Fuzzy Rank-Based Ensemble Learning for Eye Disease Classification Using Retinal Images: A Bangladeshi-Specific Dataset with Explainable AI Integration

This study focuses on identifying and categorizing ocular disorders affecting critical components of the eye, including the retina, optic nerve and blood vessels, which can lead to severe outcomes such as visual impairment or blindness if untreated. In Bangladesh, where healthcare resources are scarce, especially in rural regions, early identification of ocular illnesses is crucial to avert permanent vision impairment. To address the deficiency of region-specific datasets, this research employs retinal scans from Bangladeshi healthcare facilities to develop a transfer learning-based solution. Multiple transfer learning models, including DenseNet, EfficientNet, GoogleNet, MobileNet, ResNet50, VGG16, VGG19 and ShuffleNet, have been assessed to determine the best effective architecture for eye illness categorization. A fuzzy rank-based ensemble model was proposed, integrating predictions from these models using the Gompertz function, which adaptively adjusts the weight of each model’s predictions to improve classification accuracy. The ensemble model attained an overall accuracy of $96.16 \%$, demonstrating remarkable precision for illnesses such as Diabetic Retinopathy ($99.62 \%$) and Disc Oedema ($99.63 \%$). Moreover, Explainable AI (XAI) tools like LIME have been included to improve the transparency and interpretability of the model’s decision-making process, allowing medical practitioners to comprehend and validate predictions. By providing a reliable and interpretable AIbased solution, this research significantly improves eye disease diagnosis in Bangladesh, particularly for underserved areas.