Intestinal polyp segmentation method based on multilayer information fusion and optimization network

Abstract

The invention provides an intestinal polyp segmentation method based on a multilayer information fusion and optimization network, and the method comprises the following steps: carrying out the preprocessing of a data set, and extracting a low-level feature map and a high-level feature map through an encoder; extracting overall structure information from the low-level feature map; 4, extracting global feature information and local feature information from the advanced feature map, and obtaining enhanced features of adjacent layers; the fusion of the enhanced features of adjacent layers is realized by using three hierarchical feature integration HFIM modules; step 6, obtaining a coarse segmentation image for preliminarily segmenting the original image; optimizing the coarse segmentation image; a final polyp segmentation result is obtained; and training the network model.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image segmentation, and in particular relates to an intestinal polyp segmentation method based on multi-layer information fusion and optimization network. Background Art

[0002] Colorectal Cancer (CRC) is the third most common cancer in the world, causing great harm to human life and health. Adenomatous polyps have a high risk of malignant transformation and may develop into colon cancer. Colonoscopy can detect and remove these polyps in time, thereby effectively reducing the risk of death related to colon cancer. In clinical colonoscopy, it depends to a large extent on the professional ability of doctors. It is often a huge physical and psychological test for doctors to identify intestinal polyps while performing colonoscopy, which can easily lead to misdetection and missed detection of polyps. According to research statistics, the rate of missed polyps by doctors during colonoscopy is as high as 26% (S. Zhao, S. Wang, P. Pan, T. Xia, X. Chang, X. Yang, L. Guo, Q. Meng, F. Yang, W. Qian, et al., "Magnitude, risk factors, and factors associated with adenoma miss rate of tandem colonoscopy: a systematic review and meta-analysis," Gastroenterology, vol. 156, no. 6, pp. 1661–1674, 2019.). Therefore, it is very necessary to study the automatic segmentation method of polyps to assist doctors in identifying and diagnosing polyps, which can further reduce the workload of doctors, improve the efficiency of diagnosis, and promote the progress of medicine in the field of colon cancer.

[0003] Compared with the traditional manual segmentation method based on morphological information such as the outline, texture and shape of polyps, deep learning methods can automatically complete feature extraction, integrate multi-scale information and show strong learning and generalization capabilities, which have great advantages that cannot be replaced by traditional manual segmentation methods. As the most typical deep learning method in the field of polyp segmentation, UNet (Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: MICCAI. pp. 234–241. Springer (2015)) first introduced the U-shaped structure into medical image segmentation, and effectively solved the problem of insufficient correlation between high-level and low-level features by gradually combining features of different levels. Its innovative design has been widely used in subsequent studies and has become an important architecture in the field of medical image segmentation. SANet (Wei J, Hu Y, Zhang R, et al. Shallow attention network for polyp segmentation [C] Medical Image Computing and Computer Assisted Intervention–MICCAI 2021) introduces a color swapping strategy to reduce model overfitting caused by different color samples, allowing the model to pay more attention to the shape of polyps. IC-PolypSeg (Chen Z, Wang K, Liu Y. Efficient polyp segmentation via integrity learning [C] / / ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024: 1826-1830.) uses an attention mechanism to dynamically fuse high-level semantic features with underlying spatial information to generate a comprehensive cross-stage feature representation.DUAT (Tang F, Xu Z, Huang Q, et al. DuAT: Dual-aggregation transformer network for medical imagesegmentation [C] / / Chinese Conference on Pattern Recognition and Computer Vision (PRCV). Singapore: Springer Nature Singapore, 2023: 343-356.) uses an aggregation module to fuse local spatial features with semantic information from high-level features, thereby improving the model's ability to capture details and accurately identifying the boundaries and morphology of polyps. ColonFormer (Duc NT, Oanh NT, Thuy NT, et al. Colonformer: An efficient transformer based method for colon polyp segmentation [J]. IEEE Access, 2022, 10: 80575-80586.) implements multi-level feature fusion in the decoder, enhances the richness of feature representation, and introduces skip connections in the subsequent refinement module to further optimize the accuracy of polyp boundaries.

