Towards Trustworthy Retinal Disease Classification: An Ensemble Deep Learning Approach with Explainable AI

Accurate and interpretable diagnosis of retinal diseases from fundus and optical coherence tomography (OCT) imagery remains a critical challenge for scalable, equitable eye-care delivery. This paper proposes a four-stage ensemble deep learning framework that combines seven convolutional neural network (CNN) backbones—VGG-19, ResNet-50, DenseNet-121, InceptionV3, Xception, EfficientNet-B3 and MobileNetV2—with four ensemble aggregation strategies of increasing complexity (hard voting, soft voting, weighted averaging and stacking) and an integrated explainable AI (XAI) module (Grad-CAM, LIME, SHAP). The framework is trained and evaluated on three public benchmarks (ODIR-5K, RFMiD and Kermany OCT) using patient-level stratified 70/15/15 splits. A stacking ensemble with an XGBoost meta-learner achieves 95.14% accuracy and an AUC of 0.988 on the ODIR-5K eight-class task, a statistically significant 3.56-percentage-point improvement over the best single CNN (EfficientNet-B3, McNemar's test, p<0.001, Cohen's d = 2.31), and outperforms majority voting (Friedman χ²=23.8, p<0.001). Integration of XAI raises clinician agreement with model saliency maps from 76.4% to 91.2% (p<0.001) without degrading diagnostic accuracy (p=0.82, Wilcoxon). A lightweight three-model ensemble retains 90.45% accuracy with 95% fewer parameters than the full ensemble, offering a practical accuracy–efficiency trade-off for resource-constrained clinical deployment.