基于脑网络建模与解码分析的自闭症检测及异常脑连接机制研究

Autism is a frontier research topic in brain diseases. To date, the clinical diagnosis of autism principally relies on the subjective judgment of behavioral scales; however, the effective neuroimaging diagnosis for autism and its pathological mechanism remain unclear. To address this, based on the resting-state functional magnetic resonance imaging (fMRI) data of autism patients and healthy controls, this project intends to employ large-scale brain functional network modeling and decoding algorithms to explore the effective detection method of fMRI for autism on large dataset as well as its abnormal connectivity mechanism, which are two key scientific issues, and perform following researches: 1) Adopting whole-brain functional connectivity analysis, the project constructs the connectivity patterns for autism individuals and healthy controls, and uses two-stage model designing combines with effect size and autoencoder to develop novel fMRI decoding approach based on the connectivity patterns. 2) By means of multi-task learning and using the relevant behavioral data as auxiliary information, the project explores a joint learning algorithm that integrates disease classification and related behavioral data regression to construct an efficient fMRI detection model for autism. 3) Combined with the analysis of abnormal functional connectivity, the project utilizes dynamic causal network modeling to analyze the effective connectivity at large-scale brain network level, so as to reveal the mechanism of abnormal connectivity in autism. This project will further promote the development of fMRI data mining theory, and provide a new method and technical support for early diagnosis and effective intervention of autism.

自闭症是脑疾病研究中的一项前沿课题。目前自闭症的临床诊断主要依赖行为学量表的主观判断，缺乏有效的脑影像诊断方法，且神经层次的病理机制尚不明确。为此，本项目拟基于自闭症及健康受试的静息态功能磁共振成像（fMRI）数据，通过大尺度脑功能网络建模与解码分析，围绕探索大样本上的高效fMRI自闭症检测方法及其异常脑连接机制两个关键科学问题展开研究：基于全脑功能连接分析构建自闭症及健康受试的脑连接模式，结合效应量及自编码器学习，采用二阶模型设计探索基于脑连接特征的解码方法；基于多任务学习，以相关行为学数据为辅助信息探索融合疾病分类及相关行为数据回归的联合学习方法，构建高效的自闭症fMRI检测模型；结合异常脑功能连接分析，通过动态因果网络建模分析大尺度脑网络水平上的有效连接，揭示自闭症的异常脑连接机制。本项目将进一步推动fMRI脑功能数据挖掘理论的发展，为自闭症的早期诊断和有效干预提供新方法和技术支持。

基于功能磁共振成像（functional magnetic resonance imaging, fMRI）的脑信息解码和脑疾病诊断是脑科学研究中的前沿课题。本课题基于fMRI数据，以脑网络分析、脑信息解码和深度学习方法为理论依据，进行fMRI解码方法研究，并将其应用于自闭症分类诊断。研究基于相关分析、滑动窗分析，及高阶脑网络分析等方法，构建全脑功能连接模式、动态脑功能连接模式以及高阶脑功能连接模式；并基于脑网络特征，构建高效的特征学习与分类模型，主要包括：基于全脑功能连接模式的二阶预测模型，基于卷积神经网络与原型学习的脑连接特征学习与分类模型，基于时空信息融合分析的动态脑功能连接分类模型，基于多级脑网络特征融合的脑疾病分类模型，基于脑功能连接与行为学特征融合的脑疾病分类模型，以及基于多尺度注意力机制与深度网络的脑疾病分类模型。相关研究成果为基于fMRI的脑信息解码和脑网络分类提供新方法和技术支持。本研究在大样本自闭症fMRI数据集上进行多种交叉验证实验，验证模型的有效性，并基于显著性特征分析定位自闭症诊断相关的异常脑区和连接机制，进一步丰富自闭症诊断的影像学标记。本研究的关键技术对基于fMRI的脑功能数据挖掘理论发展具有重要科学意义。同时，面向人工智能辅助医疗的实际需求，本项目研究的脑网络分类方法将对自闭症的影像学早期诊断起重要作用，具有广阔的应用前景。