Using Computer Vision Tools To Analyse Human Motion to Study and Manage Parkinson's Disease

该项目探索如何建立计算机视觉工具来分析运动障碍症状，在研究和管理帕金森病（PD）的背景下。帕金森病是一种进行性神经退行性疾病，全球患病率不断上升。活动功能障碍，如步态或转向异常，是帕金森病的主要症状，它们也与生活质量密切相关。目前，PD临床诊断和评估的金标准是临床面对面评估。当临床医生在医院短时间内观察PD患者时，他们的行为与家庭环境不同，因此评估不能完全评估日常运动障碍。因此，频繁、客观地测量帕金森病将有利于帕金森病的研究和治疗。该项目的新颖之处在于应用最先进的深度学习计算机视觉模型来分析PD患者的特殊、有价值的、专家注释的数据集，就像使用Sphere技术收集的数据集一样。在家庭环境中，可能会收集到更深入、更个性化的信息，随着计算机视觉技术的进步，可以从这些数据中获得更有洞察力的知识。具体来说，该项目涉及训练有效的机器学习和深度学习模型，以估计定制视频数据集中的人体姿势，并对疾病进展进行下游分析，并获得其他统计见解。临床医生的专家分析可以用作基础真相注释，以训练模型进行定量和定性分析。

The project explores how to build computer vision tools to analyse motion disorder symptoms, in the context of studying and managing Parkinson's Disease (PD). Parkinson's' disease is progressive neurodegenerative disease, with an increasing global prevalence. Mobility dysfunctions, like gait or turning abnormalities, are the main symptoms of Parkinson's disease and they also tightly connect with the quality of life. Currently, the gold-standard of clinical diagnosis and evaluation for PD requires face-to-face assessment in the clinic. People living with PD, when being observed by a clinician in the hospital in the short time window, behaves differently from a home-setting and therefore the assessment is not able to fully evaluate the everyday motor disorders. Therefore, a frequent and objective measurement of PD will benefit both the study and management of Parkinson's disease. The novelty of the project is applying state-of-the-art deep learning computer vision models to analyse special, valuable, expert-annotated dataset on people with PD, like the dataset collected using Sphere technology. In a domestic-setting, potentially more in-depth, more personalized information is collected, and with the advancement in computer vision technology, more insightful knowledge could be learned from these data. Specifically, the project involves training effective machine learning and deep learning models to estimate human pose in the customized videos datasets, and to perform downstream analysis on disease progression, and to deriving other statistical insights. Clinicians' expert analysis could be used as ground truth annotations to train the model to perform quantitive and qualitative analysis as well.