The Research On Medical Image Segmentation Method Based On Deep Neural Networks

Medical image segmentation plays an important role in social reality scenarios such as medical diagnosis and clinical applications.It has very important research value and significance for the rational analysis and application of medical images.Compared to ordinary natural images,medical image segmentation tasks have problems such as variable organ shapes,multi-scale targets,blurred boundaries,and irregular edge protrusions,resulting in weak segmentation capabilities of general image segmentation algorithms for medical images.Therefore,solving the above problems to improve the segmentation accuracy of medical image segmentation tasks has research value and challenging significance.In response to the main difficulties of medical image segmentation,this article combines the characteristics of medical images to study a deep neural network-based medical image segmentation method,aiming to deeply explore the potential value of medical images and support the construction and development of the field of medical image analysis.The main content of this article is as follows:(1)A parallel encoding medical image segmentation method based on U-Net,P-Trans UNet,is proposed to address the problem of low segmentation accuracy in existing medical image segmentation methods.First,improve the hierarchical structure of the Res Net module,and replace the original 3×3 convolution in Res Net with multi-scale convolution to solve the problem of small objects and local boundary blur;secondly,learn the long-range Rely on information to solve the problem of over-segmentation of organs by the network;then,build a feature fusion enhancement module,fuse the output results of parallel encoders and strengthen the global characteristics;finally,gradually restore the resolution of the image by using cascaded upsampling and skip connections.This thesis demonstrates the excellent segmentation performance of P-Trans U-Net in medical image segmentation tasks through experiments.(2)Aiming at the problem that the existing methods require a high amount of data in the training data set and the network complexity and parameters are high,which makes it difficult to embed the segmentation network into various application scenarios,a medical image based on knowledge distillation is proposed.Segmentation method—Res-KD Unet.First,build a light student network for medical image segmentation,reduce the computational complexity of the segmentation model,so that it can be deployed on a wider range of devices;The problem of excessive feature loss of the student network and the teacher network caused by the gap in model size between the networks;then,by improving the attention-based fusion module ABF,a MABF and TABF module are respectively constructed to better integrate through The feature information of different scales introduced by the residual learning framework;finally,the hierarchical context loss module is used to calculate the loss between the multi-level aggregation features of the student network and the features of the teacher network,so as to more completely express the relationship between the student network and the teacher network hierarchy.Differences in feature information between them.This thesis demonstrates the excellent segmentation performance of Res-KD Unet with lower computational complexity and training data through experiments.