Explainable Artificial Intelligence based Deep Learning for Retinal Disease Detection

This research focuses on the automated identification of retinal diseases. To address this challenge, an artificial intelligence-based approach developed utilizing five deep learning models namely Xception, InceptionV4, EfficientNet-B4, SqueezeNet, and ResNet-264. The model leverages transfer learning to enhance its performance. It is trained on a dataset of optical coherence tomography (OCT) images to classify retinal conditions into four categories: (1) diabetic macular edema, (2) choroidal neovascularization, (3) drusen, and (4) normal. The training dataset, sourced from publicly available repositories, comprises 1,08,312 OCT retinal images covering all four categories. The proposed models achieved good results. InceptionV4 outperformed other models across multiple metrics, achieving the highest accuracy (99.50%), precision (100%), recall (100%), AUC (100%), and F1 score (100%). It surpassed SqueezeNet (accuracy: 98.00%, precision: 98.00%, recall: 98.00%), EfficientNet-B4 (accuracy: 98.50%, precision: 98.50%, recall: 98.50%), Xception (accuracy: 78.25%, precision: 80.36%, recall: 77.75%, F1 score: 99.50%), and ResNet-264 (accuracy: 87.75%, precision: 87.94%, recall: 87.50%, F1 score: 87.98%). The results highlight the effectiveness of deep learning models combined with transfer learning in achieving accurate and efficient retinal disease detection. Future research could focus on expanding the dataset and exploring hybrid architectures to enhance classification accuracy and improve generalization across various retinal conditions