the feature dimension gradually increases. In this embodiment, after the original colorectal polyp image passes through the PVT feature encoder, four levels of features from low to high will be obtained, which are marked as F1, F2, F3, and F4 respectively. The embodiment of the present invention regards F1 as a low-level shallow feature, which contains rich detail information, such as texture, edge, etc.; F2, F3, and F4 are all regarded as high-level deep features, containing abstract high-level semantic information.

[0025] S3. Input features F2, F3, and F4 into the attention-guided decoder to obtain the attention prediction feature P att .

[0026] Specifically, the attention-guided decoder in the embodiment of the present invention is used to generate a segmentation map in the early stage of the network through high-level deep features, and multiply the obtained feature information by the feature output of PVT to guide the network to focus on the area where the polyp is located before further processing the features. The attention-guided decoder fuses feature information from multiple levels through upsampling and convolution modules, with low overhead.

[0027] S4. Predicting feature P based on attention att And features F1, F2, F3, F4, combined with the residual structure to obtain features F1", F2", F3", F4".

[0028] Specifically, Figure 1 As shown, step S4 specifically includes:

[0029] S41. Attention prediction feature P att Multiply feature F1 by point to get feature F1', and add feature F1 and feature F1' to get feature F1".

[0030] S42. Attention prediction feature P att Multiply feature F2 by point to get feature F2', and add feature F2 and feature F2' to get feature F2";

[0031] S43. Attention prediction feature P att Multiply feature F3 by point to get feature F3', and add feature F3 and feature F3' to get feature F3".

[0032] S44. Attention prediction feature P att Multiply feature F4 by point to get feature F4', and add feature F4 and feature F4' to get feature F4".

[0033] S5. Input features F1”, F2”, F3”, F4” into the AHA module to obtain features A1, A2, A3.

[0034] Specifically, the AHA module includes three AHA units with the same structure, and step S5 includes:

[0035] S51. Input feature F1" and feature F2" into the first AHA unit to obtain feature A1;

[0036] S52. Input feature F2" and feature F3" into the second AHA unit to obtain feature A2;

[0037] S53. Input feature F3" and feature F4" into the third AHA unit to obtain feature A3.

[0038] Specifically, Figure 2As shown, each AHA unit in the AHA module constructed by the present invention adaptively adjusts attention from two dimensions, channel and space, so that the features have stronger robustness and representation ability. In the AHA unit, the lower-level features are first merged with the higher-level features to highlight the boundary features in the feature map, exclude the noise in the tissue around the polyp, and use low-overhead self-attention operations to effectively extract the global information of the features. Then, the features are further refined by convolution kernels of three different scales, and the channel and space attention operations are completed in turn, so as to adaptively suppress the redundancy in the features, highlight the features containing useful information, and enhance the network's understanding of the features.

[0039] Specifically, the processing process of each AHA unit includes:

[0040] S511. Set feature F i " and feature F i+1 "Fuse to get the intermediate feature A i ';where i=1,2,3.

[0041] Specifically, step S511 includes:

[0042] The feature F from a higher layer i+1 "The up-sampled features are obtained by up-sampling through 3×3 SC (Spatial and Channel Reconstruction Convolution) convolution, and the up-sampled features are combined with the features of the current level F i "Multiply them to get the fusion features;

[0043] The fused features are passed through 1×1 convolution and Softmax layer to obtain the first feature;

[0044] The fused features are passed through two fully connected layers to obtain the second feature, and a ReLu activation function is connected between the two fully connected layers;

[0045] Multiply the first feature by the second feature to get the third feature; add the third feature to the fusion feature to get the intermediate feature A i '.

[0046] S512. The intermediate feature A i 'Through 3×3 SC convolution, 5×5 convolution and 3×3 dilated convolution, feature A is obtained 3×3 , A 5×5 , A D .

[0047] S513. Feature A 5×5 With feature A DThe first reshaped feature is obtained by concatenating through the Concat function, and then undergoing global average pooling, two layers of fully connected layers, and the Sigmoid activation function.

[0048] S514. The first reshaped feature is respectively combined with feature A 5×5 , Feature A D The dot product obtains the first dot product feature and the second dot product feature; the first dot product feature and the second dot product feature are added to obtain the feature A with channel attention information. C .

