Research Of Medical Image Semantic Segmentation Technology Based On Deep Learning

Semantic segmentation of medical images can segment the lesions,organs or substructures of the human body in medical images,and plays an important role in locating and identifying diseased areas and formulating medical plans.At present,clinical segmentation of medical images mainly relies on manual work,but manual segmentation is an extremely cumbersome and highly professional work,and its segmentation accuracy is easily affected by human energy,emotions,etc.In recent years,researchers have proved through practice that deep learning-based medical image semantic segmentation has shown strong potential in the field of medical image segmentation.However,due to serious noise interference in medical images and the limitation of convolution operations that easily lose contextual information,the accuracy of current deep learning-based medical image semantic segmentation methods is still limited.Therefore,this thesis studies the existing deficiencies in medical image semantic segmentation.The main work and contributions are as follows:1.In order to solve the problem of inaccurate segmentation of medical images due to the influence of noise in noisy backgrounds,a medical image semantic segmentation algorithm(USN-net)based on non-local attention skip connection module is proposed.The algorithm model uses the non-local attention skip connection module to capture the long-distance dependencies of shallow network feature maps,and generates correlation scores between pixel positions to give corresponding attention to pixel-level features,so that the skip connection splicing of feature encoding can be achieved.Pay more attention to the region of interest in medical images,reduce the influence of noise in irrelevant regions in medical images on the encoding of high-semantic features,and improve the accuracy of segmentation results.Through experiments on the kidney dynamic imaging dataset and the multi-organ abdominal CT dataset,it is verified that USN-net has better semantic segmentation performance.2.In order to solve the problem of inaccurate multi-scale image segmentation due to the lack of context semantics in medical image semantic segmentation,a medical image semantic segmentation algorithm based on local and global fusion(UTSN-net)was proposed.In the coding process of this model,the low-level features with highresolution information extracted by convolution operation are used to suppress noise through the non-local attention skip connection module,and finally the accurate highresolution positioning information is saved to the non-local attention skip connection module.In the output of the local attention skip connection module,the deep Transformer operation is used to extract the global context information and save it into the deep features.Shallow features are comprehensively considered to improve the accuracy of segmentation results.Finally,through experiments on the kidney dynamic imaging dataset and the multi-organ abdominal CT dataset,it is verified that UTSN-net has better semantic segmentation performance.