MobViGNet: An Explainable AI-based Hybrid Transfer Learning Approach for Eye Disease Classification from External Eye Images

Abstract

The application of artificial intelligence (AI) in medical diagnostics, particularly in ophthalmology, holds great promise for enhancing early detection and improving patient outcomes. However, the complexity of diagnosing common eye diseases from external eye images, along with issues such as limited labeled datasets and the need for model transparency, pose significant challenges. To address these, we propose MobViGNet, an innovative hybrid transfer learning model for multi-categorical detection of common eye diseases, such as cataracts, conjunctivitis, and normal conditions. Our propose MobViGNet, integrating the strengths of MobileNetV2 and VGG19 to achieve efficient and effective eye disease classification from external eye images. The lightweight, depth wise separable convolutions of MobileNetV2 and deep hierarchical feature extraction of VGG19 are integrated through parallel feature learning and subsequent concatenation and custom dense layers to construct a compact yet robust feature representation. Both networks are initialized with pre-trained ImageNet weights and frozen base layers to allow prior knowledge, while the additional layers are fine-tuned with the Adam optimizer and sparse categorical cross-entropy loss. MobViGNet outperforms individual and state-of-the-art networks like Xception, InceptionV3, and VGG19, achieving 99.85% accuracy, 99.77% F1-score, and 0.15% error rate on 5-fold cross-validation. Explainability is incorporated into the model with LIME and Grad-CAM, with interpretable predictions that are crucial for clinical trust. The model also showed excellent performance when tested on an external dataset, confirming its ability to generalize beyond the training data.