Self-Supervised Learning for Facial Skin Lesion Classification

Due to its sensitivity, facial skin lesion diagnosis and treatment represent a huge challenge for patients and dermatologists. Although deep-learning-based approaches have shown remarkable advances in addressing diseases and skin cancers, facial skin lesion detection has lacked effort in this direction and has not been investigated thoroughly. This paper addresses this problem by proposing DermTech, a self-supervised deep learning framework. Our framework consists of different parts that contribute to precisely identifying five common facial skin conditions; Rosacea, Eczema, Acne, Milia, and Melasma. DermTech utilizes a Vision Transformer (ViT) combined with contrastive learning techniques to enhance feature representation. Additionally, it introduces reconstruction loss, which is combined with contrastive loss in a novel multi-tasking framework, to enable strong classification abilities for various skin issues. Our experiments show that each proposed contribution leads to improvement in performance and reaches a total accuracy of 76.70%.