An Explainable AI Ensemble Model for Acquired Vitelliform Lesion Identification

Retinal diseases gradually weaken eyesight and may even potentially lead to blindness. Recognizing changes in the retina based on OCT imaging allows for the detection of diseases at their early stages and thus for making an appropriate diagnosis. In this study, acquired vitelliform lesions (AVL), drusen, and healthy cases are identified utilizing various CNN-based architectures, such as the Convolutional Gated Recurrent Units U-Net, Deep CNN-GRU Network, Residual Attention CNN. The dataset consisting of OCT images was created using photographs gathered from two research centres and publicly available OCT database. The single models obtained to be very effective in recognition the retinal diseases, obtaining the accuracy of 93.80%, 94.03%, and 95.18% for the Convolutional Gated Recurrent Units U-Net, Deep CNN-GRU Network, Residual Attention CNN, respectively. These models outperformed the pre-trained deep learning architectures, VGG-16, ResNet-18, InceptionV3, and DenseNet-121. In order to further enhance the performance of AVL, drusen, and normal cases identification up to 97.09% accuracy, bagging, boosting, and stacking of ensemble learning methods for all models are applied. Moreover, in order to eliminate the black box effect and indicate on what basis the classifier makes conclusions, two interpretability techniques are employed: Gradient Weighted Class Activation Maps (Grad-CAM) and Shap values. This study gives the in-depth insight into AVL, Drusen, and normal cases identification. Moreover, it provides the most effective CNN-based architecture with high accuracy to support ophthalmologists.