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

Abstract

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