A Study on Explainability of Deep Learning Model for Image Classification Using CycleGAN

In recent years, there has been growing interest in Computer-Aided Diagnosis (CAD) and the widespread application of Convolutional Neural Networks (CNNs) for image-based diagnosis. However, CNN-based diagnostic approaches have faced challenges in terms of interpretability. Previous studies have proposed the use of CycleGAN in analyzing the classification process of CNNs, as it offers the potential for enhanced interpretability. While successful in obtaining interpretability for simpler disease images, such as cardiac hypertrophy, it has been more challenging to achieve interpretability in complex tasks like gender classification in brain imaging. Therefore, this study aims to employ human facial images, a task easily comprehensible to humans, to conduct gender classification using CycleGAN and analyze the criteria and tendencies involved in gender transformation. Through the analysis of pre- and post-transformation images, as well as pixel value differences, we aim to explore the relationship between gender differences and image disparities. Findings indicate notable changes in factors such as hair length, eye size, and skin color. Moreover, the inclusion of overall image characteristics, beyond individual facial features, suggests the possibility of using external means such as wigs or cosmetics to potentially deceive CycleGAN’s gender classification.