Research On Optimized Mutil-Atlas Hippocampus Segmentation

The hippocampus,a brain structure located between the thalamus and the medial temporal lobe of the human brain,is responsible for memory,learning,emotional control and spatial orientation.At the same time,some neurological diseases,such as Alzheimer's disease and depression,are associated with changes in the shape and volume of the hippocampus.Due to the irregular structure of the brain hippocampus,fuzzy edges,it is difficult to distinguish it from the adjacent tissues,resulting in poor segmentation accuracy and low segmentation efficiency.The segmentation based on single atlas is easy to cause serious segmentation error.The segmentation method based on multi-atlas registration can make full use of the image information of multiple individuals,overcome the image difference,and the algorithm has strong applicability and high accuracy.However,due to the complexity of the traditional multi-map hippocampal segmentation algorithm and the tedious calculation process,the main research content is to improve the efficiency and accuracy of the algorithm.In this paper,on the basis of studying the traditional multi-atlas hippocampus segmentation,image registration algorithm based on ANTs and the improved multi-atlas fusion algorithm based on U-Net is proposed.The multi-atlas segmentation method mainly includes three processes:image preprocessing,image registration and label fusion.(1)Based on the image preprocessing stage,the skull of the brain magnetic resonance image first,by calculating the mutual information of gradient similarity of map,the selection,selection and target map more floating image group of joint,reduce error greatly label image's influence on the accuracy of label fusion,and uses bounding box algorithm to extract the hippocampus centered magnetic resonance images as interested area,can effectively reduce the data size.(2)In image registration phase,this article is mainly based on ANTs improved registration method is proposed,in view of the traditional "rough" hybrid registration method of defects,brigadier general ANTs with rigid registration,affine registration and deformation model algorithm combined into a system,through the optimization of the interpolation function and similarity measure,realize the optimization of ANTs registration method,and the resampling method and the improved differential homeomorphism Demons were analyzed,through the improvement of ANTs registration method in improving registration accuracy at the same time,the registration time is shortened by around 50%.(3)The label fusion is proposed based on the theory of deep learning improved U-Net MRI hippocampal map segmentation algorithm,based on the U-Net network was improved,in the input image and its corresponding floating image using the stochastic gradient descent method for training,introduce the residual structure U-Net further enhance network performance,precise segmentation hippocampal formation.The U-Net network was improved by using the label fusion technology to achieve multi-atlas hippocampus segmentation.Combining with the optimization of the early registration process,the high-precision multi-atlas hippocampus segmentation was achieved.Compared with the MV,WV,STAPLE and PBM algorithms,the segmentation accuracy of this algorithm was improved by 1.1%and 12%.In this paper,the improved multi-atlas segmentation algorithm can take into account the efficiency and accuracy of segmentation at the same time.