[0004] Although the existing deep learning models have significantly improved the effect of processing polyps, they still have limitations. (1) The CNN-based feature extraction network has difficulty understanding global context information, which makes it difficult to adapt to the complex background and multi-scale changes of polyp images. (2) Although the Transformer-based feature extraction network enhances the processing of global information, it is easy to lose information during the transmission of context information. There is an urgent need to study new algorithms to make up for the shortcomings of existing methods. Summary of the invention

[0005] The purpose of the present invention is to provide an intestinal polyp segmentation method based on multi-layer information fusion and optimization network, optimize the segmentation performance of polyp segmentation, and provide more efficient auxiliary support for clinical diagnosis.

[0006] To achieve the above-mentioned purpose, the present invention provides the following technical solutions:

[0007] A method for segmenting intestinal polyps based on multi-layer information fusion and optimization network includes the following steps:

[0008] Step 1: Get the data set and divide it into training set and test set;

[0009] Step 2: Preprocess the data set and extract the feature map through the encoder. The extracted feature map is M i , where M i , i=2, 3, 4 are high-level feature maps, M1 is a low-level feature map;

[0010] Step 3: Extract overall structural information C1 from the low-level feature map;

[0011] Step 4: Extract global and local feature information from high-level feature maps to obtain enhanced features of adjacent layers. The method is as follows:

[0012] For the high-level features M, i=2, 3, 4 extracted by the encoder, the global features and local feature information are further extracted through the contextual fine feature processing CFFP module respectively;

[0013] The contextual fine feature processing CFFP module includes a global feature extraction module and a local feature extraction module;

[0014] Global feature extraction module, used to extract global feature information C gi :Advanced Features M i , i = 2, 3, 4, the features are initially processed through 1×1 convolution, and then the global attention to high-level features is strengthened through the graph convolution network. Subsequently, the influence of irrelevant features is suppressed through channel attention and convolution kernel, and the global feature information C is output. gi ;

[0015] Local feature extraction module, used to extract local information C li :Advanced Features M i , i = 2, 3, 4, through three series of 1 × 1 and 3 × 3 convolution kernels to further extract deep information, through the sigmoid activation function, and with the high-level feature M i , i = 2, 3, 4 are multiplied and added to output local feature information C li ;

[0016] The enhanced feature C of the adjacent layer is obtained by adding the global feature information and the local feature information. i ;

[0017] Step 5: Use the three-level feature integration HFIM module to realize the enhanced features C of adjacent layers i The method is:

[0018] The hierarchical feature integration HFIM module enhances the adjacent layer feature C i The add and concat operations are used at the same time, and the feature representation capability is further enhanced through 1×1 and 3×3 convolution kernel optimization, corresponding to the output H i , i=1,2,3;

[0019] Step 6: Convert the output of step 5 to H i , i = 1, 2, 3 are added together and fed to the multi-scale context attention MSCA module together with the features after the convolution operation of M1 to eliminate the noise information introduced in the high-level feature fusion process and output the rough segmentation map O4 for the initial segmentation of the original image;

[0020] Step 7: Optimize the coarse segmentation map O4 and use the adaptive channel pyramid ACAP to extract the high-level features M from the encoder i , i = 2, 3, 4 for feature representation, and output A representing the high-level information in the deep features of the encoder i , i=2,3,4;

[0021] Step 8: Use three skip reverse attention SRA modules to represent the high-level information in the deep features of the encoder A i , i=2,3,4 further improve the edge details and obtain the coarse segmentation map O i , i = 1, 2, 3, 4, the rough segmentation image O1 is upsampled 4 times to obtain the final polyp segmentation result;

[0022] Step 9: Train the network model.

[0023] Furthermore, the method of step 2 is as follows: preprocess the images of the data set including resizing, multi-scale magnification and reduction, input the preprocessed images into an encoder based on SegFormer-b2, and extract feature maps.

[0024] Furthermore, in step 2, the encoder is an encoder based on SegFormer-b2.

