An Evaluation of Self-supervised Pre-training for Skin-Lesion Analysis

Self-supervised pre-training appears as an advantageous alternative to supervised pre-trained for transfer learning. By synthesizing annotations on pretext tasks, self-supervision allows models large amounts of pseudo-labels before fine-tuning them the target task. In this work, we assess diagnosing skin lesions, comparing three self-supervised pipelines a challenging baseline, five test datasets comprising in- and out-of-distribution samples. Our results show that is competitive both in improving accuracies reducing variability outcomes. Self-supervision proves particularly useful low training data scenarios (<1500 <150 samples), where its ability stabilize outcomes essential provide sound results.