Study On Medical Image Segmentation Algorithm Based On Attention

Medical image segmentation can assist doctors in medical diagnosis.In practical applications,medical imaging devices vary,the size and shape of targets are irregular,and various artifacts exist in images,which still pose some challenges to medical image segmentation.U-Net is currently one of the most widely used networks in the field of medical image segmentation.However,convolutional operations have local characteristics when extracting features,which can affect the performance of medical image segmentation.This article addresses these issues from two perspectives: increasing receptive field and integrating multi-scale features.Compared to convolutional neural networks,Transformers have a self-attention mechanism that can effectively obtain global information.Therefore,this article applies Transformers to the field of medical image segmentation to compensate for the shortcomings of convolutional neural networks.Since the common difficulty in skin lesion segmentation and optic disc and cup segmentation is unclear boundaries and varying target sizes,this article conducts experiments around these two segmentation tasks,adding additional attention modules to both convolutional neural networks and Transformers,further verifying the effectiveness of attention mechanisms in medical image segmentation tasks and making the following contributions:(1)For the problems of varying sizes,different positions,fuzzy boundaries,and irregularities of skin lesion regions,this article proposes a skin lesion segmentation method based on attention and dual-encoding convolutional neural networks.This method first integrates and shares feature information extracted by two encoders through a dual-encoding structure,thereby fully utilizing multi-scale features.Then,it uses a complementary attention module(CAM)to increase the receptive field and capture contextual information.CAM can highlight the target area while suppressing the background area,adapting to skin lesion image segmentation tasks with different sizes,positions,and irregular edges.The effectiveness of this method was verified through experiments on the ISIC2017 and ISIC2018 datasets.(2)For the problems of the similarity between optic disc and cup regions and background regions,fuzzy edges,and high noise in optic disc and cup segmentation tasks,this article proposes a dual-branch and Transformer-based optic disc and cup segmentation method,designing three modules to improve the representation ability of features.Among them,the Scale-Aware-Feature Fusion Module(SCA-FFM)is used to collect semantic and positional information of optic disc and cup from high-level features;the Identification Module(IM)is used to capture hidden optic disc and cup information in low-level features;and the Graph Convolution Domain-Feature Fusion Module(GCD-FFM)integrates high-level semantic features and low-level features.The effectiveness of this algorithm was verified through experiments on the Drishti-GS1,RIM-ONE-r3,and REFUGE datasets.