[0025] Furthermore, the method of step three is as follows: for the low-level feature M1 extracted by the encoder, the overall structural information C1 of the polyp is extracted through two serially connected 3×3 convolutions, batch normalization and Relu activation functions.

[0026] Furthermore, the expression of the local feature information output by the local feature extraction module in step 4 is as follows:

[0027]

[0028] Among them, CBR3 represents 3×3 convolution, BN and ReLu, and CBR1 represents 1×1 convolution, BN and ReLu. M stands for multiplication; n Is to input L k After treatment with CBR3 and CBR1, k Add together the results of the above combination steps three times, where L in the first combination step is kRepresents high-level features M i , i = 2, 3, 4, L in the last two combination steps k Represents the output of the previous combination step, and its expression is as follows:

[0029] M n =(CBR3(CBR1(L k ))+L k ) ×3

[0030] Furthermore, in step 5, the output H i , i=1,2,3, the expression is as follows:

[0031] H i =C addition +C concat

[0032] C addition =Conv1(Conv3(X i +C i-1 ))

[0033] C concat =Conv1(Concat(X i , C i-1 ))

[0034] Conv1 represents 1×1 convolution, Conv3 represents 3×3 convolution, and X i Indicates the subsequent CFFP module or HFIM

[0035] Output of the module.

[0036] Further, in step six, the multi-scale contextual attention MSCA module performs the following steps:

[0037] The average pooling layers with scales of 1, 3, and 5 are used as K and V in the self-attention mechanism. The average pooling layer helps the model capture a wider range of global features and enhances the multi-scale feature perception ability of the self-attention mechanism. As Q in the self-attention mechanism, K is multiplied with the reshaped C1 and Q respectively and passed through the softmax layer, and then multiplied with V to complete the self-attention operation; the output of this module is a rough segmentation map O4 for the initial segmentation of the original image.

[0038] Furthermore, in step 7, the high-level features M extracted by the encoder are processed using the adaptive channel pyramid ACAP. i , i=2,3,4 The method of feature representation is: M i, i = 2, 3, 4. Four series of dilated convolution kernels with dilation rates of 1, 3, 5, and 9 are used to extract multi-scale detail information of the original feature map from the four branches respectively, and the original number of channels is restored after concat. Channel attention is used for further processing of each branch to adapt the representation ability of different branches to polyps of different scales; the results of adjacent branches are added and merged in turn, and Ai representing the high-level information in the deep features of the encoder is output. ， i=2,3,4.

[0039] Furthermore, the method of step eight is:

[0040] Three skip reverse attention SRA modules are used to focus on A in step 6, which represents the high-level information in the deep features of the encoder. i , i=2,3,4 further improve the edge details, and the output result is S i , i = 2, 3, 4, the specific formula of the jump reverse attention SRA module is as follows:

[0041] S i =O i ×R i +O i

[0042] Where R i is the output of reverse attention, expressed as:

[0043] R i =1-sigmoid(A i )

[0044] The output S4 of the SRA module of this layer is added to the previously obtained coarse segmentation map O4 to obtain the coarse feature map O3 output by this layer; the coarse segmentation maps O1 and O2 are obtained in the same way;

[0045] The rough segmentation image O1 is upsampled 4 times to obtain the final polyp segmentation result.

[0046] Furthermore, the method of step nine is: use the Adam algorithm to guide the gradient descent process. The intersection-over-union loss function IoU Loss and the focal loss function Focal Loss are selected as the total loss function of the model, and the coarse segmentation map O of each layer is used. i , i = 1, 2, 3, 4 calculate the total loss of the model and perform back propagation to guide the model to update parameters.

[0047] The present invention has the following beneficial effects:

[0048] (1) The present invention proposes a new segmentation algorithm for intestinal polyps. This algorithm creatively uses a pre-segmentation network to compensate for the information loss at different layers when context information is fused in existing methods, and uses an optimized network to accurately extract the edge contour of the polyp, further improving the segmentation performance.

