Evaluating Uncertainty in Deep Q-Network Ensembles for Trustworthy Anomaly Detection in Medical Imaging

Abstract

Reliable anomaly detection is crucial for safe AI deployment in clinical imaging, yet most systems offer limited insight into prediction uncertainty or failure modes—key factors in medical decision-making. We analyze the uncertainty characteristics of a patch-level Deep Q-Network anomaly detection framework (DQN_AD) for brain MRI, trained with few annotated abnormal cases and designed to generalize to highly imbalanced clinical datasets. Our study links uncertainty to model errors, calibration, anomaly scores, spatial correspondence with ground truth, and selective evaluation. Results show that high-uncertainty predictions consistently coincide with error-prone regions, providing a strong signal for identifying potential failures. This study establishes the foundation for uncertainty-aware, reinforcement learning–based anomaly detection models that enhance reliability, interpretability, and clinical usability in large-scale MRI analysis.