Colorectal polyp image segmentation method combining mixed attention and semantic complement

Abstract

The application belongs to the field of colorectal image segmentation, and particularly relates to a colorectal polyp image segmentation method combining mixed attention and semantic complementation, comprising the following steps: constructing and training a colorectal polyp segmentation model, inputting a colorectal polyp image to be segmented into the trained colorectal polyp segmentation model to obtain a polyp segmentation image; the colorectal polyp segmentation model comprises a PVT feature encoder, an AHA module, an SCDB module, an attention-guided decoder and an MSP module; the model can effectively improve the segmentation accuracy of polyp edges, complement local features and global features, and enhance the generalization of the network.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of colorectal image segmentation, and particularly relates to a colorectal polyp image segmentation method combining mixed attention and semantic complementation. BACKGROUND

[0002] Colorectal cancer is one of the most common cancers worldwide, posing a serious threat to human health. It 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, and the key lies in early detection and prevention of polyp canceration. However, during colonoscopy, sometimes polyps are missed due to the subjectivity of manual examination and reliance on the experience of clinicians. 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 as early as possible, especially before they turn malignant. Doctors carefully observe each part of the colon during colonoscopy in order to timely remove precancerous lesions that may turn into cancer, significantly reducing the risk of colorectal cancer. However, the identification of small or flat polyps (such as lesions less than 10 mm in diameter) is particularly challenging. The complex topology of the colon makes navigation and treatment 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 changing structure, making complete resection more difficult, which not only consumes time but is also susceptible to subjective judgment by doctors. These characteristics make the colorectal polyp segmentation task a challenging task. Existing polyp segmentation models mainly include multi-scale feature fusion-based colorectal polyp segmentation, attention mechanism-based colorectal polyp segmentation, and Transformer-based colorectal polyp segmentation. However, most current methods cannot efficiently combine the features output by the Transformer with the features output by the CNN. SUMMARY

[0004] To solve the above problems, the application provides a colorectal polyp image segmentation method combining mixed attention and semantic complementation, which comprises constructing and training a colorectal polyp segmentation model, inputting a colorectal polyp image to be segmented into the trained colorectal polyp segmentation model to obtain a polyp segmentation map; the colorectal polyp segmentation model comprises 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 comprises the following steps:

[0006] S1. Obtain an original colorectal polyp image training data set, which comprises a plurality of groups of images; each group of images comprises an original colorectal polyp image and a corresponding polyp segmentation label thereof;

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

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

[0009] S4. Obtain features F1'', F2'', F3'' and F4'' based on the attention prediction feature P att and the features F1, F2, F3 and F4 in combination with a residual structure;

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

[0011] S6. Obtain a feature F4''' by performing 1*1 convolution on the feature F4'', and input the features F4''', A1, A2 and A3 into an SCDB module to obtain features S1, S2 and S3;

[0012] S7. Input the features F4''', S1, S2 and S3 into an MSP module to obtain feature maps P1, P2, P3 and P4, calculate a loss by using a loss function, and train model parameters by back propagation until the model parameters converge.

[0013] The colorectal polyp segmentation model has the following beneficial effects:

[0014] The present invention adopts the PVT (Pyramid Vision Transformer) feature encoder based on the Transformer mechanism as the backbone network of the model to extract robust multi-scale feature information, including low-level features F1 and high-level features F2~F4. The feature information is optimized in the spatial and channel dimensions through the AHA (Adaptive hybrid attention) module; the global and local features are complemented by the SCDB (Semantic Complementary Dual Branch) module; the network is guided to focus on important areas in the image in the early stage by generating preliminary segmentation through the attention-guided decoder; and the segmentation results of multiple models are dynamically weighted by the MSP (Multi-scale Prediction) module. In response to the shortcomings of existing polyp segmentation models 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 polyp edges, let local features and global features complement each other, 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 Schematic diagram of the structure of the adaptive hybrid attention AHA unit of the present invention;

[0017] Figure 3 Schematic diagram of the structure of the semantically complementary double-branch SCDB unit of the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0019] The present invention provides a colorectal polyp image segmentation method that combines 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 a raw colorectal polyp image training data set, which includes a plurality of groups of images; each group of images includes a raw colorectal polyp image and a corresponding polyp segmentation label.

