A Comparative Study Of Optimized Multi-Atlas Hippocampal Segmentation Algorithm

The hippocampus(Hippocampus),also known as the hippocampus of the brain,is mainly responsible for learning and memory functions.If the hippocampus in the human brain is damaged,it can lead to brain diseases such as schizophrenia and depression.The structural volume of the hippocampus occupies a relatively small space in the magnetic resonance imaging(MRI)of the human brain,and its shape and topological structure are also more complicated,resulting in the unclear boundary of the hippocampus structure in the human brain MRI image.Therefore,extracting the hippocampus more accurately from the human brain will be more helpful for the measurement of the hippocampus volume and the analysis of its morphology,which is of great significance for the diagnosis and treatment of brain diseases.This article mainly uses multi-atlas medical image segmentation algorithm to segment the hippocampus of the brain to provides a reference for medical research and auxiliary diagnosis.The multi-atlas medical image segmentation algorithm includes three stages:image preprocessing,image registration,and image label fusion.Image registration and image label fusion are the two key stages of the multi-atlas medical image segmentation algorithm.This article will improve and optimize the above two stages to improve the accuracy and efficiency of hippocampus segmentation.The main research contents are as follows:(1)The image preprocessing stage mainly studies skull removal,atlas selection and Region of Interest(RoI)extraction.Skull removal uses the Brain Surface Extractor(BSE)algorithm proposed by Stefan Bauer.The atlas selection uses Normalized Mutual Information(NMI)to calculate the similarity between the MRI to be segmented and other MRIs,and select the atlas most similar to the target image as the final fusion atlas.The region of interest of the image is extracted from the two image libraries with the size of 60*74*67 and 50*50*50 using the bounding box algorithm.(2)In the image registration process,in order to solve the problem of low segmentation accuracy of multi-atlas medical images,four registration methods are used in turn,which are resampling,Advanced Normalization Tools(ANTs)registration,resampling and differential homeomorphic demons registration,and ANTs and differential homeomorphic demons registration.The comparative analysis of the experimental results shows that in the registration stage,after ANTs registration replaces resampling,and then combined with the differential homeomorphism Demons registration,the segmentation results obtained are more accurate.(3)In the stage of image label fusion,five fusion algorithms are compared,which are weighted averaging(MV)algorithm,graphcut label fusion(GM)algorithm based on generative model constraint,metric learning(ML)algorithm,random forests(RF)algorithm and integrated semi supervised label propagation and random forests(RF-SSLP)algorithm.The experimental results show that using ANTs registration instead of resampling,and then combined with the differential homeomorphism Demons registration can improve the accuracy of the five fusion algorithms of MV,GM,ML,RF and RF-SSLP respectively,and reduce the amount of segmentation time.And through the comparison and analysis of the above five fusion algorithms,it is found that the label propagation fusion segmentation algorithm based on the semi-supervised random forest combining ANTs and the differential homeomorphic Demons registration has the highest segmentation accuracy and the shortest segmentation time.In summary,the combination of ANTs and Demons registration is used in the image registration stage,and the RF-SSLP algorithm is used in the image fusion stage.Compared with MV algorithm,GM algorithm,ML algorithm and RF algorithm,this multi-atlas medical image segmentation combination algorithm has the highest segmentation accuracy improvement.The time reduction is the fastest.