Enhancing Generalization in Training Unbalanced OCT Datasets Using Customized Ensemble Learning

In medical image classification, class imbalance poses a significant challenge to model performance and generalizability. This study introduces an innovative ensemble approach, combining under-sampling and ensemble learning techniques, to improve the generalization of deep learning models in retinal Optical Coherence Tomography (OCT) imaging. The dataset consists of 84,495 OCT images categorized into four classes. To address class imbalance, the dataset was systematically partitioned into balanced subsets for training, ensuring fair evaluation. Experimental results demonstrate the effectiveness of our ensemble approach, with improvements in classification performance based on key metrics such as AUC ROC and AUC PR across multiple classes. The robustness of the ensemble model is validated through evaluation on a test set, providing a more accurate assessment of its generalization capabilities. These findings highlight the efficacy of ensemble learning in addressing class imbalance and enhancing the reliability of deep learning models in medical image analysis. The study emphasizes the importance of strategic data restructuring and balanced sampling strategies to mitigate class imbalance effects, leading to more robust and clinically useful models.