眼动追踪测量在神经行为障碍与退行性疾病早期诊断中的作用研究

Early detection of neurobehavioral and neurodegenerative disorders is critical to disease management and cure. However, existing behavioral tests typically rely on a specific set of cognitive operations (e.g., sorting, logical inference, or the anti-saccade task). Like neuroimaging, often this is too costly and cumbersome to apply to large populations (for diagnostic or screening), especially when patients are not living near a large research facility. Subjects who cannot understand or follow the task instruc-tions are also excluded from these behavioral probes (e.g., infants and cognitively impaired patients)..Our vision is that one ought to be able to quantitatively assess brain function and disease state through much simpler and ecologically-relevant testing. We focus on deciphering how patterns of eye movements, probed by having patients watch a few minutes of TV, can provide a window into brain function and dysfunction. The main promise is to eliminate confounds introduced when requiring patients to follow complex instructions or when running contrived experiments in the lab. Importantly, our approach may be more natural and more sensitive for testing young children and the elderly. To extract “biometric signatures” from the complex eye movement data obtained during natural free viewing, we both examine statistical properties of eye movements (distributions of saccade amplitudes, fixation durations, etc) and investigate the extent to which the eye may be guided towards salient stimuli (as computed by the saliency model). Pilot results demonstrate robust classification of subjects by age and disease groups: child vs. young adult vs. elderly; controls vs. Pakinson’s patients; controls vs. Attention Deficit Hyperactivity (ADHD)..We aim to develop new techniques for computational modeling and comparison with behavior. To test and validate these techniques, we will contrast them with other known state-of-the-art tools, including CANTAB and antisaccade. We will first do this with normal participants across a large age span, and then we will contrast to specific clinical groups of children or elderly with oculomotor / cognitive disturbances to test the veracity of the classifier (ADHD, PD, and extending to Alzheimer’s Disease (AD)). If successful, our approach will demonstrate new ways to quantitatively assess brain function from natural behavior – free viewing – with a long-term goal of yielding better and more cost-effective patient screening and diagnosis.

神经行为障碍和神经退行性疾病的早期检测对于疾病的控制和治疗有着至关重要的作用。然而现有的临床检测方法在应用到大范围的神经行为障碍与退行性疾病的病人人群中时，通常成本过高和难于实施。本项目拟在通过更简单、更符合生态学规律的测试来定量地评估大脑功能和疾病状态。我们重点研究如何能够通过分析患者在观看几分钟电视过程中的眼动模式来了解大脑机能及障碍。从这些在自然观察情况下得到的复杂眼动数据中，我们检测眼动的统计特征，例如扫视振幅的分布、注视时长等，同时通过视觉显著性模型，研究眼睛被导向显著性刺激的程度。运用神经计算模型及行为比较等新技术，我们可以达到抽取出“生物计量特征”的目的。我们研究的初步结果显示出这种新颖的研究手段能够有效地区分受测者年龄以及疾病类型。此研究将揭示用眼球追踪来定量评估大脑机能的新技术，并建立低成本的简易的和早期的检测诊断神经障碍与退行性疾病的方法。

针对常规深度学习网络框架下提取的图像区域特征存在冗余、不够精确的问题，本研究提出了一种增强深度特征的显著性检测模型，通过采用RBM（限制玻尔兹曼机），实现对中间特征进行进一步精细化处理，提升模型对于图像中的显著目标进行检测能力；为了获取更多局部和全局层次的高层次语义特征，采用深度残差网络作为特征提取基本结构，有效解决网络产生的退化问题，实现了均匀目标区域以及精确目标边缘的显著目标检测；为了有效解决图像显著性检测时对尺度敏感的问题，设计一种基于生成对抗网络的多尺度对抗特征学习的图像显著性检测算法，通过该对抗网络结构提取更强表示的特征，在基于VGG16网络结构的基础上，设计了一个新的层即关联层，实现了更加快速精确的判断一幅图像是生成的显著图还是显著图真值数据的目标。