MARVEL boosts medical AI safety with 37% fewer false positives on unknown cases

MARVEL, proposed by A.S. Anudeep and Vaanathi Sundaresan, tackles out-of-distribution detection for long-tailed medical datasets—a critical gap in clinical AI safety. Most existing OOD methods assume balanced data and coarse OOD sources; MARVEL works on diverse, imbalanced medical data. Its three-component architecture includes a Nonlinear von Mises-Fisher (NvMF) classifier that learns non-linear decision boundaries (with a theoretical link to cosine classifiers), a multi-expert framework where margin-aware NvMF classifiers specialize in different label-distribution regions to handle imbalance, and a dedicated outlier expert trained to separate inliers from outliers.

Evaluated on three retinal, skin lesion, and colorectal histopathology datasets (RFMiD, ISIC2019, NCTCRC), MARVEL consistently beats state-of-the-art OOD methods. It achieves mean FPR95 reductions of 8.45%, 13.02%, and 36.90% respectively, with comprehensive ablations confirming each component's contribution. This allows automated diagnostic systems to reliably flag unfamiliar cases for clinician review, reducing false alarms and improving trust in AI-assisted workflows. Code is publicly available.