[0049] S515. Set feature A 3×3 With feature A C The points are added and fused point by point, and then processed through ReLu activation function, 1×1 convolution, Sigmoid activation function and reshaping operation to obtain the second reshaped feature with spatial attention information.

[0050] S516. The second reshaped feature is respectively combined with feature A 3×3 , Feature A C The third and fourth dot product features are obtained by dot product; the third and fourth dot product features are added together and then convolved by 1×1 to obtain feature A i .

[0051] S6. Perform 1×1 convolution on feature F4” to obtain feature F4”’, and input feature F4”’, A1, A2, and A3 into the SCDB module to obtain features S1, S2, and S3.

[0052] Specifically, the SCDB module includes three SCDB units with the same structure, and step S6 specifically includes:

[0053] S61. Input feature F4"' and feature A3 into the first SCDB unit to obtain feature S1;

[0054] S62. Input feature S1 and feature A2 into the second SCDB unit to obtain feature S2;

[0055] S63. Input feature S2 and feature A1 into the third SCDB unit to obtain feature S3.

[0056] The SCDB (Semantic Complementary Decoder) unit in the example of the present invention combines the features extracted by the convolutional neural network and the Transformer, fuses information of different scales and abstract levels, optimizes the feature representation in the polyp image, ensures that the global and local features complement each other during the decoding process, and further improves the accuracy and robustness of polyp segmentation.

[0057] Preferably, if Figure 3As shown in the figure, the SCDB unit mainly consists of two branches. In the first branch, the input features A1~A3 are divided into multiple blocks, and parallel fully connected layers are used to obtain query Q, key K and value V, respectively, to perform self-attention operations. Since Q and K contain context information, they are processed into spatial information descriptors and channel information descriptors through pooling, fully connected layers and Sigmoid activation functions, and dot multiplication is performed with the self-attention features to obtain feature P t In the second branch, the input features are processed by 1×1, 3×3 and 5×5 convolutions respectively, and concatenated by the Concat function, and then successively passed through convolution, batch normalization and ReLu activation function to obtain the feature P c . t With P c Perform dot multiplication, adjust to a suitable scale through upsampling, and perform dot multiplication with the output of the higher-level SCDB module to obtain the corresponding outputs S1~S3.

[0058] like Figure 3 As shown in the figure, each SCDB unit receives two input features of different scales and processes them to obtain an output feature. The specific process includes:

[0059] S61. Obtain feature P t :

[0060] The input features with smaller scale are divided into multiple sub-blocks to form a sub-block set, and the sub-block set is passed through three fully connected layers to obtain the query Q, key K and value V; self-attention operation is performed on the query Q, key K and value V to obtain the self-attention feature;

[0061] Multiply the query Q and the key K to obtain the context information feature; pass the context information feature through 1 global average pooling, 2 fully connected layers and 1 Sigmoid activation function to obtain the spatial information descriptor; pass the context information feature through 3×3 SC convolution and Sigmoid activation function to obtain the channel information descriptor;

[0062] After multiplying the spatial information descriptor with the self-attention feature, the multiplication result is multiplied again with the channel information descriptor to obtain the feature P t ;

[0063] S62. Obtain feature P c :

[0064] The smaller-scale input features are subjected to 1×1 convolution to obtain convolution features, and the convolution features are subjected to 1×1 convolution, 3×3 convolution, and 5×5 convolution to obtain the first convolution feature, the second convolution feature, and the third convolution feature;

[0065] The first convolution feature, the second convolution feature and the third convolution feature are concatenated to obtain the total concatenation feature, and the total concatenation feature is subjected to 3×3 SC convolution, batch normalization layer and ReLu activation function to obtain feature P c ;

[0066] S63. Set feature P t With feature P c After the dot multiplication, upsampling is performed, and the upsampling result is multiplied with another input feature (i.e., the input feature with a larger scale) to obtain the output feature.

[0067] S7. Input the features F4'', S1, S2, and S3 into the MSP module to obtain weighted feature maps P1-P4. Construct a loss function based on the feature maps P1-P4, calculate the loss through the loss function, and back-propagate the training model parameters until the model parameters converge. When performing the actual segmentation task, merge the feature maps P1-P4 as the original polyp segmentation map P.

