Research On Multi-atlas Segmentation Algorithm Of MRI Hippocampus Based On ANTs Registration

The hippocampus is an important part of the human brain.It is located between the thalamus and the medial temporal lobe.It is mainly responsible for learning,memory and spatial location.Studies have shown that many kinds of mental diseases,such as Alzheimer's disease,temporal lobe epilepsy,schizophrenia and so on,are closely related to the changes of hippocampal volume and morphology.Precise segmentation of the hippocampus and accurate measurement of its volume in Magnetic Resonance Imaging(MRI)can provide assistance for clinicians in the diagnosis and treatment of brain diseases.The research on the segmentation of the hippocampus from magnetic resonance image is also of great significance for the research of various mental diseases.Due to the blurred boundaries of the anatomical structures of the hippocampus,amygdala and other brains in the magnetic resonance image,the contrast is low,the hippocampus is small and the anatomical structure is complex,the traditional segmentation method cannot achieve efficient and accurate segmentation of the hippocampus.High-precision segmentation of hippocampus from brain magnetic resonance image has also become the focus and difficulty of academic research.Therefore,this paper focuses on the segmentation method based on multi-atlas registration,using the prior information provided by the atlas to achieve accurate segmentation of the hippocampus.This paper analyzes the characteristics and segmentation framework of multi-atlas segmentation method in detail,including three stages:image preprocessing,image registration and label fusion.In the stage of image preprocessing,we focus on skull removal,atlas selection and region of interest extraction to ensure the accuracy of subsequent registration and label fusion.In the registration stage,Advanced Normalization Tools(ANTs)are used to replace resampling process in the coarse registration.The local neighborhood similarity of each pixel is used as the objective function of registration,and the deformation field is effectively constrained and smoothed to achieve better registration effect and then the diffeomorphic Demons algorithm is used for fine registration,which improves the accuracy and speed of registration.In the label fusion stage,we studied a variety of label fusion algorithms,using weighted voting,STAPLE,patch-based fusion algorithm(PBM),Graph Cuts algorithm,non-local patch-based sparse constraint algorithm(NLP),Local label learning(LLL)fusion strategies,the label images obtained by ANTs-based diffeomorphic Demons registration are fused,and the segmentation results of each algorithm were compared and analyzed.The experimental results show that the ANTs-based diffeomorphic Demons registration algorithm combined with the weighted selection(WV)fusion algorithm can obtain a better segmentation effect,the mean Dice coefficients of the left and right hippocampus are 0.8989 and 0.9096.