Deep Learning for Comprehensive Retinal Health: Beyond Automated Fundoscopy - Classification, Segmentation, and Retinal Image Reconstruction

This study presents a deep learning approach aimed at expanding the scope of retinal health analysis beyond standard fundoscopy, with a focus on classification and segmentation tasks. Results highlight the effectiveness of models like Inception, ResNet, and MobileNet in accurately classifying retinal features, with Inception demonstrating superior performance (accuracy: 96%, precision: 90%). Despite MobileNet's slightly lower accuracy, its lightweight architecture proves suitable for realtime applications like smartphone fundoscopy (accuracy: 91%). Moreover, the study proposes leveraging deep learning-based on Semi-supervised segmentation methods, such as GANs, autoencoders and U-N et, to address challenges in dataset segmentation without ground truth annotations. These findings significantly contribute to advancing automated retinal image analysis for efficient diagnosis of eye diseases, offering valuable insights for future research and clinical applications.