[0049] (2) The algorithm proposed in this invention is trained and tested on real public data sets. The test results on some test sets are as follows: Figure 7 As shown in Figure 2, it can be seen that the algorithm we proposed performs very well in terms of segmentation accuracy. Compared with the visual segmentation diagram of other models, Figure 6 As shown, the algorithm proposed in the present invention is closer to the real segmented image than other algorithms, which proves the effectiveness and excellence of the present invention in segmenting polyps. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a structural diagram of the intestinal polyp segmentation method based on multi-layer information fusion and optimization network proposed by the present invention;

[0051] Figure 2 This is the implementation schematic diagram of the CFFP module;

[0052] Figure 3 This is the implementation schematic diagram of the HFIM module;

[0053] Figure 4 This is the internal structure diagram of the MSCA module;

[0054] Figure 5 This is the implementation structure diagram of the ACAP module;

[0055] Figure 6 It is a comparison diagram of the segmentation results of the present invention and other networks;

[0056] Figure 7 This is the test result of the present invention on a part of the test set. DETAILED DESCRIPTION

[0057] The present invention provides a method for segmenting intestinal polyps based on multi-layer information fusion and optimization network. The overall framework diagram of the system can be found in Figure 1 , we will refer to Figures 1 to 5 , a detailed explanation of the entire method:

[0058] Step 1: Get dataset images to prepare for training and testing the network. We take 900 images from the Kvasir-SEG dataset and 550 images from the ClinicDB dataset as training sets, and the remaining images from the Kvasir_SEG and ClinicDB datasets and all images from the CVC-ColonDB, ETIS, and CVC-300 datasets for testing. These images are sample data obtained from real colonoscopy operations, with high clinical authenticity, and can effectively simulate actual application scenarios.

[0059] Step 2: Resize the images in the dataset e The image is resized to 352×352 pixels and is enlarged and reduced using a multi-scale factor of {0.75, 1, 1.25}. It is then fed into the encoder based on SegFormer-b2 and the output feature map is represented as M i . Where M i , i=2, 3, 4 are high-level feature maps, and M1 is a low-level feature map.

[0060] Step 3: For the low-level feature M1 extracted by the encoder, after passing it through two series of 3×3 convolutions, batch normalization and Relu activation function, the overall structural information of the polyp can be well extracted. After this processing, the output of M1 is C1. i , i = 2, 3, 4, and further extract global and local information through the contextual and fine feature processing (CFFP) module. Figure 2 The upper part shows the global feature extraction module, which initially processes features through 1×1 convolution, and then strengthens the global focus on high-level features through the graph convolution network. Subsequently, channel attention and convolution kernels are used to suppress the influence of irrelevant features. The final output is C gi , which is expressed as follows:

[0061] C gi =CBR3(CA(CBR1(GConv(CBR1(M i )))))

[0062] Among them, CBR3 represents 3×3 convolution, BN and ReLu, CBR1 represents 1×1 convolution, BN and ReLu, CA represents channel attention mechanism, and GConv represents graph convolution module. Figure 2 The lower part shows the local feature extraction module used to extract local information C li , Advanced Information M iThe deep information is often contained in the image. This deep information can be further extracted through three series of 1×1 and 3×3 convolution kernels, and the sigmoid activation function is used to combine it with the high-level feature M. i , i = 2, 3, 4 are multiplied and added, and more attention is paid to its effective information. The expression is as follows:

[0063]

[0064] Among them, CBR3 represents 3×3 convolution, BN and ReLu, and CBR1 represents 1×1 convolution, BN and ReLu. Represents multiplication. n Is to input L k After treatment with CBR3 and CBR1, k Add together the results of the above combination steps three times, where L in the first combination step is k Represents high-level features M i , i = 2, 3, 4, L in the last two combination steps k Represents the output of the previous combination step. Its expression is as follows:

[0065] M n =(CBR3(CBR1(L k ))+L k ) ×3

[0066] The enhanced feature C is obtained by adding the global and local information i , which is expressed as follows:

