PROTON: Prototype-Based Test-Time Online OOD Detection for Medical VLMs

PROTON (PROtotype-based Test-time ONline OOD detection) is a novel post-hoc module designed to enhance out-of-distribution (OOD) detection for medical vision-language models (VLMs) without requiring model modifications or additional training data. By maintaining an online prototype bank and adaptively integrating prototype distance with Maximum Concept Matching (MCM) scoring, PROTON achieved significant improvements on the FLAIR + FIVES benchmark: +23.9 AUROC for covariate shifts, +8.8 for semantic shifts, and +8.1 for far-OOD detection. This advancement is critical for practitioners as it enables reliable zero-shot clinical image classification in the presence of varying input distributions.