Generalizing a Small Facial Image Dataset Using Facial Generative Adversarial Networks for Stroke’s Facial Weakness Screening

Stroke is a medical emergency resulting from disruption of blood supply to different parts the brain which leads facial weakness and paralysis as control center. leading cause long-term disability significantly changes patient’s life. This paper introduces use image dataset containing neutral smiling expressions classify common sign stroke. Our “real dataset" comprises face images normal subjects stroke patients. However, increase dataset, we added another known “FaceGAN dataset". additional contains pair synthesized public datasets were augmented generate two at eight age groups. The faces divided into left right side using landmark detection technique corrected for geometric distortions through affine transformation matrix Delaunay triangulation. An autoencoder model composed ConvNeXt encoder ConvNet decoder was trained used fine-tune classification our proposed architecture. Results four-fold cross validation showed that less prone overfitting when with FaceGAN an average AUC 0.76 F1-score 71.19%, compared without data only achieved 61.54%. study shows can efficiently generalize models programs small detection. work be further improved optimized clinical application in future.