Application Of Multi-atlas Medical Image Segmentation

With the development of science and technology,human beings have divided the brain structure into multiple functional areas,each of which governs different functions of human beings.Therefore,in disease research,it is often necessary to extract corresponding regions of interest(ROI)for diagnosis and prediction through changes in their characteristics such as volume.For the segmentation of the ROI functional area,although manual segmentation can obtain higher segmentation accuracy,it is time-consuming and labor-intensive and has poor repeatability.Therefore,the use of computer technology for fully automatic segmentation technology has become a research hotspot.Multi-map image segmentation technology has achieved good results in the field of pattern recognition and medical image segmentation.This article will examine how to get segmentation results faster and more accurately.First of all,this paper will introduce the current multi-map based segmentation algorithm and its advantages and disadvantages,such as Majority Voting(MV),Global Weighted Voting(GWV),and Local Weighted Voting.(Local Weighted Voting,LWV),Non-Local Mean(NLM)and Sparse Patch-based Method(SPBM).Secondly,GPU parallel acceleration is carried out for NLM and SPBM,two algorithms with better performance in the current segmentation algorithm based on multi-graph,so as to obtain segmentation results faster and meet clinical requirements more effectively.Finally,there are two key problems in the current multi-graph label fusion framework:(1)the significant difference between the features in the image domain and the binary domain of anatomical labels is ignored.The optimized weight for minimizing the difference of different grayscale does not necessarily mean that it is the best choice for tag fusion;(2)Lack of knowledge of the block and evaluation of the map block..The current tagging approach only exploits the relationship between the target block and the map block,ignoring that unrelated map blocks may dominate the voting process.To solve these problems,we propose a novel multi-graph tag fusion framework to simultaneously describe various image blocks(including target block and map block)and estimate the final optimized intrinsic tag fusion weight to minimize the risk of false tags.We have evaluated the tag fusion method we proposed and segmented the deep gray matter and ROl of the hippocampus,basal ganglia region and the whole brain.Compared with other advanced multi-graph tag fusion methods,our method has achieved significant results and improved the segmentation accuracy.