Medical Image Segmentation Based On Multi-level Feature Fusion

With the continuous development of medical imaging technology,we have realized early detection and early diagnosis and treatment for more and more diseases,and greatly improved the efficiency of disease treatment,through the analysis and processing of medical images.Among them,medical image segmentation can extract the ROI from complex medical images,laying a foundation for subsequent quantitative and qualitative analysis and processing.However,manual segmentation based on professional doctors is time-consuming and laborious,and traditional computer image segmentation techniques have low accuracy.Therefore,more and more scholars have focused on deep learning and proposed a series of medical image segmentation networks based on convolutional neural networks(CNN).In order to further enhance the application of deep learning in the field of medical image segmentation,we carry out research on the field of deep learning models and feature learning,and propose the fusion of shallow texture features and deep semantic features to achieve accurate segmentation of medical images.This thesis proposes a novel framework based on densely connected networks,called a multipath adaptive fusion network,to enhance the fusion of shallow texture features and deep semantic features for medical image segmentation.In the downsampling process of the proposed framework,dense blocks with short connections are used.It can not only apply dense connections between each layer to make full use of local features,but also directly connect between previous dense blocks and each layer of the current dense block,thereby achieving an effective continuous memory(CM)mechanism,To spread and retain shallow texture features.During the upsampling process,a novel multipath adaptive fusion unit has been proposed to effectively fuse shallow texture features and deep semantic features,instead of using traditional convolutional neural network(CNN)methods to discard texture features directly.By evaluating the proposed framework at BRATS 2015,it can be found that compared with most popular models,the proposed framework further improves the performance of medical image segmentation.At the same time,compared with the current mainstream deep learning segmentation network frameworks,it can obtain ideal results with fewer parameters.In addition,based on the completion of medical image segmentation,weexplored the security of medical images based on multi-level feature fusion,proposed a medical image encryption/decryption scheme,and a region of interest(ROI)segmentation in a private environment The network has further enhanced the application of multi-level feature fusion in the field of medical imaging.