[0067] C i =C gi +C li

[0068] Step 4: Enhance the feature C of the adjacent layers in step 3 i In order to integrate the features and avoid information loss as much as possible, the present invention uses three hierarchical feature integration modules (Hierarchical Feature Integration Module, HFIM) to complete the above functions. Its specific structure is as follows Figure 3 As shown, for the feature C of the adjacent layer i The add and concat operations are used at the same time, and the feature representation capability is further enhanced through 1×1 and 3×3 convolution kernel optimization. The corresponding output H i , i=1,2,3, the expression is as follows:

[0069] H i =C addition +C concat

[0070] C addition =Conv1(Conv3(X i +C i-1 ))

[0071] C concat =Conv1(Concat(X i , C i-1 ))

[0072] Conv1 represents 1×1 convolution, Conv3 represents 3×3 convolution, and X i Indicates the output of the subsequent CFFP module or HFIM module.

[0073] Step 5: Convert the output of step 4 to H i , i = 1, 2, 3 are added together and fed to the Multi-Scale Contextual Attention (MSCA) module together with the features after the convolution operation of M1 to eliminate the noise information introduced in the process of high-level feature fusion. Figure 4 As shown, The average pooling layers with scales of 1, 3, and 5 are used as K and V in the self-attention mechanism. The use of average pooling layers helps the model capture a wider range of global features and helps enhance the multi-scale feature perception ability of the self-attention mechanism. As Q and K in the self-attention mechanism, they are multiplied with the reshaped C1 and Q respectively and passed through the softmax layer, and then multiplied with V to complete the self-attention operation. This module can well integrate low-level information and aggregated high-level information, reducing the interference of noise in the original image. The output of this module is a rough segmentation map O4 for the initial segmentation of the original image. The specific expression of O4 is as follows:

[0074] O4=R(softmax(QK+R(C1)K))V+R(H)

[0075]

[0076] R stands for reshape. The above steps 3 to 5 show the process of the pre-segmentation network proposed in this patent. Although the overall shape of the polyp can be observed from the output result of the pre-segmentation network, that is, the output O4 of step 5, the segmentation map does not handle the edge details well, and the edge optimization still needs to be completed through subsequent steps.

[0077] Step 6: To optimize the rough segmentation map O4 obtained in step 5, first use the adaptive channel attention pyramid (ACAP) to represent the high-level features M, i = 2, 3, 4 extracted by the encoder. Specifically, M i , i=2,3,4. Four series of dilated convolution kernels with dilation rates of 1, 3, 5, and 9 are used to extract multi-scale detail information of the original feature map. The original number of channels is restored after concat. Channel attention is used for further processing of each branch to adapt the representation ability of different branches to polyps of different scales. The results of adjacent branches are added and merged in turn to fully represent the high-level information in the deep features of the encoder. The final output of this module is A i , i=2,3,4.

[0078] Step 7: Use three Skip-Reverse Attention (SRA) modules to represent the information A after the feature representation in step 6 i , i=2,3,4 further improve the edge details, and the output result is S i , i = 2, 3, 4, its specific structure is as follows Figure 1 As shown on the left. The specific formula of jump reverse attention is as follows:

[0079] S i =O i ×R i +O i

[0080] Where R i is the output of reverse attention and can be expressed as:

[0081] R i =1-sigmoid(A i )

[0082] Among them A i It is the output of the ACAP module. Through the processing of the SRA module, it can well combine the features of different layers and guide the improvement of edge details.

[0083] To obtain the rough segmentation graph O of each layer i , i = 1, 2, 3, taking O3 as an example, the output S4 of the SRA module of this layer is added to the previously obtained coarse segmentation map O4 to obtain the coarse feature map O3 output by this layer. The coarse segmentation maps O1 and O2 can be obtained by following the same steps.

[0084] Steps 6 and 7 show the process of optimizing the network in the present invention. The final polyp segmentation result is obtained by upsampling O1 by 4 times.

