Out-of-Distribution Detection in Medical Imaging Benchmarks using Concentration-Free Density Estimation

The reliable deployment of deep learning systems in medical imaging requires effective mechanisms to detect inputs that fall outside the training distribution. Recent efforts have highlighted the unique challenges posed by medical data, including subtle semantic shifts and domain-specific covariate variations across histopathology, endoscopy, and MRI benchmarks. However, many existing post-hoc OOD detection methods have been primarily designed and validated on generic image domains, raising questions about their robustness in medical contexts. In this work, we investigate the integration of CF-OOD, a concentration-free density estimation approach based on reverse nearest neighbor statistics, within medical imaging benchmarks. The method builds upon the Concentration-Free Outlier Factor (CFOF), which is specifically designed to mitigate distance concentration effects in high-dimensional spaces, making it well-suited for deep latent representations. We evaluate the proposed approach across multiple medical OOD scenarios, including covariate-shifted in-distribution data, near-OOD, and far-OOD settings. Our study provides an empirical analysis of concentration-free density estimation in medical feature spaces and investigates its behavior compared to established post-hoc detectors.