An Ensemble Transfer Learning Framework for Multi-class Retinal Fundus Disease Classification.

Retinal fundus images provide vital diagnostic information for identifying various eye diseases. Early detection is crucial, as millions of individuals lose their vision each year from conditions that could have been treated if diagnosed in time. Although deep learning has shown strong potential in automated retinal disease diagnosis, existing approaches suffer from several limitations. Many focus only on a single disease, instead of multiclass classification, which limits their real-world use. Several models rely on small or imbalanced datasets without proper balancing, which can create biased models. In addition, few studies use hyperparameter optimization or explainable AI, reducing performance and clinical trust. To overcome these challenges, this study presents an ensemble transfer learning framework for multi-class retinal fundus disease classification using the “Retinal Fundus Image 50k” (RFID50K) dataset, comprising 50,000 high-resolution images across seven disease categories. We apply the Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance and implement hyperparameter optimization (HPO) for improved model training. Multiple pre-trained models, including ResNet152V2, DenseNet121, InceptionV3, and VGG19, are integrated with a custom classifier head. Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to provide interpretability by highlighting image regions influencing predictions. The proposed ensemble framework outperforms most state-of-the-art models, achieving an accuracy of 99.8%, and demonstrates strong reliability and robustness in classifying retinal fundus diseases.