基于异构数据融合及脑网络构建的阿尔兹海默病早期诊断关键技术

Alzheimer's Disease (AD) is an irreversible neurodegenerative disease. How to diagnose it early is the key problem to slow down the course of AD effectively. At present, the two common diagnostic methods have their own defects: (1)the image-based computer-aided diagnosis is difficult to train network because of the small number of effective samples; (2)the clinical test based on neuropsychology cannot achieve early diagnosis. In view of this, this project studies the key technologies of early diagnosis of AD based on heterogeneous data fusion and brain network construction. Complete the early diagnosis data set by fusing multi-modal images with clinical diagnosis. Improve the accuracy of network training based on small sample data by building transfer learning model of AD early diagnosis. Explore the pathology and course of AD by constructing feature-based brain network and making features visible. This study will provide new ideas for early computer-aided diagnosis of AD and new technologies for development of treatment. It has important value of theoretical research and practical application.

阿尔茨海默病(Alzheimer's Disease, AD)是一种无法逆转的神经退行性疾病，如何早确诊是有效减缓AD病程的关键问题。目前常见的两种诊断方式均存在各自的不足：(1)基于影像学的计算机辅助诊断方式因为有效样本少极易出现网络训练困难；(2)基于神经心理学的临床测试方式无法实现早期确诊。鉴于此，本项目研究基于异构数据融合及脑网络构建的AD早期诊断关键技术。将多种模态图像与临床诊断相融合，完备早期诊断数据集；搭建迁移学习AD早期诊断模型，提高小样本数据的网络训练准确率；构建基于特征的脑网络模型，实现特征可视化，探究AD病理及病程。本研究将为AD早期计算机辅助诊断提供新思路，为治疗方法的开发提供新技术，具有重要的理论研究与实用价值。

近年来，图像融合广泛应用于各种领域，在医学诊断上具有很好的使用价值。本项目从 AD 研究权威网站 ADNI 数据库收集患者的脑部结构影像(MRI)和功能影 像(PET、DTI)，建立用于神经网络预训练的预训练数据集，对 AD 数据集中的图像进行图像增强的预处理。研究前期使用卷积神经网络对图像进行融合，后期对于融合网络进行改进，得到了更好的融合效果。本项目使用卷积神经网络分别对单模态数据集和多模态融合数据集进行分类，实验证明，图像融合能够提升图像分类的准确率。本项目提出的方法进一步提高了AD图像分类的准确率，对 AD 的早期诊断与病情发展研究有着重要的理论价值和现实意义。