一种基于深度学习神经网络的自闭症谱系障碍诊断模型设计

Autism Spectrum Disorder(ASD) is a representative disease of pervasive developmental disorders. It is a neurodevelopmental disorder characterized by impaired social interaction, verbal and non-verbal communication, and restricted and repetitive behavior. ASD does a great harm to the human health, and the causes of the disease still remain unclear. Therefore, an effective ASD diagnosis and evaluation method is urgently needed. In this project, we will develop a high-accuracy machine learning model to assist doctors to diagnose ASD. The four main aims in this study are listed as follows (1) Collecting suitable neuroimaging datasets from the public available dataset NDAR (National Database for Autism Research). The datasets should contain sufficient samples, and the quality should be satisfactory. (2) Design an artificial intelligence based diagnosis model which is able to conduct combined analysis of both sMRI and rs-fMRI images. (3) Parallelize the model on the GPU (Graphics Processing Unit) for accelerating the computational speed. (4) Develop a method of locating ASD related biomarkers based on the neural network by using the knowledge of graph theory. We expect to build an ASD diagnosis model with higher accuracy, and locate the disease related biomarkers which are meaningful in the clinics. We expect this study will give new insights on the mechanism of ASD, and lead to better therapeutic treatments for ASD patients in the future.

人类自闭症谱系障碍（ASD）是广泛性发育障碍的代表性疾病。病患具有情绪表达困难、社交互动障碍以及沟通困难等特点。ASD可严重危害人类健康且致病原因仍未明确，因此急需有效的诊断方法和准确的评估标准。本项目旨在设计一套基于人工智能算法的高精度ASD诊断模型，帮助医生对疾病进行诊断。此外，设计一套基于神经影像学图像的ASD生物标记物提取法，协助医生进行病理分析。研究内容包括（1）从适合进行ASD研究的公共数据集NDAR（National Database for Autism Research）下载样本数量充足，图像质量过关的子数据集，并对所得数据进行预处理。（2）设计能整合rs-fMRI及sMRI图像进行分析的、基于人工智能神经网络的ASD诊断模型。（3）将诊断模型在GPU上并行化以提升计算速度。（4）基于计算机图论知识设计基于人工智能神经网络架构的神经影像学生物标记物定位方法。预期此项研究能够获得更加准确的ASD诊断模型，并定位出对ASD临床诊断有意义的生物标记物。从而对疾病的机制加深认识，并为治疗提供更精确的切入点。

人类自闭症谱系障碍（ASD, Autism Spectrum Disorder）是广泛性发育障碍的代表性疾病，目前在儿童中呈高发增长趋势。由于ASD致病原因尚未明确，因此诊断和治疗都较为困难，但是早期的诊断和干预可以有效提高ASD患者的预后。本项目旨在利用磁共振影像（MRI， Magnetic Resonance Imaging）数据，设计一套基于人工智能算法的ASD诊断模型，帮助医生对疾病进行诊断。为了构建高精度的ASD诊断模型，我们首先从北美公共数据集NDAR（National Database on Autism Research）和ABIDE（Autism Brain Imaging Data Exchange）下载了适合进行ASD研究的充足样本，开发了针对sMRI和rs-fMRI图像数据的预处理流程，在保证数据质量的同时比其他预处理流程具有更快的速度；针对两种图像数据，分别设计了基于HOG算法的手工特征提取法和基于卷积神经网络（CNN, Convolutional Neural Network）的自动特征提取法，构建了基于深度神经网络的ASD诊断模型，并基于所得数据设计用来调试模型参数、评估模型性能的交叉验证方案；提出了基于朴素贝叶斯方法和基于深度学习神经网络的两种ASD相关生物标志物定位法，分析了生物标志物与临床症状、治疗反应及临床预后的关联。