Deep-learning-based Automatic Hippocampal Sub-region Segmentation Of Magnetic Resonance Brain Image

The hippocampus plays an important role in the spatial positioning and long-term memory of the brain.Clinical studies have shown that changes in the morphological structures of the hippocampus and its subregions are closely related to various brain diseases,such as Alzheimer’s disease.Quantitative analysis of changes in the morphological structures of the hippocampus and its subregions could help early diagnosis of these diseases.Magnetic resonance imaging(MRI)is an important technic to study the morphology of the hippocampus.Therefore,it is very important to accurately segment the hippocampus and its subregions from the magnetic resonance imaging of the brain.However,the boundaries of the hippocampal sub-regions are complicated,and the brightness differences between different sub-regions are not obvious,so it is difficult to segment the hippocampal sub-regions based on MRI.At present,manual segmentation of the hippocampus based on MRI imaging coats much time and requires experienced experts to perform,which is difficult to meet the current needs of medical big data.Research on how to make full use of the latest technologies to improve the segmentation accuracy of the hippocampal sub-regions while stabilizing segmentation performance has become the focus of scholars’ research.Therefore,this thesis focuses on solving these problems,and carries out related research on the automatic segmentation of MRI hippocampal sub-regions based on deep learning methods.The main contents are summarized as follows.1.To improve the MRI hippocampus sub-region segmentation accuracy and stabilize the segmentation performance,considering the problems of large volume differences among the sub-regions of the hippocampus,wide distribution along the long axis,and complicated spatial structures of the sub-regions,a combination of ResNet50,Self-Attention and Spectral Normalization two-dimensional U-shaped network(SASNUnet-ResNet50)was proposed.The self-attention mechanism can better highlight the connections between distant voxels in the hippocampal sub-region feature map,and have a better extraction effect for long-distance features;spectral normalization enables various sub-regions output by the convolution layer.It is more balanced and could avoid the small proportion of sub-regions in the hippocampus that are too small,so as to better restore the image details.Making the model training more stable,and improve the generalization ability of the model.The experimental results show that the proposed method improves the segmentation accuracy of the hippocampus with stable segmentation performance.2.In view of the rich multi-scale spatial structure of the human brain magnetic resonance image,this thesis proposes a three-dimensional fully convolution DenseNet including a feature pyramid(Feature PyramidNetworks,FPN)structure,which is simply called FPN-Dense VoxNet.The network mainly integrates the FPN structure into a DenseNet with a 3D convolution kernel,which better integrates the bottom and high-level features of the neural network to realize the full use of multi-level features.In the case of deep network layers,it could still restore shallow features and restore image details,thus could finally improve the accuracy of the subdivision of the hippocampus.3.This article compares the proposed method with the existing method on the public datasets.The experimental results show that the method proposed in this thesis obtains a higher accuracy on the hippocampal subregion segmentation.The effectiveness of the method proposed in this article is also verified from the visual effects.The related experiments of the proposed method are executed in the public data sets of CobraLab and Kulaga-Yoskovitz,and the segmentation accuracy is superior to the results of the existing methods.