[0022] Specifically, the present application adopts the Kvasir-SEG dataset as the raw colorectal polyp image training data set. The Kvasir-SEG dataset is a colorectal polyp segmentation dataset containing endoscopic images, a total of 1000 groups of images, each group of images containing 1 endoscopic image containing a polyp (i.e. the raw colorectal polyp image as the sample image) and 1 true value label mask image (i.e. the polyp segmentation label) corresponding to the endoscopic image, which labels the accurate position of the polyp in the endoscopic image, and the image resolution ranges from 332x487 to 1920x1072. According to the experimental training needs, the size of all sample images and true value label mask images is adjusted to 352x352 in the network training experiment during the implementation of the present application, 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 raw colorectal polyp image into the PVT feature encoder to obtain four features, F1, F2, F3, and F4.

[0024] Specifically, as shown in Figure 1 In the implementation of the present application, a 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 inputting an image into the PVT feature encoder, a series of multi-scale features at 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, the raw colorectal polyp image will obtain four levels of features from low to high after passing through the PVT feature encoder, which are marked as F1, F2, F3, and F4. In the present embodiment, F1 is regarded as a low-level shallow feature, which contains rich detailed information such as texture and edge; F2, F3, and F4 are all regarded as high-level deep features, which contain abstract high-level semantic information.

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

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

[0027] S4. input the attention prediction feature P into the feature F1, F2, F3, F4 to obtain features F1”, F2”, F3”, F4”. att and features F1, F2, F3, F4, and obtain features F1”, F2”, F3”, F4” by combining the residual structure.

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

[0029] S41. multiply the attention prediction feature P by the feature F1 to obtain a feature F1’, and add the feature F1 and the feature F1’ to obtain a feature F1”; att

[0030] S42. multiply the attention prediction feature P by the feature F2 to obtain a feature F2’, and add the feature F2 and the feature F2’ to obtain a feature F2”; att

[0031] S43. multiply the attention prediction feature P by the feature F3 to obtain a feature F3’, and add the feature F3 and the feature F3’ to obtain a feature F3”; att

[0032] S44. multiply the attention prediction feature P by the feature F4 to obtain a feature F4’, and add the feature F4 and the feature F4’ to obtain a feature F4”; att

[0033] S5. input the 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 the feature F1” and the feature F2” into the first AHA unit to obtain a feature A1;

[0036] S52. input the feature F2” and the feature F3” into the second AHA unit to obtain a feature A2;

[0037] S53. input the feature F3” and the feature F4” into the third AHA unit to obtain a feature A3.

[0038] Specifically, as shown in Figure 2 ​​​​As shown, each AHA unit in the AHA module constructed by the present application adaptively adjusts attention from two dimensions of channels and spaces, so that the features have stronger robustness and representation ability. In the AHA unit, the relatively low-level features are first fused with the higher-stage features to highlight the boundary features in the feature map, exclude the noise in the tissue around the polyp, and effectively extract the global information of the features by using a low-overhead self-attention operation. Then, the features are further refined by three convolution kernels of different scales, and the attention operations in the channels and spaces are sequentially completed, so as to adaptively suppress the redundancy in the features, highlight the features containing useful information, and strengthen the understanding of the network to the features.

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

[0040] S511. Fusing the features F i and F i+1 to obtain an intermediate feature A i '; wherein i = 1, 2, 3.

[0041] Specifically, step S511 specifically includes:

[0042] The features F i+1 from the higher layer are obtained by 3x3 SC (Spatial and Channel Reconstruction Convolution) convolution and then up-sampling to obtain up-sampling features, and the up-sampling features are multiplied with the features F i of the current level to obtain fused features.

[0043] The fused features are obtained by 1x1 convolution and Softmax layer to obtain first features;

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

[0045] The first features are multiplied with the second features to obtain third features, and the third features are added with the fused features to obtain the intermediate feature A i '.

