Medical Out-of-Distribution Analysis Challenge 2024

Biomedical image computing relies on annotated examples, posing challenges due to cost and scalability issues. Out-of-distribution (OOD) detection, specifically anomaly detection, offers a solution by generalizing to new conditions without labeled data. However, the absence of a standardized medical dataset hinders fair comparisons between approaches.
 
To address this, we propose the Medical Out-of-Distribution Challenge, providing two datasets (brain MRI and abdominal CT) with over 600 scans each. The training set contains scans without identified anomalies, while the test set includes both natural and synthetic anomalies for a controlled comparison. Despite improvements in previous editions, current algorithm performance is not medically relevant, signaling room for methodological advancements.
 
For evaluation, we suggest an object-centric metric for pixel-level tasks, aligning with medical objectives. By maintaining categories, renewing synthetic samples, and using an object-centric metric, the challenge aims to facilitate fair comparisons of approaches in a realistic and clinical setting.