Mitigating catastrophic forgetting in medical imaging via incremental learning

Deep learning models tend to forget previously learned information when trained on new tasks; this phenomenon, known as catastrophic forgetting, continues to be a major bottleneck in many applications of medical image analysis. This paper benchmarks five continual incremental learning (CIL) approaches: Elastic Weight Consolidation (EWC), Learning without Forgetting (LwF), Memory Aware Synapses (MAS), Synaptic Intelligence (SI), and Variational Continual Learning (VCL) in endoscopic image classification problems. We sequentially trained on two medical datasets, LIMUC (Task 1) and Kvasir (Task 2) using DenseNet121 and ResNet50 architecture as the backbone architecture. Performance was evaluated via macro-averaged accuracy, precision, recall, F1-score, catastrophic forgetting metrics and computational efficiency. Across both models, SI and MAS consistently delivered strong classification performance on Task 2, with DenseNet121 achieving 92.30% and 91.57% precision, and ResNet50 achieving 91.02% and 85.44%, respectively. MAS showed the best trade-off between stability and plasticity, reducing catastrophic forgetting to 7.83% (DenseNet121) and 2.51% (ResNet50) while maintaining over 85% Task 2 accuracy. In contrast, VCL had the lowest forgetting (4.61% and 1.93%), but suffered from poor Task 2 generalization, with accuracies of only 12.00% and 12.17% respectively. These findings highlight that MAS offers a balanced continual learning strategy across architectures, SI is preferable when maximizing adaptation to new tasks, and VCL is optimal when prior knowledge retention is prioritized. This work provides practical guidance for designing continual learning pipelines in medical imaging systems requiring long-term adaptability without performance degradation.