Symmetric Contrastive Loss for Out-of-Distribution Skin Lesion Detection

Detecting out-of-distribution (OOD) data has been a challenging task for deep learning models trained with real-life datasets. This work studies OOD detection in medical images where inter-class difference (e.g., variations in visual appearance across separate diseases) outweighs intra-class difference (e.g., same disease but on different locations or people). To improve OOD detection performance, we propose a self-supervised learning approach that can better capture inter-/intra-class variance using a novel symmetric contrastive loss. Two large-scale, publicly-available skin lesion datasets, HAM10000 and DermNet, are adopted in our study. Comprehensive experiments, including three different distributional shifts, disease-specific OOD detection, as well as an adversarial attack, are conducted to validate the effectiveness of our approach.