[0046] S512. The intermediate feature A i is respectively obtained by 3x3 SC convolution, 5x5 convolution and 3x3 hole convolution to obtain features A 3×3 , A 5×5 and A D .

[0047] S513. The features A 5×5 and A DThe first remodeling feature is obtained by concatenation, global average pooling, two fully connected layers, and Sigmoid activation function operations.

[0048] S514. The first remodeling feature is respectively multiplied with the feature A 5×5 , the feature A D , to obtain the first point multiplication feature and the second point multiplication feature; the first point multiplication feature and the second point multiplication feature are added to obtain the feature A C with channel attention information.

[0049] S515. The feature A 3×3 is pointwisely added with the feature A C , and then processed by a ReLu activation function, a 1*1 convolution, a Sigmoid activation function and a remodeling operation to obtain the second remodeling feature with spatial attention information.

[0050] S516. The second remodeling feature is respectively multiplied with the feature A 3×3 , the feature A C , to obtain the third point multiplication feature and the fourth point multiplication feature; the third point multiplication feature and the fourth point multiplication feature are added and then processed by a 1*1 convolution to obtain the feature A i .

[0051] S6. The feature F4” is processed by a 1*1 convolution to obtain a feature F4”’, and the feature F4”’, A1, A2 and A3 are input into an 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. The feature F4”’ and the feature A3 are input into a first SCDB unit to obtain the feature S1;

[0054] S62. The feature S1 and the feature A2 are input into a second SCDB unit to obtain the feature S2;

[0055] S63. The feature S2 and the feature A1 are input into a third SCDB unit to obtain the feature S3.

[0056] The SCDB (Semantic Complementary Decoder) unit in the example of the present application 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 in the decoding process, and further improves the accuracy and robustness of polyp segmentation.

[0057] Preferably, as Figure 3As shown, the SCDB unit mainly includes 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 the query Q, the key K and the value V, and the self-attention operation is performed. Since the Q and the K contain context information, they are processed into spatial information descriptors and channel information descriptors through pooling, a fully connected layer and a Sigmoid activation function, and are multiplied with the self-attention feature to obtain the feature P t . In the second branch, the input features are processed through 1x1, 3x3 and 5x5 convolutions respectively, and are spliced through a Concat function, and then are sequentially processed through a convolution, a batch normalization and a ReLu activation function to obtain the feature P c . P t is multiplied with P c , is adjusted to a proper scale through upsampling, is multiplied with the output of the SCDB module at a higher level, and the corresponding output S1-S3 is obtained.

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

[0059] S61. Obtain the feature P t :

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

[0061] The query Q is multiplied with the key K to obtain a context information feature. The context information feature is processed through one global average pooling, two fully connected layers and one Sigmoid activation function to obtain a spatial information descriptor. The context information feature is processed through a 3x3 SC convolution and a Sigmoid activation function to obtain a channel information descriptor;

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

[0063] S62. Obtain the feature P c :

[0064] The input feature of a smaller scale is processed through a 1x1 convolution to obtain a convolution feature. The convolution feature is processed through a 1x1 convolution, a 3x3 convolution and a 5x5 convolution respectively to obtain a first convolution feature, a second convolution feature and a third convolution feature;

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

[0066] S63. The feature P t is multiplied by the feature P c , and the result is upsampled. The upsampled result is multiplied by another input feature (i.e., an input feature with a larger scale) to obtain an output feature.

[0067] S7. The features F4”’, S1, S2 and S3 are input into an MSP module to obtain weighted feature maps P1-P4. A loss function is constructed according to the feature maps P1-P4, the loss is calculated by the loss function, and the model parameters are trained by back propagation until the model parameters converge. In actual segmentation tasks, the feature maps P1-P4 are combined as an original polyp segmentation map P.

