Facial Palsy Detection Through Spatiotemporal Landmark Analysis with Attention Mechanisms

This paper presents a novel approach to facial palsy detection using spatiotemporal landmark analysis with attention mechanisms. Unlike previous methods, we utilize 68 facial landmarks to capture subtle facial variations and employ an end-to-end learning approach integrating feature extraction and sequence classification. We use Principal Component Analysis (PCA) on the DigiFace-1M dataset for facial feature extraction. Our self-attention-based sequence learning model emphasizes frames with subtle motion discrepancies. Evaluated on the Toronto NeuroFace (TNF) dataset using five-fold cross-validation, our method demonstrates robust performance, offering a proof-of-concept for clinical diagnosis and treatment planning of conditions like Bell's palsy, Parkinson's disease, and stroke.