Ensemble uncertainty estimation improves skin cancer malignancy prediction

Widespread access to imaging technologies and stronger machine learning (ML) architectures for dermatology tasks such as malignancy prediction have spurred a race to develop models to assist in the automated diagnosis of skin cancer. However, high diagnostic performance on benchmarking datasets quickly deteriorates when models are challenged with data from disparate clinical sources. Generalization gaps stem from the high variability in skin lesion images due to lighting, capture angle, imaging technology and patient phenotype among other factors, impeding the safe application of diagnostic ML models in practice. In this study, we apply a novel multi-criterion uncertainty-estimation approach to detect out-of-distribution skin lesion images from four publicly available datasets across five countries. Using our method, Supervised Autoencoders for Generalization Estimates (SAGE), we quantify likeness of images from patients in Argentina, Brazil, Austria, North Macedonia, Turkey, Australia and the United States to the popular HAM10000 benchmarking dataset and identify problematic image artifacts affecting the reliability of predictions in a pre-clinical setting. We show how filtering images based on SAGE score thresholds can improve the performance of a separate malignancy prediction model and how our approach is robust to variations in image modality and the introduction of new diagnostic classes, providing users with a powerful tool for interrogating key differences between their data and the training distribution of an ML model before clinical implementation.