Improving Facial Acne Segmentation through Semi-Supervised Learning with Synthetic Images

Facial acne is a very common skin condition that can worsen or leave scars if left untreated. Hence, deep learning-based methods have been proposed to automatically detect acne. However, acquiring medical images like acne is difficult, and generating labels without expert advice is challenging. This study generated synthetic images using GAN and then improved acne segmentation performance through semi-supervised learning methods. To validate this, acne images were acquired using skin analysis equipment, and these were divided into labeled images with ground truth and unlabeled images without it. The GAN was then trained using the labeled images. By using the generated GAN model for semi-supervised learning to train an acne segmentation model, a performance of 71.47% was achieved, surpassing the 70.62% obtained through supervised learning alone. Furthermore, the performance reached 71.48% when trained on real unlabeled images, demonstrating that GAN use can produce results that are comparable to those obtained with real images.