Leveraging Diffuser Data Augmentation to Enhance ViT-Based Performance on Dermatoscopic Melanoma Images Classification

Abstract

The worldwide prevalence of skin cancer, particularly malignant melanoma, has greatly increased. As a consequence, optimal disease management for early diagnosis and intervention is required for improved patient prognosis. To facilitate this, new technologies such as Machine Learning and Deep Learning have been employed to analyze dermatoscopic images. Some applications involve abnormal tissue structure identification and lesion classification. In this study, we introduce a Denoising Diffusion Probabilistic Models (DDPM) approach for data augmentation (DA) to enhance the predictive classification capability of Vision Transformer (ViT) for dermatoscopic image analysis. The ViT was used as deep architecture to evaluate the contribution of diffuser-based and classical DA techniques. The classifier was trained to distinguish between Melanoma, Nevus and Seborrheic Keratosis patches. Experimental trials performed on a publicly available melanoma dataset showed enhanced performance of Diffuser-based DA over classical DA. Furthermore, this study presents, analyzes, and discusses the findings, emphasizing the practicality and efficiency of the proposed method.