[0068] Specifically, according to the attention prediction feature P att The loss function of the colorectal polyp segmentation model is constructed based on the feature maps P1 to P4 and the polyp segmentation map labels corresponding to the original colon polyp image. The parameters of the colorectal polyp segmentation model are updated with the minimum loss function as the optimization goal to complete the training of the colorectal polyp segmentation model. Among them, the loss function L calculation formula is:

[0069] L=L main +L binary

[0070] L main =L P1 +L P2 +L P3 +L P4

[0071]

[0072] L binary =L dice (P att ,G)+L ce (P att ,G)

[0073] Among them, L main represents the main loss function, L binary represents the relay supervision loss function; L P1 Represents the first-level loss function, L P2 Represents the second-level loss function, L P3 Represents the third-level loss function, L P4 Represents the fourth-level loss function; represents the weighted binary cross entropy loss function, represents the weighted intersection-over-union loss function, L dice () represents the Dice coefficient loss function; L ce () represents the cross entropy loss function. G represents the polyp segmentation map label corresponding to the original colon polyp image.

[0074] Preferably, in one embodiment of the present invention, the Pytorch framework is used to implement this model and trained on a Tesla V100 GPU with 32GB of video memory. Our model requires 150 epochs to complete training, and the batchsize of each training is set to 4. We use the Adam optimizer and set the initial learning rate to 1e-4, and use the poly learning rate as a strategy to adjust the learning rate during training. All input images are resized to 352×352 during the training and testing phases. The data augmentation strategies used are random rotation, vertical flipping, and horizontal flipping.

[0075] In the present invention, unless otherwise clearly stipulated and limited, the terms such as "installation", "setting", "connection", "fixation" and "rotation" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral one; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0076] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A colorectal polyp image segmentation method combining hybrid attention and semantic complementarity, characterized in that: Constructing and training a colorectal polyp segmentation model, inputting the colorectal polyp image to be segmented into the trained colorectal polyp segmentation model to obtain a polyp segmentation map; the colorectal polyp segmentation model includes a PVT feature encoder, an AHA module, an SCDB module, an attention-guided decoder, and an MSP module; The training process of the colorectal polyp segmentation model includes the following steps: S1. Obtaining an original colorectal polyp image training dataset, which includes multiple groups of images; each group of images includes an original colorectal polyp image and its corresponding polyp segmentation map label; S2. Input the original colorectal polyp image into the PVT feature encoder to obtain four features, namely F1, F2, F3, and F4; S3. Input features F2, F3, and F4 into the attention-guided decoder to obtain the attention prediction feature P att ; S4. Predicting feature P based on attention att And features F1, F2, F3, F4, combined with the residual structure to obtain features F1", F2", F3", F4"; S5. Input the features F1”, F2”, F3”, and F4” into the AHA module to obtain features A1, A2, and A3; S6. Perform 1×1 convolution on feature F4” to obtain feature F4”’, input feature F4”’, A1, A2, and A3 into the SCDB module to obtain features S1, S2, and S3; S7. Input the features F4”’, S1, S2, and S3 into the MSP module to obtain the feature maps P1, P2, P3, and P4, calculate the loss through the loss function, and back-propagate the training model parameters until the model parameters converge.

2. The colorectal polyp image segmentation method combining hybrid attention and semantic complementarity according to claim 1, characterized in that: Step S4 specifically includes: S41. Attention prediction feature P att Multiply feature F1 by point to get feature F1', and add feature F1 and feature F1' to get feature F1". S42. Attention prediction feature P att Multiply feature F2 by point to get feature F2', and add feature F2 and feature F2' to get feature F2"; S43. Attention prediction feature P att Multiply feature F3 by point to get feature F3', and add feature F3 and feature F3' to get feature F3". S44. Attention prediction feature P att Multiply feature F4 by point to get feature F4', and add feature F4 and feature F4' to get feature F4".

3. The colorectal polyp image segmentation method combining hybrid attention and semantic complementarity according to claim 1, characterized in that: The AHA module includes three AHA units with the same structure, and step S5 includes: S51. Input feature F1" and feature F2" into the first AHA unit to obtain feature A1; S52. Input feature F2" and feature F3" into the second AHA unit to obtain feature A2; S53. Input feature F3" and feature F4" into the third AHA unit to obtain feature A3.

