Multi-Level Ensemble Learning for Retinal Disease Classification Using a Weighted Strategy

The classification of retinal diseases is vital for the early detection and treatment of vision-related disorders. This study proposes a multi-level ensemble learning approach employing deep learning models for accurate classification of retinal diseases. We have performed binary classification on the Retinal Fundus Multi-Disease Image Dataset (RFMiD). To address class imbalance, a weighted training strategy is applied, where the class having a lower number of samples has been applied a higher class weight. We have used various convolutional neural network (CNN) models, including a custom CNN and many pretrained models like VGG16, DenseNet121, Xception, InceptionV3, ResNet50, etc. We have proposed an ensemble meta learning approach that integrates high-performing convolutional neural networks (CNNs) within a meta-learning framework (logistic regression), leveraging predictions from multiple models to boost overall performance. The proposed ensemble model achieves superior results compared to individual CNN classifiers, with an accuracy of 91%, precision of 92%, recall of 91%, and F1-score of 92%. These findings highlight the effectiveness of ensemble learning and the importance of handling class imbalance in medical image classification.