Dynamic memory to alleviate catastrophic forgetting in continual learning with medical imaging

Deep learning (DL) models in medical imaging are susceptible to performance degradation due to domain shifts from evolving technology and protocols. This paper proposes a dynamic memory (DM) approach for continual learning to address this issue. DM uses a rehearsal method, retaining a diverse subset of training data to mitigate catastrophic forgetting while adapting to new domains. A pseudo-domain (PD) model enhances DM by detecting style clusters and balancing memory. Experiments on cardiac segmentation in MRI and lung nodule detection in CT demonstrate DM's advantage over baseline methods, showing consistent performance improvement and reduced forgetting.

Matthias Perkonigg, Johannes Hofmanninger, Christian J. Herold, James A. Brink, Oleg Pianykh, Helmut Prosch, Georg Lang