Ensemble of Self-supervised Learning Methods for Robust Skin Disease Image Diagnosis Leveraging Unlabeled Data

Abstract

Deep learning technologies have led to remarkable improvements in medical image analysis, but the collection and annotation of medical data remain challenging. This study leverages self-supervised learning and unlabeled images from the National Skin Disease Database of Japan (NSDD) to enhance skin disease classification. By generating pre-trained models using three self-supervised learning methods, and comparing them to a baseline pre-trained model on ImageNet, we found that pre-training with unlabeled images increased transfer learning performance by 1-2% on labeled images in the NSDD. We also demonstrated the effectiveness of the pre-trained models across four public dermatological image datasets. We further implemented a hybrid pre-training approach that combines self-supervised and supervised learning, and observed accuracy improvements in certain datasets. Since the optimal strategy is task and dataset-specific, we proposed an ensemble of transfer learned models from diverse pre-training methods for robust classification. The ensemble approach consistently enhanced accuracy across various datasets, with an improvement by up to 8.3% from the baseline, confirming the effectiveness of self-supervised pre-training with unlabeled dermatological images even when the images are mainly from one ethnic group.