A Survey on Computer Vision-Based Automatic Assessment of Stroke and Facial Palsy

Abstract

Timely and accurate diagnosis is a critical component of stroke treatment and management. Recent advancements in computer vision technologies have shown the potential to aid with the process before the patient has access to radiology. This survey curates current research on computer vision-based approaches for stroke and other facial pathology screening based on the patient's physical symptoms such as facial asymmetry and eye movements. Our work aims to identify gaps in the current state of research and put forward a critical review that could be linked to limited data availability, reproducibility of published methods, and robustness of proposed methods on unseen data. A catalogue of methods utilising machine learning, deep neural networks, and classical computer vision techniques are compiled together with the descriptions of datasets applicable for research in computer vision-based stroke and facial palsy screening. Based on the prevalent state of research where novel techniques are trained and evaluated on datasets that are not released to the public we provide an independent analysis of the state-of-the-art methods based on experimental implementations. The analysis demonstrates a fair comparison between methods through evaluations on identical settings, with the results suggesting a general theme of problems with generalization.