Facial Paralysis Detection exploiting Landmark Asymmetry using MobileNetV3-small and GCNN Ensemble

Facial paralysis, or facial palsy, is a condition characterized by weakness or dysfunction of the facial muscles due to damage to the facial nerve. Accurate and early diagnosis of the condition is important for effective medical intervention and timely recovery. Conventional diagnostic methods and isolated deep learning models are often limited in capturing both global facial features and subtle asymmetries, which reduces their reliability in clinical applications. To address this limitation, facial landmark detection is employed to capture geometric and structural information from regions of the face most affected by paralysis. Specifically, the dlib 68 facial landmark model enables accurate key point localization, facilitating asymmetry analysis. Building on this, we designed a hybrid deep learning model that combines MobileNetV3-Small and a Graph Convolutional Neural Network (GCNN). MobileNetV3-Small is used to efficiently extract global image-level information, while the GCNN utilizes the predefined 68 facial landmarks to model structural dependencies and asymmetry patterns. The combined representation of these two modalities enables accurate classification of facial paralysis. Experiments were conducted on the publicly available dataset “FIASNAS,” which contains 1,259 face images of stroke-affected individuals and 2,511 images of healthy individuals. The experimental results show that the proposed approach is promising and achieves good performance.