Research On Medical Image Segmentation Method Based On Deep Neural Network And Attention Mechanism

Medical image segmentation is a crucial technology in medical image analysis,which can separate the target tissues in complex medical images,providing strong support for medical diagnosis and treatment.Medical image segmentation is the first step in computer-aided detection,which helps doctors diagnose and treat medical conditions.However,due to medical images’ special and diverse properties,medical image segmentation still faces many challenges and difficulties.Deep neural networkbased medical image segmentation algorithms have emerged to solve these problems.By utilizing the powerful learning ability of deep learning models,they can automatically learn complex feature representations and improve the accuracy and efficiency of medical image segmentation.Therefore,this article mainly explores highperformance medical image segmentation algorithms based on deep networks,aiming to improve the accuracy and efficiency of medical image segmentation,help doctors accurately locate lesion areas,and provide better medical diagnosis and treatment assistance.In order to achieve this goal,this article first provides a detailed literature review and analysis of deep neural network-based medical image segmentation methods in recent years,comparing their similarities and differences and introducing their fundamental principles and contributions.In addition,this article also explores the difficulties and characteristics faced by medical image segmentation and how to use the learning ability of deep neural networks to solve these problems.Based on the above analysis,this article further proposes two different convolutional neural networkbased medical image segmentation algorithms,including:(1)This article proposes a triple-attention convolutional neural network-based medical image segmentation algorithm(TANet)to address the challenges of scale variation and difficulty segmenting small targets in medical images.The algorithm uses three mechanisms,scale attention,position attention,and channel spatial attention,to weigh the image features from different perspectives and improve segmentation accuracy.Firstly,TANet dynamically selects the most suitable scale features for the segmentation task from multi-scale features using the scale attention mechanism,which helps the model handle objects of different scales.Secondly,TANet uses position attention to enhance the correlation of similar features in the global view and channel attention to capture the correlation between channels.This mechanism enhances the representation of global contextual features from different dimensions,helping the model to handle complex scenes and segmentation tasks and improve segmentation accuracy,for instance,addressing the difficulty of segmenting small targets.(2)To improve the generalization and precise prediction ability of medical image segmentation models,this paper proposes a Transformer-based and CNN medical image segmentation algorithm(SST-Net)that combines Transformer with CNN.The algorithm uses Transformer as the encoder because Transformer is a powerful sequence modeling method that can capture global information and help the model better understand the image.However,Transformer may have some shortcomings in processing local details.Therefore,SST-Net also uses a CNN-based local feature processing module to provide missing local detail information.In addition,SST-Net uses a progressive feature fusion approach to reduce the difference between features at different depths.Two algorithms were experimented with and analyzed on the polyp and skin lesion datasets;the results showed that they performed well in most evaluation metrics and surpassed other current medical image segmentation algorithms.