[0068] Specifically, a loss function of the colorectal polyp segmentation model is constructed according to the attention prediction feature P att , the feature maps P1-P4 and a polyp segmentation map label 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 objective, and the training of the colorectal polyp segmentation model is completed. The loss function L is calculated according to the following formula

[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] Wherein, L main represents a main loss function, L binary represents a relay supervision loss function; L P1 represents a first level loss function, L P2 represents a second level loss function, L P3 represents a third level loss function, and L P4 represents a fourth level loss function. represents a weighted binary cross-entropy loss function, represents a weighted cross-entropy loss function, L dice represents a Dice coefficient loss function; L ce represents a cross-entropy loss function. G represents the polyp segmentation label corresponding to the original colon polyp image.

[0074] Preferably, an embodiment of the present application implements this model using the Pytorch framework and trains it on a Tesla V100 GPU with 32GB of video memory. Our model requires 150 epochs to complete training, and the batchsize size is set to 4 for each training. We use the Adam optimizer and set the initial learning rate to 1e-4, while using the poly learning rate as the strategy for adjusting the learning rate during training. The size of all input images is adjusted to 352x352 in the training and testing stages. The data enhancement strategy adopted is random rotation, vertical flip and horizontal flip.

[0075] In the present application, unless otherwise explicitly specified and limited, the terms "mounting", "setting", "connecting", "fixing", "rotating" and the like should be understood broadly, for example, can be fixedly connected, can also be detachably connected, or integrated; can be mechanically connected, can also be electrically connected; can be directly connected, can also be indirectly connected through an intermediate medium, can be the internal communication of two elements or the interaction relationship between two elements, unless otherwise explicitly limited, those skilled in the art can understand the specific meaning of the above-mentioned terms in the present application according to the specific circumstances.

[0076] Although the embodiments of the present application have been shown and described, those skilled in the art can understand that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for colon polyp image segmentation combining mixed attention and semantic complementarity, characterized in that, The colorectal polyp segmentation model is constructed and trained, and the colorectal polyp image to be segmented is input into the trained colorectal polyp segmentation model to obtain a polyp segmentation image; the colorectal polyp segmentation model comprises 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 comprises the following steps: S1. Obtain an original colorectal polyp image training data set comprising a plurality of groups of images; each group of images comprises an original colorectal polyp image and a corresponding polyp segmentation image label; S2. Input the original colorectal polyp image into the PVT feature encoder to obtain four features F1, F2, F3, and F4; S3. Inputting the features F2, F3, F4 into the attention-guided decoder to obtain the attention prediction feature P att ; S4. Predicting features P based on attention att and features F1, F2, F3, F4, features F1", F2", F3", F4" are obtained in combination with the residual structure; S5. Input the features F1'', F2'', F3'', and F4'' into the AHA module to obtain features A1, A2, and A3; The AHA module comprises three AHA units having the same structure, and the processing process of one AHA unit comprises: S511. merge the features F i ” and the features F i+1 ” to obtain intermediate features A i ’; wherein i = 1, 2, 3. S512. The intermediate feature A i A is obtained by 3x3 SC convolution, 5x5 convolution, and 3x3 hole convolution, respectively 3×3 , A 5×5 , A D ; S513. Feature A 5×5 with Feature A D Concatenate, then go through global average pooling, two fully connected layers, Sigmoid activation function to get the first remodeling features; S514. The first reshaped feature is respectively multiplied with the feature A 5×5 , the feature A D Dot product to obtain the first dot product feature and the second dot product feature; add the first dot product feature and the second dot product feature to obtain the feature A C ; S515. Feature A 3×3 with Feature A C Point-by-point addition fusion, and then processed by ReLu activation function, 1x1 convolution, Sigmoid activation function and reshaping operation to obtain the second reshaped feature; S516. The second reshaped feature is respectively multiplied with the feature A 3×3 , the feature A C to obtain a third point multiplication feature and a fourth point multiplication feature; and the third point multiplication feature and the fourth point multiplication feature are added and then subjected to a 1x1 convolution to obtain the feature A i ; S6. Obtain feature F4''' by performing 1x1 convolution on feature F4'', and input features F4''', A1, A2, and A3 into the SCDB module to obtain features S1, S2, and S3; The SCDB module comprises three SCDB units having the same structure, each SCDB unit receives two input features of different scales to obtain one output feature, and the specific process comprises: S61. Obtain feature P t : divide the input feature of the smaller scale into a plurality of sub-blocks to form a sub-block set, and input the sub-block set into three fully connected layers to obtain a query Q, a key K, and a value V; perform self-attention operation on the query Q, the key K, and the value V to obtain a self-attention feature; multiply the query Q and the key K to obtain a context information feature; perform 1 global average pooling, 2 fully connected layers, and 1 Sigmoid activation function on the context information feature to obtain a spatial information descriptor; and perform 3x3 SC convolution and a Sigmoid activation function on the context information feature to obtain a 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 : perform 1x1 convolution on the input feature of the smaller scale to obtain a convolution feature, and perform 1x1 convolution, 3x3 convolution, and 5x5 convolution on the convolution feature to obtain a first convolution feature, a second convolution feature, and a third convolution feature; The first convolutional feature, the second convolutional feature, and the third convolutional feature are spliced to obtain a total spliced feature, and the total spliced feature is subjected to a 3*3 SC convolution, a batch normalization layer, and a ReLu activation function to obtain a feature P c ; S63. The feature P t with the feature P c point multiplication, and the output feature is obtained by multiplying the up-sampling result with the input feature of larger scale. S7. Input features F4''', S1, S2, and S3 into the MSP module to obtain feature maps P1, P2, P3, and P4, calculate the loss by using a loss function, and train the model parameters by back propagation until the model parameters converge.