4. The colorectal polyp image segmentation method combining hybrid attention and semantic complementarity according to claim 3, characterized in that: The processing of an AHA unit includes: S511. Set feature F i " and feature F i+1 "Fuse to get the intermediate feature A i ';where i=1,2,3; S512. The intermediate feature A i 'Through 3×3 SC convolution, 5×5 convolution and 3×3 dilated convolution, feature A is obtained 3×3 , A 5×5 , A D ; S513. Feature A 5×5 With feature A D After concatenation, the first reshaped feature is obtained through global average pooling, two fully connected layers, and Sigmoid activation function; S514. The first reshaped feature is respectively combined with feature A 5×5 , Feature A D The dot product obtains the first dot product feature and the second dot product feature; the first dot product feature and the second dot product feature are added to obtain feature A C ; S515. Set feature A 3×3 With feature A C The points are added and fused point by point, and then processed by ReLu activation function, 1×1 convolution, Sigmoid activation function and reshaping operation to obtain the second reshaped feature; S516. The second reshaped feature is respectively combined with feature A 3×3 , Feature A C The third and fourth dot product features are obtained by dot product; the third and fourth dot product features are added together and then convolved by 1×1 to obtain feature A i .

5. The colorectal polyp image segmentation method combining hybrid attention and semantic complementarity according to claim 4, characterized in that: Step S511 specifically includes: The feature F i+1 "The up-sampled features are obtained by up-sampling through 3×3 SC convolution, and the up-sampled features are combined with the feature F i "Multiply them to get the fusion features; The fused features are passed through 1×1 convolution and Softmax layer to obtain the first feature; The fused features are passed through two fully connected layers to obtain the second feature, and a ReLu activation function is connected between the two fully connected layers; Multiply the first feature by the second feature to get the third feature; add the third feature to the fusion feature to get the intermediate feature A i '.

6. The colorectal polyp image segmentation method combining hybrid attention and semantic complementarity according to claim 1, characterized in that: The SCDB module includes three SCDB units with the same structure, and step S6 specifically includes: S61. Input feature F4"' and feature A3 into the first SCDB unit to obtain feature S1; S62. Input feature S1 and feature A2 into the second SCDB unit to obtain feature S2; S63. Input feature S2 and feature A1 into the third SCDB unit to obtain feature S3.

7. The colorectal polyp image segmentation method combining hybrid attention and semantic complementarity according to claim 6, characterized in that: Each SCDB unit receives two input features of different scales and processes them to obtain an output feature. The specific process includes: S61. Obtain feature P t : The input features with smaller scale are divided into multiple sub-blocks to form a sub-block set, and the sub-block set is passed through three fully connected layers to obtain the query Q, key K and value V; self-attention operation is performed on the query Q, key K and value V to obtain the self-attention feature; Multiply the query Q and the key K to obtain the context information feature; pass the context information feature through 1 global average pooling, 2 fully connected layers and 1 Sigmoid activation function to obtain the spatial information descriptor; pass the context information feature through 3×3 SC convolution and Sigmoid activation function to obtain the channel information descriptor; After multiplying the spatial information descriptor with the self-attention feature, the multiplication result is multiplied again with the channel information descriptor to obtain the feature P t ; S62. Obtain feature P c : The smaller-scale input features are subjected to 1×1 convolution to obtain convolution features, and the convolution features are subjected to 1×1 convolution, 3×3 convolution, and 5×5 convolution to obtain the first convolution feature, the second convolution feature, and the third convolution feature; The first convolution feature, the second convolution feature and the third convolution feature are concatenated to obtain the total concatenation feature, and the total concatenation feature is subjected to a 3×3 SC convolution, a batch normalization layer and a ReLu activation function to obtain the feature P c ; S63. Set feature P t With feature P c After the dot multiplication, upsampling is performed, and the upsampling result is multiplied with the input feature with a larger scale to obtain the output feature.

8. The colorectal polyp image segmentation method combining hybrid attention and semantic complementarity according to claim 1, characterized in that: The loss function L is constructed based on the feature graphs P1 to P4 and is expressed as L=L main +L binary L main =L P1 +L P2 +L P3 +L P4 L binary =L dice (P att ,G)+L ce (P att ,G) Among them, L main represents the main loss function, L binary represents the relay supervision loss function; L P1 represents the first-level loss function, L P2 Represents the second-level loss function, L P3 Represents the third-level loss function, L P4 Represents the fourth level loss function; represents the weighted binary cross entropy loss function, represents the weighted intersection-over-union loss function, L dice () represents the Dice coefficient loss function; L ce () represents the cross entropy loss function, and G represents the polyp segmentation map label corresponding to the original colon polyp image.

Images