Dynamic model merging for class-incremental learning on imbalanced ocular disease image dataset

Deep Learning, a subset of artificial intelligence that is primarily used to discern patterns within collections of data, has demon- strated strong potential for its usage in medical image analysis. However, as more and more new data is introduced into a trained Deep Learning model, previously learned information is gradually forgotten, a phenomenon known as catastrophic forgetting. This can be a hindrance to its long-term effectiveness. Class Incremental Learning (CIL) allows models to learn new classes sequentially while retaining prior knowledge. Dynamic Model Merging (DynaMMo) is an efficient CIL method that merges lightweight task- specific adapters. Despite its efficiency, DynaMMo still faces imbalance challenges, especially with limited memory constraints and imbalance within a dataset. While DynaMMo partially addresses this during classification head fine-tuning through under- sampling, the initial adapter training stage remains susceptible. Therefore, we propose an extension of the DynaMMo framework for eye disease classification using fundus images by incorporating Class-Balanced Focal Loss. Our approach combines focal loss with effective-number weighting during the adapter training stage, then evaluated on an imbalanced subset of the RFMiD dataset. Our approach significantly improves latest accuracy, average accuracy, and macro F1-scores under limited memory constraints. An ablation study confirms that the combined benefits of focal loss and effective-number weighting are crucial for improved perfor- mance.