Simulation of Facial Palsy Using Cycle GAN with Skip-Layer Excitation Module and Self-Supervised Discriminator

Abstract

The Yanagihara method is used to evaluate facial nerve palsy based on visual examinations by physicians. Examples of scored images are important for educational purposes and as references, however, due to patient privacy concern, actual facial images of real patients cannot be used for educational purposes. In this paper, we propose a solution to this problem by generating facial images of a virtual patient with facial nerve palsy, that can be shared and utilized by physicians. To reproduce the patient facial expression in a public face image, we propose a method to generate a swapped face image using the improved Cycle Generative Adversarial Networks (Cycle GAN) with a skiplayer excitation module and a self-supervised discriminator. Experimental results demonstrate that the proposed model can generate more coherent swapped faces that are similar to the public face identity and patient facial expressions. The proposed method also improves the quality of generated swapped face images while keeping them identical to the source (genuine) face image.