【作者】 黄敏；

【导师】 梁岩；

【作者基本信息】 电子科技大学 ， 信息与通信工程， 2023， 硕士

【摘要】 孤独症谱系障碍(Autism Disorder Spectrum,ASD),是一种以社会交流障碍、刻板行为和狭隘兴趣为主要特征的神经发育障碍性疾病,具有高致残率,影响儿童健康成长。近年来ASD的发病率逐年增加,美国疾病控制与预防中心的一项ASD有关研究报告显示,平均每36名8岁儿童中就有一名被确认患有ASD,而ASD临床诊断存在耗时长、主观性强等问题,导致在经济不够发达的地区由于经济、医疗手段等局限,许多ASD儿童没有得到及时诊断,使得我国儿童的ASD患病率报告远远低于这一结果。因此,迫切需要构建一种快速、经济、有效的客观筛查方法,实现对ASD儿童的大规模排查。以往研究发现,ASD儿童具有非典型的注视模式,因此,将眼动追踪技术应用于ASD的辅助诊断具有很大的潜力。本文以自然情绪为刺激范式,对ASD儿童的眼动轨迹特性进行探索,研究眼动轨迹的特征提取方法,实现对ASD儿童显著特性的客观量化表征,并研究不同的建模方法,构建一种基于眼动轨迹的ASD儿童自动辅助筛查模型。本文的主要研究内容如下:1、探索了ASD儿童感知自然情绪的眼动轨迹特性。通过对照实验,采集了ASD儿童和正常对照组观察自然场景情绪视频的眼动追踪数据。基于机器视觉的脸部感兴趣区域划分标准,实现对刺激材料的感兴趣区域划分。基于视觉心理学理论,研究眼动轨迹特征提取方法,验证ASD儿童与TD儿童的眼动轨迹差异性。结果表明ASD儿童有着异常的扫描策略,且与情绪内容具有相关性。2、提出了眼动轨迹的两种特征提取方法。利用显著性检测技术自动提取刺激材料感兴趣区域,通过注意力矩阵与视频显著性区域进行融合表征眼动轨迹的注意力空间分布,另外,通过编码眼球扫描的显著性区域表征眼动轨迹时序视觉信息。研究了基于残差网络对两种特征进行表征学习的方法。基于两种特征建立的ASD筛查模型取得了79.49%和81.82%的准确率,分别比基线模型提高了2.91%和5.24%。3、构建了一个基于眼动轨迹特征融合的ASD儿童辅助筛查模型。研究基于卷积神经网络以及全连接神经网络实现对眼动轨迹的动力学信息特征、注意力空间分布特征以及时序视觉信息特征进行融合的方法,并探索了视频级眼动轨迹测试结果的融合策略。本文所提出的ASD儿童辅助筛查模型取得了88.01%的准确率,比基线模型提高了11.43%。更多还原

【Abstract】 Autism Spectrum Disorder(ASD)is a neurodevelopmental disorder characterized by impaired social communication,stereotyped behavior,and narrow interests,which can negatively impact the healthy growth of children and has a high disability rate.The incidence of ASD has been on the rise in recent years,with one out of every 44 8-yearold children confirmed to have ASD according to a report by the Centers for Disease Control and Prevention.However,the clinical diagnosis of ASD is limited by subjectivity and time-consuming procedures,which may delay timely diagnosis due to economic and medical limitations in underdeveloped areas,leading to a reported prevalence of ASD among children in certain regions that is lower than the actual prevalence.Therefore,there is an urgent need to develop a rapid,economical,and effective objective screening method to facilitate large-scale screening of ASD children.Previous studies have shown that children with ASD exhibit atypical gaze patterns,suggesting that eye-tracking technology may have great potential to aid in the diagnosis of ASD.In this study,we utilized natural emotion as a stimulus paradigm to investigate the characteristics of eye movement trajectories in children with ASD.We conducted a study on the feature extraction method of eye movement trajectories and achieved an objective quantitative representation of the salient characteristics in children with ASD.Furthermore,we explored different modeling methods to construct an eye-tracking-based automatic assisted screening model for children with ASD.The main research contents are as follows.1.We conducted a controlled experiment to explore the characteristics of eyetracking in children with ASD while perceiving natural emotions.We collected eyetracking data of children with ASD and a normal control group while they observed emotional videos of natural scenes.The areas-of-interest division of stimulus materials was established based on machine vision facial areas-of-interest division standards.We studied the feature extraction method of eye movement trajectory using the theory of visual psychology to verify the difference in eye track between ASD children and typically developing children.The results suggest that children with ASD exhibit abnormal scanning strategies that correlate with emotional content.2.We proposed two feature extraction methods for eye movement trajectories.The first method utilized saliency detection technology to automatically extract the region of interest of the stimulus material,and the attention spatial distribution of the eye track was represented by fusing the attention matrix with the video saliency region.The second method encoded the temporal visual information of eye trajectories by representing the salience regions of eye scans.We studied the method of representation learning for these two types of features based on residual network.The proposed ASD screening model based on the two features achieved 79.49% and 81.82% accuracy,respectively,which were 2.91% and 5.24% higher than the baseline model.3.We constructed an auxiliary screening model for ASD children based on eye-track feature fusion.We studied the method of integrating dynamic information features,spatial attention distribution features,and time-series visual information features of eye movement trajectories using convolutional neural network and fully connected neural network.We also explored the fusion strategy of video-level eye movement test results.The proposed ASD children’s auxiliary screening model achieved an accuracy rate of88.01%,which was 11.43% higher than the baseline model.更多还原