Dynamic model merging for class-incremental learning on imbalanced ocular disease image dataset
Dean Hans Felandio Setiadi Saputra, Aldrich Reinhart Wahyudi, Jakendra Lathaniel Sulistijo, Hidayaturrahman
Procedia Computer Science 269 (2025), pp. 649–658 · ICCSCI 2025 conference proceedings
The problem
A study on dynamic model merging techniques for class-incremental learning applied to imbalanced ocular disease image datasets, addressing the challenges of continual learning in medical imaging with limited and imbalanced data.
How we approached it
Extended DynaMMo with class-balanced focal loss during adapter training, combining focal loss with effective-number weighting on an imbalanced RFMiD subset.
The results
With a 25-sample memory buffer, latest accuracy increased from 58.03% to 69.30% (+11.27 percentage points), average accuracy from 71.46% to 73.73%, and macro F1 from 0.4784 to 0.5962. The experiment used ResNet18 on 10 RFMiD classes across five incremental tasks. At 200 samples, latest accuracy slightly decreased (71.46% to 71.22%) while macro F1 improved (0.6693 to 0.6866). The benefit is strongest under the smaller memory constraint; these are research test-set results, not clinical validation.
| Memory samples | Latest accuracy Baseline → ours | Macro F1 Baseline → ours |
|---|---|---|
| 25 | 58.03% → 69.30% | 0.4784 → 0.5962 |
| 200 | 71.46% → 71.22% | 0.6693 → 0.6866 |
My contribution: research methodology, experiments, and drafting the methods, results and discussion, as recorded in the paper’s author contribution statement.