2. The method of claim 1, wherein the method is a colorectal polyp image segmentation method combining mixed attention and semantic complementation. Step S4 specifically comprises: S41. The attention prediction feature P att and the feature F1 is multiplied by the feature F1' to obtain the feature F1", and the feature F1 and the feature F1" are added. S42. The attention prediction feature P att and the feature F2 is point-multiplied to obtain a feature F2', and the feature F2 and the feature F2' are added to obtain a feature F2"; S43. The attention prediction feature P att and the feature F3 is multiplied by the feature F3' to obtain a feature F3", and the feature F3 and the feature F3" are added. S44. The attention prediction feature P att and the feature F4 is point-multiplied to obtain a feature F4', and the feature F4 and the feature F4' are added to obtain a feature F4".

3. The method of claim 1, wherein the method is a colorectal polyp image segmentation method combining mixed attention and semantic complementation. Step S5 comprises: S51. Input the feature F1'' and the feature F2'' into the first AHA unit to obtain the feature A1; S52. Input the feature F2'' and the feature F3'' into the second AHA unit to obtain the feature A2; S53. Input the feature F3'' and the feature F4'' into the third AHA unit to obtain the feature A3.

4. The method of claim 1, wherein the method is a colorectal polyp image segmentation method combining mixed attention and semantic complementation. Step S511 specifically comprises: The feature F i+1 is obtained by upsampling the feature after 3x3 SC convolution, and the fusion feature is obtained by multiplying the upsampling feature and the feature F i . perform 1x1 convolution and a Softmax layer on the fusion feature to obtain a first feature; perform two fully connected layers on the fusion feature, and connect a ReLu activation function between the two fully connected layers; The first feature is multiplied with the second feature to obtain a third feature; the third feature is added with the fusion feature to obtain an intermediate feature A i ’ 5. The method of claim 1, wherein the method further comprises: Step S6 specifically comprises: S61. Input the feature F4”’ and the feature A3 into a first SCDB unit to obtain a feature S1; S62. Input the feature S1 and the feature A2 into a second SCDB unit to obtain a feature S2; S63. Input the feature S2 and the feature A1 into a third SCDB unit to obtain a feature S3.

6. The method of claim 1, wherein the method further comprises: According to the feature maps P1-P4, a loss function L is constructed, which is represented 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) wherein, 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 cross-entropy loss function, L dice represents the Dice coefficient loss function; L ce represents the cross-entropy loss function, and G represents the polyp segmentation label corresponding to the original colon polyp image.

Images