Continual Learning in Medical Image Analysis: Appropriate Use and Lifecycle Governance.
Authors
Abstract
AI models for medical imaging are typically trained once and deployed unchanged; however, as scanners, protocols, and patient populations continue to evolve, model performance may degrade over time. Continual learning offers a way to incorporate newly acquired data while limiting the loss of previously learned diagnostic knowledge; therefore, it is particularly suited to clinical environments in which historical images cannot be freely retained or reused. This review outlines the continual learning scenarios and methodological families most relevant to radiology, examines when continual updating is preferable to periodic retraining or simple recalibration, and summarizes the applications in classification, segmentation, detection, and multi-joint radiographic grading. It then addresses the clinical constraints that determine feasibility: privacy and regulatory limits on data retention, delayed and noisy labels, class imbalance, and evaluation requirements beyond the average accuracy. Therefore, continual learning should be implemented as a governed, auditable maintenance process rather than as a single algorithmic choice.