自闭症的早期诊断对自闭症儿童的发育改善非常重要。由于传统量表辅助诊断的方法存在专业性较强客观性不足的缺点，近年来，利用客观便捷的眼动技术对儿童进行自闭症早期辅助诊断已成为自闭症领域的研究热点，但如何利用眼动技术挖掘有效的眼动生理指标以提高自闭症早期诊断的客观性和准确性，是该领域研究所面临的挑战。针对这一问题，本文提出并实现了一种基于瞳孔响应和注视行为的自闭症早期辅助诊断系统，该系统可客观高效地对3-6岁儿童做自闭症辅助诊断。本文从瞳孔响应和注视行为这两个眼动的研究方向出发，首先设计了基于瞳孔响应和注视行为的眼动实验，然后在现有的3-6岁自闭症和正常儿童的眼动数据集上针对瞳孔响应和注视行为分别提取了相对瞳孔响应和兴趣区相对注视率两种眼动特征，对这两种眼动特征做了显著性和相关性分析，发现了自闭症儿童异常的相对瞳孔响应，即更小的收缩与潜伏收缩极限，更大的扩张与潜伏扩张极限以及更大扩张幅度与速度，还发现了自闭症儿童对兴趣区表现出更少的相对注视率，这些异常的眼动表现证明了本文的眼动实验在诱发两类儿童差异化的瞳孔响应和注视行为上的有效性。使用两种眼动特征结合机器学习算法分别独立构建了基于瞳孔响应的自闭症分类模型和基于注视行为的自闭症分类模型，采用留一交叉验证在现有的数据集上分别实现了93.33%的分类准确率、93.84%的AUC值和84.21%的分类准确率、79.50%的AUC值，均优于前人在现有数据集上达到的水平。对两个模型进行了融合，融合模型和基于瞳孔响应和注视行为的眼动实验共同构成本文所提出的基于瞳孔响应和注视行为的自闭症早期辅助诊断系统。开展眼动实验采集了新的3-6岁自闭症和正常儿童的眼动数据，并通过融合模型给出预测结果，模拟了系统真实使用的过程，同时使用该全新眼动数据集作为测试集对模型的性能进行了评估，实验结果表明，基于瞳孔响应和注视行为的融合模型在测试集上实现了93.94%的分类准确率，和96.92%的AUC值，优于近年来领域内相关研究工作的水平，使得本文所提出的基于瞳孔响应和注视行为的自闭症早期辅助诊断系统极具临床应用的可行性和潜力。
Early diagnosis of autism is critical for developmental improvement in autism. Due to the shortcomings of traditional scale-assisted diagnosis methods, which are highly professional and lack objectivity, in recent years, eye-tracking technology to assist early diagnosis of autism in children has become a research hotspot in the field of autism. However, it is a challenge for researchers to use eye-tracking technology to mine effective eye movement physiological indicators to improve the objectivity and accuracy of early diagnosis of autism. This thesis proposes and implements an early auxiliary diagnosis system for autism based on pupillary responses and gaze behaviors, which can objectively and efficiently assist autism diagnosis in children aged 3-6.This thesis first designs an eye-tracking experiment based on pupillary responses and gaze behaviors. Then two types of eye-tracking features, relative pupillary response features and relative fixation rate features in the area of interest, are extracted from the existing dataset. The significance and correlation analysis of features were carried out. The abnormal relative pupillary responses and less relative fixation rate in children with autism were found, which demonstrate the eye-tracking experiment's effectiveness in evoking differentiated pupillary responses and gaze behaviors between children with and without autism. The pupillary response based classification model and the gaze behavior based classification model were constructed by using two types of eye-tracking features combined with machine learning algorithms. The two models have an average classification accuracy of 93.33% and 84.21%, an average AUC of 93.84% and 79.50%, respectively, with the Leave-one-out cross-validation method. The two models are fused, and the fusion model and the eye-tracking experiment together constitute the eye-tracking system for the early diagnosis of autism. The performance of models are evaluated on a new test dataset. The experimental results show that the prediction results of the fusion model achieve an average classification accuracy of 93.94% and an average AUC of 96.92%, which is better than previous works. The proposed early diagnosis system for autism based on pupillary responses and gaze behaviors has great feasibility and potential for clinical application.
