Research On Dermoscopy And Polyp Images Segmentation Method Based On Deep Learning

Skin cancer and colorectal cancer are two types of cancers with high mortality rates today,and the early clinical manifestations are in the form of skin lesions and polyps,respectively.The segmentation of skin lesions and polyps serves as the basis for key steps such as earlystage trait judgment and disease classification and provides a reliable basis for subsequent disease analysis and treatment planning.At the same time,accurate lesion segmentation results can help to improve the accuracy of the diagnosis of diseases and the judgment of benign and malignant by physicians.Therefore,whether it is from the perspective of promoting the research and development of automatic segmentation of medical images,or from the perspective of assisting clinicians in diagnosis,it is of great practical significance to research the segmentation of skin lesions and intestinal polyps.In this paper,a systematic study is carried out on the segmentation algorithm of dermoscopy and polyp images.The main work is as follows:1.Aiming at the problems of blurred boundary,uneven color distribution,and irregular shape of skin lesions in dermoscopic images,a dermoscopic image segmentation model MAUNet based on multi-scale attention fusion is proposed.MAU-Net is a segmentation network based on U-Net.By analyzing the difference between the pixels around the skin lesions and the pixels between tissues,and using the boundary features of the skin lesions,a multi-scale attention module(MA)is constructed.which fuses different levels of features during feature extraction and gives important targets a certain weight.Therefore,the importance of channel and spatial pixel features in the decoder is strengthened to improve the segmentation accuracy.Ablation experiment results show that the MA module helps to enhance the segmentation performance of the model.On the ISIC2017 dermoscopy dataset,the average intersection ratio,pixel accuracy,and Kappa value obtained by MAU-Net are 83.61%,93.58%,and 81.70%,respectively,which are 5.27%,2.01%,and 6.83% higher than U-Net,respectively.2.To overcome the complex surface condition of skin lesions and the problem of reduced segmentation accuracy caused by uneven pixel brightness,a dermoscopic image segmentation model Ms F-Seg Former based on multi-scale fusion and channel attention is proposed.By combining the edge characteristics of skin lesions at different scales,a multi-scale fusion module is designed to strengthen the extraction of skin lesion edge information.To ensure the accuracy and stability of the segmentation of skin lesions,a multi-layer encoder is used to extract the detail features and depth features of the skin lesions,and the multi-resolution feature information is effectively fused through a multi-scale fusion module.And combined with the channel attention mechanism to strengthen the features of the skin lesions,effectively improve the segmentation performance of the model.The experimental results show that the model can achieve better segmentation results on the dermoscopy segmentation task.3.Aiming at the problems of poor algorithm segmentation stability and incomplete segmentation due to the diverse external tissue structures and complex lesion surfaces in polyp images,a segmentation model BFNet based on bilateral feature fusion is proposed.The model has a double-branch structure,one with a narrower channel and a shallower structural layer,which is used to focus on the connection between adjacent pixels;The other branch introduces two modules,RFB and DFB,with wider channels and deeper structure layers to obtain highlevel semantic contextual information.Finally,the guided aggregation layer is used to efficiently combine the features of the two branches,and effectively fuse the multi-level spatial features and depth features.It can be seen from the ablation experiments that the segmentation results of different combinations of the two branches are better than the segmentation results of the single branch,indicating that the dual-branch structure can obtain more complementary information and enhance the integrity of the segmentation edges.Comparative experiments on three public medical image datasets show that the proposed algorithm can effectively process various complex medical images.Especially in the Kvasir-SEG polyp dataset for automatic polyp detection,the average Dice and average intersection ratios obtained by BFNet reached92.3% and 86.2%,respectively,which were 4.7% and 5.2% higher than U-Net(Res Net34),respectively.