[0085] Step 8: Train the network model. According to the division of the training set and the test set mentioned in step 1 and the processing of steps 2 to 7, use the Adam algorithm to guide the gradient descent process. The intersection over union loss function (IoU Loss) and the focal loss function (Focal Loss) are selected as the total loss function of the model to guide the training process of the model. The formula is as follows:

[0086] L(O, G)=0.5L iou (O, G)+0.5L focal (O, G)

[0087] Where O represents the coarse segmentation map, G represents the segmentation mask Ground Truth, L iou represents the intersection-over-union loss function, L focal represents the focal loss function.

[0088] We use the coarse segmentation map O of each layer i , i = 1, 2, 3, 4 calculate the total loss of the model, and perform back propagation to guide the model to update the parameters. The formula is as follows:

[0089]

[0090] Among them, L(O i , G) represents the coarse segmentation map O output in step six and step seven i , the loss between i=1,2,3,4 and the segmentation mask G.

[0091] Aiming at the problems faced by doctors in polyp segmentation during actual colonoscopy, the present invention proposes a method for intestinal polyp segmentation based on multi-layer information fusion and optimization network. The use of a pre-segmentation network makes up for the information loss problem between different layers and integrates more comprehensive context information. In addition, the present invention also uses an optimization network to finely extract the edge contour of the polyp, thereby improving the accuracy of the segmentation result. The present invention effectively solves the problems of information loss and unclear edges in existing methods and has strong potential for clinical application.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for segmenting intestinal polyps based on multi-layer information fusion and optimization network, comprising the following steps: Step 1: Get the data set and divide it into training set and test set; Step 2: Preprocess the data set and extract the feature map through the encoder. The extracted feature map is M i , where M i , i=2, 3, 4 are high-level feature maps, M1 is a low-level feature map; Step 3: Extract overall structural information C1 from the low-level feature map; Step 4: Extract global and local feature information from high-level feature maps to obtain enhanced features of adjacent layers. The method is as follows: High-level features M extracted by the encoder i , i = 2, 3, 4, further extract global and local feature information through the contextual fine feature processing CFFP module; The contextual fine feature processing CFFP module includes a global feature extraction module and a local feature extraction module; Global feature extraction module, used to extract global feature information C gi :Advanced Features M i , i = 2, 3, 4, the features are initially processed through 1×1 convolution, and then the global attention to high-level features is strengthened through the graph convolution network. Subsequently, the influence of irrelevant features is suppressed through channel attention and convolution kernel, and the global feature information C is output. gi ; Local feature extraction module, used to extract local information C li :Advanced Features M i , i = 2, 3, 4, through three series of 1 × 1 and 3 × 3 convolution kernels to further extract deep information, through the sigmoid activation function, and with the high-level feature M i , i = 2, 3, 4 are multiplied and added to output local feature information C li ; The enhanced feature C of the adjacent layer is obtained by adding the global feature information and the local feature information. i ; Step 5: Use the three-level feature integration HFIM module to realize the enhanced features C of adjacent layers i The method is: The hierarchical feature integration HFIM module enhances the adjacent layer feature C i The add and concat operations are used at the same time, and the feature representation capability is further enhanced through 1×1 and 3×3 convolution kernel optimization, corresponding to the output H i , i=1,2,3; Step 6: Convert the output of step 5 to H i , i = 1, 2, 3 are added together and fed to the multi-scale context attention MSCA module together with the features after the convolution operation of M1 to eliminate the noise information introduced in the high-level feature fusion process and output the rough segmentation map O4 for the initial segmentation of the original image; Step 7: Optimize the coarse segmentation map O4 and use the adaptive channel pyramid ACAP to extract the high-level features M from the encoder i , i = 2, 3, 4 for feature representation, and output A representing the high-level information in the deep features of the encoder i , i=2,3,4; Step 8: Use three skip reverse attention SRA modules to represent the high-level information in the deep features of the encoder A i , i=2,3,4 further improve the edge details and obtain the coarse segmentation map O i , i = 1, 2, 3, 4, the rough segmentation image O1 is upsampled 4 times to obtain the final polyp segmentation result; Step 9: Train the network model.

2. The intestinal polyp segmentation method according to claim 1, characterized in that: The method of step 2 is as follows: preprocess the images in the dataset including resizing, multi-scale magnification and reduction, input the preprocessed images into the encoder based on SegFormer-b2, and extract the feature map.

