DeB5-XNet: An Explainable Ensemble Model for Ocular Disease Classification using Transfer Learning and Grad-CAM

Vision is ones' window to the world, enabling individuals to fully engage in and enjoy various aspects of daily life. Diseases can result in loss of vision. Developing an automated diagnostic system is crucial to support the limited number of ophthalmologists in managing the growing number of patients with severe ocular diseases. This research proposes a multiclass diagnostic system for identifying three ocular diseases: Diabetic Retinopathy (DR), Glaucoma (G), and Cataract (C). The study begins by identifying Contrast-Limited Adaptive Histogram Equalization (CLAHE) in LAB color space as an effective image enhancement technique for fundus photographs (FP). By leveraging the unique architectures of various pre-trained models, different features of an image are identified. We evaluate combinations of seven pre-trained models and find that the combination of DenseNet121 and EfficientNetB5 (DeB5-XNet) achieves the highest test accuracy of 95%, with a significant reduction in false negatives compared to individual models. Grad-CAM is then employed to provide a visual explanation of our ensemble model's predictions, demonstrating that the features captured by the model are aligned closely with those utilized by ophthalmologists for diagnosis.