Handling Class Imbalance Problem for Diabetic Retinopathy Fundus Image Classification

Diabetic Retinopathy (DR) is a leading cause of vision impairment, necessitating early detection through automated deep learning models. This study explores the use of pre-trained Convolutional Neural Networks (CNNs), specifically EfficientNet B5 and B7, for DR classification especially using imbalanced fundus image dataset from the Asia Pacific Tele-Ophthalmology Society (APTOS) 2019. A key challenge in this task is the class imbalance in DR severity levels, which can reduce model performance. To address this, we employ class weighting, L2 regularization, data augmentation, and upsampling techniques to enhance model generalization. Through extensive experimentation, EfficientNet B5 and B7 architectures were fine-tuned over multiple iterations to optimize performance. Results indicate that EfficientNet B7 achieved a slightly superior test accuracy of 93.1% compared to 92.7% for EfficientNet B5. Both models attained an overall F1-score of 92%, with B7 demonstrating better generalization by achieving the highest minimum class-wise F1-score. While EfficientNet B5 is more suitable for deployment in resource-constrained environments due to its lower batch size, B7 is better suited for large-scale applications. Comparative analysis with related studies reveals that dataset size and model architecture significantly impact performance. The findings underscore the potential of transfer learning and hyperparameter tuning in improving DR classification and early diagnosis, contributing to more accessible and efficient ophthalmic screening solutions.