Research On Label Fusion Algorithm Based On Multi-atlas Segmentation

There are many complex structures with different functions in human brain,such as hippocampus,amygdala,superior temporal gyrus,cerebellum,brainstem,caudate nucleus,and other key brain structures which are closely related to a variety of brain diseases,and their precise segmentation is a prerequisite for quantitative analysis of physicians in clinical diagnosis,so the multi-atlas segmentation technology has become the focus of current research at home and abroad.Multi-atlas segmentation technology mainly includes two key steps,image registration and label fusion.This method need register multiple atlases with the target image and select the appropriate fusion algorithm to fuse the registered atlases to get the final segmentation result.In order to make the result of segmentation more accurate,it is necessary to select the appropriate label fusion algorithm to let the registered atlases achieve high precision in the fusion process,and it can effectively extract the information in each initial segmentation and get the most representative results as the final segmentation results.The widely used Label fusion algorithm are Majority Voting,STAPLE,COLLATE,and so on.MV does not take into account the differences between the different segmented images,STAPLE algorithm does not use the priori information of the images.In order to obtain higher segmentation accuracy,this paper first preprocess the brain MR images,including skull-strip,filtering,grayscale normalization and histogram matching,and then this paper researches and improves the label fusion algorithm based on multi-atlas registration,which mainly includes the following points:(1)Around the human brain MR images,we research and analysis the widely used MV fusion algorithm and STAPLE algorithm,and use this two methods to fuse multiple tissues of the brain images that have been registered.We use the similarity measure of fusion result and golden standard as the evaluation criterion to compare the fusion result of the two methods with the fusion result of the optimal single atlas method.(2)We proposed a new improved Weight-Voting fusion algorithm based on the MV algorithm,which uses the similarity measure between atlases and the target image as the weight of the image fusion,we fuse multiple tissues of the brain images that have been registered,and compare the algorithm proposed in this paper with the optimal single atlas,MV and STAPLE fusion algorithms respectively.The experimental results show that the Multi-atlas based segmentation method has better effect than the optimal single atlas segmentation method,and the performance of the new improved fusion algorithm proposed in this paper is better than the optimal single atlas,MV and STAPLE fusion algorithm,and also the effectiveness and accuracy of the proposed algorithm in medical image segmentation is explained.