Research On Hippocampus Multi-atlas Fusion Algorithms Based On Deep Learning Network

Hippocampus is an important physiological organ in human and vertebrate brains.It has a significant relationship with human and animal physiological functions such as spatial cognition and memory.Through clinical experiments,it was found that the occurrence of mental disorders such as temporal lobe epilepsy,Alzheimer’s disease and schizophrenia often accompanied by the increase and decrease of hippocampal volume.Therefore,the study of hippocampal volume provides a powerful help for the diagnosis and treatment of various mental disorders.To study the volume of hippocampus,we need to segment the hippocampus from the brain structure map.This paper studies the algorithm of hippocampal segmentation and fusion based on deep learning network.The main contents are as follows:Firstly,because of the special image format,irregular shape,small size and low contrast between the edge and surrounding tissues,human hippocampus medical image segmentation technology is difficult to segment,and the traditional segmentation method achieves satisfactory results.The segmentation algorithm based on atlas registration uses the prior information provided by atlas to achieve the separation of target images,mainly using brain MRI image segmentation.However,the traditional single-atlas fusion algorithm has the drawbacks of incomplete extraction and utilization of atlas information.In order to fully extract and utilize the atlas features of the image to be fused and improve the fusion accuracy of the hippocampus atlas,a multi-atlas fusion algorithm based on deep learning network is proposed.Secondly,in the process of spectral preprocessing,the resampling method and differential homeomorphism algorithm are used to advance the atlas.After registration,the registered atlas are decomposed into base and detail parts,and the accuracy of traditional registration algorithm is improved by weighted average fusion of base atlas.Finally,the VGG deep learning network is used to extract features from detail atlas and fuse them hierarchically.The final result is obtained by weighted average fusion of detail and base atlas,which improves the accuracy of atlas fusion.In this paper,the multi-map image of hippocampus after registration is fused by depth learning method.Firstly,the image of hippocampus is preprocessed,including skull removal,extraction of region of interest(ROI),resampling rough registration and differential homozygous Demons precise registration.After registration,the registration map is specially matched by depth learning network.Feature extraction and image fusion are completed.Because the base and detail features of the atlas are fully utilized,the deficiency of inadequate utilization of the detail features of the atlas by traditional algorithms can be improved.The experimental results show that the accuracy of atlas fusion is higher than that of traditional fusion algorithm,and the evaluation indexes are improved.