CASIA OpenIR  > 毕业生  > 博士学位论文
基于结构磁共振影像的脑区分割与阿尔茨海默症早期预测研究
赵元兴
Subtype博士
Thesis Advisor刘成林
2022-08-17
Degree Grantor中国科学院大学
Place of Conferral自动化研究所
Degree Name工学博士
Degree Discipline计算机应用技术
Keyword细粒度整脑分割 半监督学习 阿尔茨海默症早期预测 轻度认知障碍转化预测 自适应全局网络
Abstract

阿尔茨海默症(Alzheimer`s Disease, AD)是一种中枢神经系统变性疾病,是最常见的痴呆原因也是老年人最常见的致死原因。由于目前尚无有效药物可以减缓或治愈AD,所以公认最有效的治疗方法即为在疾病发展早期减缓或阻止疾病的进展。因此,开展AD早期预测算法的研发,提高疾病的早期诊断率与准确率具有十分重要的意义。

本文研究基于结构磁共振影像(structural Magnetic Resonance Imaging,sMRI)的脑区分割与AD早期预测方法。具体来说,首先研究细粒度脑区分割方法,并在此基础上进行AD早期预测,即预测轻度认知障碍(Mild Cognitive Impairment, MCI)患者是否会在未来36个月内转化为AD。论文的主要创新性研究成果如下:

提出了一种基于自集成网络与多视角半监督的细粒度脑区分割方法来克服三维分割网络显存消耗过大和标记样本不足的问题。本方法基于全卷积网络进行三维脑区分割,通过图像块适配模块和自集成分类模块可以在有限显存消耗下学习到人脑大范围结构信息和复杂纹理信息。并且在多视角半监督方法中,通过独立训练不同子空间的分割模型组成集成模型,在大量无标记样本上指导分割模型的训练,降低样本标记不足造成的影响。实验表明,本方法在速度和精度方面均优于对比方法。在TITAN X GPU上分割一张人脑sMRI图像的时间是3秒,远快于基于配准的方法和基于深度学习的方法,适合于大规模数据处理场景。同时本方法经修改后,参加多模态脑肿瘤分割竞赛(Multimodal Brain Tumor Segmentation Challenge,BraTS),在510支注册队伍中获得第二名。

提出了一种基于自适应全局网络与对称一致性损失的细粒度脑区分割方法,进一步克服三维分割网络显存消耗过大和标记样本不足的问题。本方法基于全卷积网络进行三维脑区分割,通过自适应全局网络结构在有限显存资源下学习人脑完整的全局结构信息,并利用对称一致性损失,将人脑两半球对应脑区在无标记样本上进行相互学习,降低样本标记不足造成的影响。实验表明,本方法在速度和精度方面均优于对比方法。在TITAN X GPU上分割一张人脑sMRI图像的时间在1秒以内,远快于其它分割方法,适合于大规模数据处理场景。

提出了一种基于区域集成网络与关系正则化损失的AD早期预测方法来克服病灶区域小和标记样本不足的问题。本方法基于深度神经网络进行AD早期预测,利用上文设计的细粒度脑区分割算法构建区域集成网络,独立学习不同脑区的图像特征,避免关键小脑区特征难以学习的问题。同时本方法利用关系正则化损失,通过建模疾病不同阶段样本间关系扩充训练集合并正则化训练过程。实验表明,本方法在精度方面优于对比方法,并且可以定量评价不同脑区对于预测的贡献,对预测结果进行解释。

提出了一种基于分层注意力网络的AD早期预测方法来克服深度神经网络模型难以解释的问题。本方法基于深度神经网络进行AD早期预测,根据人脑解剖学知识和脑区分割结果,通过分层注意力网络对人脑结构进行分解,建立细粒度脑区与预测结果间的关系,对预测结果进行解释。实验表明,本方法可以更细致地描述不同脑区与疾病间关系,并且分类性能与粗粒度脑区预测模型保持一致。

Other Abstract

Alzheimer's disease (AD), occurring in the central nervous system, is the most common cause of dementia and death among the elderly. No treatments available today can slow down or reverse the progression of this fatal disease. It is of great significance to develop algorithms for the early prediction of AD to improve the early prediction rate and accuracy.

This thesis studies the methods for brain region segmentation and early prediction of AD using structural magnetic resonance images (sMRI). Concretely, we propose the fine-grained whole-brain segmentation methods to split the whole brain into the non-overlapped regions. Based on these regions, the AD early prediction methods are proposed to predict whether the Mild Cognitive Impairment (MCI) would convert to AD within the following 36 months.The major contributions are as follows:

A whole-brain segmentation method based on self-ensemble network and multiview semi-supervised learning is proposed to overcome the lacking of annotated data and the limitation of Graphics Processing Unit (GPU) memory. Segmentation is performed using fully convolutional networks. A novel self-ensemble architecture and a patch adaptation module are proposed to take the advantage of brain's structure and texture information, and a multiview semi-supervised learning method is proposed to exploit unlabeled data. Extensive experiments demonstrate that the proposed method dramatically outperforms previous methods in both speed and accuracy. The model segments an sMRI within 3s on a TITAN X GPU, much faster than multi-atlas-based methods and previous deep learning methods. Also, this model was applied to participate in the Multimodal Brain Tumor Segmentation Challenge and won the second place among 510 teams.

A whole-brain segmentation method based on adaptable global network and symmetric consistency loss is proposed. The proposed method segments the whole brain based on fully convolutional networks. It learns the brain's global structure under the limited computation resource with a memory-efficient framework and uses a symmetry consistency loss to learn the model parameters from plenty of unlabeled data by a symmetry consistency loss. Extensive experiments conducted on four datasets demonstrate that the proposed method outperforms previous methods and has achieved state-of-the-art performance. In terms of speed, the model segments a raw MRI image within 1s on TITAN X GPU, much faster than previous 3D deep learning methods.

An AD early prediction method based on the region ensemble model and relation regularized loss is proposed to address the subtleness of the lesion and the limitation of labeled data. To improve the influence of the small key regions in the training set, a region ensemble model based on brain region segmentation is proposed, and the relation of progression stages is taken into a relation regularization loss. Experiments on public datasets for AD early prediction demonstrate that the proposed method has achieved state-of-the-art performance, and the model allows interpreting prediction according to the contribution of each brain region.

An AD early prediction method is proposed based on a hierarchical attention model to improve the interpretability of the diagnosis model. Based on the brain anatomy and segmentation, the method uses hierarchical attention model to establish the relationship between the fine-grained brain regions, so as to allow structure-based interpretation in prediction. Experiments on public datasets for AD early prediction demonstrate that the proposed method can interpret the prediction using the contribution of fine-grained regions while the disease prediction performance is comparable to the coarse-grained diagnosis model.

Pages124
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/49668
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
赵元兴. 基于结构磁共振影像的脑区分割与阿尔茨海默症早期预测研究[D]. 自动化研究所. 中国科学院大学,2022.
Files in This Item:
File Name/Size DocType Version Access License
赵元兴 毕业论文.pdf(5394KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[赵元兴]'s Articles
Baidu academic
Similar articles in Baidu academic
[赵元兴]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[赵元兴]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.