Continual Learning in Medical Imaging: A Systematic Review of Methods, Applications, and Clinical Translation

Deep learning supports many tasks in medical imaging, but most systems assume a fixed data source and a single training event. In clinical practice, new scanners arrive, protocols change, patient populations shift, and new diagnostic questions appear. A frozen model loses accuracy as the data shift, while a model retrained only on new data forgets what it knew (catastrophic forgetting). Continual learning updates a model on new data while keeping prior knowledge. Following PRISMA 2020, we searched five sources with database-specific search strings and included 77 primary studies. We formally define the task-, domain-, and class-incremental scenarios in one notation and organize methods into five families: regularization, replay, dynamic architectures, foundation-model adaptation, and hybrids. We appraise the design and reporting of every included study on ten items, and the main findings hold when preprints are excluded. Only 7 of the 77 studies test on an external source, and none reports a clinical evaluation. Evaluations differ too much to rank the families across studies; so, we match families to clinical constraints by mechanism and draw a practical decision guide from head-to-head results on public data. We relate the methods to the regulation of model updates and give specific recommendations.