3. The intestinal polyp segmentation method according to claim 1, characterized in that: In step 2, the encoder is an encoder based on SegFormer-b2.

4. The intestinal polyp segmentation method according to claim 1, characterized in that: The method of step three is as follows: for the low-level feature M1 extracted by the encoder, the overall structural information C1 of the polyp is extracted through two serially connected 3×3 convolutions, batch normalization and Relu activation functions.

5. Step 2 of the intestinal polyp segmentation method according to claim 1 is characterized in that: The expression of the local feature information output by the local feature extraction module in step 4 is as follows: Among them, CBR3 represents 3×3 convolution, BN and ReLu, CBR1 represents 1×1 convolution, BN and ReLu, M stands for multiplication; n Is to input L k After treatment with CBR3 and CBR1, k Add together the results of the above combination steps three times, where L in the first combination step is k Represents high-level features M i , i = 2, 3, 4, L in the last two combination steps k Represents the output of the previous combination step, and its expression is as follows: M n =(CBR3(CBR1(L k ))+L k ) ×3 6. The intestinal polyp segmentation method according to claim 1, characterized in that: In step 5, the output H i , i=1,2,3, the expression is as follows: H i =C addition +C concat C addition =Conv1(Conv3(X i +C i-1 )) C concat =Conv1(Concat(X i ,C i-1 )) Conv1 represents 1×1 convolution, Conv3 represents 3×3 convolution, and X i Indicates the output of the subsequent CFFP module or HFIM module.

7. The intestinal polyp segmentation method according to claim 1, characterized in that: In step 6, the multi-scale contextual attention MSCA module performs the following steps: The average pooling layers with scales of 1, 3, and 5 are used as K and V in the self-attention mechanism. The average pooling layer helps the model capture a wider range of global features and is used to enhance the multi-scale feature perception ability of the self-attention mechanism. As Q in the self-attention mechanism, K is multiplied with the reshaped C1 and Q respectively and passed through the softmax layer, and then multiplied with V to complete the self-attention operation; the output of this module is a rough segmentation map O4 for the initial segmentation of the original image.

8. The intestinal polyp segmentation method according to claim 1, characterized in that: In step 7, the high-level features M extracted by the encoder are processed using the adaptive channel pyramid ACAP. i , i=2,3,4 The method of feature representation is: M i , i = 2, 3, 4. Four series of dilated convolution kernels with dilation rates of 1, 3, 5, and 9 are used to extract multi-scale detail information of the original feature map from the four branches respectively, and the original number of channels is restored after concat. Channel attention is used for further processing of each branch to adapt the representation ability of different branches to polyps of different scales; the results of adjacent branches are added and merged in turn, and A representing the high-level information in the deep features of the encoder is output. i , i=2,3,4.

9. The intestinal polyp segmentation method according to claim 1, characterized in that: The method of step eight is: Three skip reverse attention SRA modules are used to focus on A in step 6, which represents the high-level information in the deep features of the encoder. i , i=2,3,4 further improve the edge details, and the output result is S i , i = 2, 3, 4, the specific formula of the jump reverse attention SRA module is as follows: S i =The i ×R i +O i Where R i is the output of reverse attention, expressed as: R i =1-sigmoid(A i ) The output S4 of the SRA module of this layer is added to the previously obtained coarse segmentation map O4 to obtain the coarse feature map O3 output by this layer; the coarse segmentation maps O1 and O2 are obtained in the same way; The rough segmentation image O1 is upsampled 4 times to obtain the final polyp segmentation result.

10. The intestinal polyp segmentation method according to claim 1, characterized in that: The method of step nine is: use the Adam algorithm to guide the gradient descent process. Select the intersection-over-union loss function IoU Loss and the focal loss function FocalLoss as the total loss function of the model, and use the coarse segmentation map O of each layer i , i = 1, 2, 3, 4 calculate the total loss of the model and perform back propagation to guide the model to update parameters.

Images