OptiVerse AI: Explainable Deep Learning and Three-Dimensional Retinal Reconstruction from Fundus Images for Early Retinal Disease Detection

Abstract

Diabetic Retinopathy (DR) is a major cause of blindness, but early detection is hindered by poor access to retinal imaging services and lack of interpretability of existing artificial intelligence (AI) systems. In this work, we introduce OptiVerse AI, an end-to-end architecture of explainable deep learning along with computational three-dimensional retinal reconstruction based on conventional color fundus images. The proposed system relies on transfer learning based on EfficientNetB7 for multi-stage diabetic retinopathy classification, leveraging adaptive class weighting and data augmentation for improved robustness in imbalanced datasets. We apply Gradient-weighted Class Activation Mapping (Grad-CAM) for achieving model interpretability via visualization of retinal regions involved in diagnostic decisions. Moreover, we propose a pipeline for reconstructing retinal depth maps from learned features in order to generate three-dimensional point clouds and meshes without Optical Coherence Tomography (OCT). The system has been developed based on a database of more than 90,000 retinal fundus images and achieved the accuracy of more than 95%. It provides a cloud platform for combining disease detection, model interpretation and structural analysis of retina, thus being a scalable and cost-effective solution potentially useful for improving diabetic retinopathy screening and clinical decision-making.