Integrating Convolutional and Transformer Networks for Efficient Retinal Disease Screening

Vision impairment worldwide is significantly caused by retinal disorders like diabetic retinopathy, glaucoma, and age-related macular degeneration. To prevent permanent vision loss, timely and accurate diagnosis is essential, together with suitable medical treatment. The manual analysis of retinal fundus images utilized in regular screenings is frequently laborious and susceptible to inconsistency. Besides these weaknesses, such approaches often miss significant subtle indicators in extensive and complex image datasets. In this study, we create an automated system utilizing deep learning to identify retinal diseases through publicly accessible datasets EyePACS and ODIR-5K. We suggest a combined model utilizing Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) to enhance feature extraction and classification performance. The initial noteworthy aspect of this project is image preprocessing, which includes noise elimination, image resizing, and normalization to ensure uniformity for model inputs. The second contribution involves applying random oversampling techniques to enhance the dataset for additional disease categories, thus more effectively equipping the model to learn from minority classes. The data refinement was employed to train and assess both the traditional CNNs and the hybrid CNN-ViT model. The suggested architecture achieved an accuracy of 97.9% on the EyePACS dataset, showcasing its classification effectiveness for various retinal diseases. Additionally, techniques for visualizing attention maps.