Research On Medical Image Segmentation Algorithm Based On CNN And Attention Mechanism

Medical image segmentation is a complex and critical step in the field of medical image processing and analysis.The purpose is to segment medical images with special meanings and extract relevant features to provide a reliable basis for clinical diagnosis and pathology research,and assisting doctors to make a more accurate decision.In recent years,due to the application of deep learning algorithms in medical image segmentation,medical image segmentation technology has made significant progress.Based on this,this thesis has made in-depth research to improve the diagnostic accuracy of chest X-ray images.Two methods have been proposed: Multi-Pillar convolutional neural Network(MPN)and Lesion Attentive Network(LAN).Due to the special structure of the human body,the visual characteristics of different locations in the chest X-ray image vary greatly.The basic assumption of the Weight Sharing technology in traditional neural networks is that different regions have similar distributions.Therefore,this thesis proposes a Multi-Pillar convolutional neural Network to solve this problem.MPN can be seen as a structure with multiple pillars supporting the top,and multiple branch modules at the bottom,which independently process input data from different locations and regions,and then fuse the feature maps of each branch module to form fusion features.The resulting single-stream network further encodes global contextual information from the fused features.The AUC score obtained by this method on the NIH chest X-ray data set is better than other algorithms in this thesis,which verifies the effectiveness of the proposed method.Considering that the target lesion information accounts for a small proportion of the entire image,in order to improve segmentation efficiency and diagnostic accuracy,Lesion Attentive Network is proposed to improve the depth model: the Hard Attention and Soft Attention imitate how the radiologist analyzes the image and pays attention to the most disease-relevant areas in the input image,which improves the efficiency of the model in extracting features;and the addition of Large Decision Margin Loss(LDM Loss)can in turn encourage the model to mine more accurate and persuasive prediction capabilities,thereby improving its generalization performance.Extensive experiments were performed on the NIH chest X-ray data set.The results show that the proposed method is superior to other existing algorithms compared in this thesis on the verification set and test set,confirming the effectiveness of the proposed network model and improved algorithm.