Revolutionizing diabetic maculopathy detection with MobileNet, GAN-enhanced imaging, and Graph Neural Networks: A multimodal AI approach for precision ophthalmology

Abstract

Diabetic Maculopathy (DM) is a serious complication of diabetes that damages the small blood vessels in the macula, threatening central vision. Timely detection is essential for effective intervention and vision preservation. Traditionally, ophthalmologists have relied on labor-intensive manual examinations of retinal fundus images, which may delay diagnosis and treatment. This study proposes a modified MobileNet deep learning model for the automated detection and classification of DM at different stages, enhanced by the integration of clinical data and Optical Coherence Tomography (OCT) images. Synthetic fundus images were generated using Generative Adversarial Networks (GANs) to address data scarcity and class imbalance, focusing on underrepresented classes such as Severe maculopathy. External datasets, including Messidor and EyePACS, were also incorporated to validate the model’s robustness and generalizability across diverse populations. The proposed model was trained on a unified dataset encompassing fundus images specifically annotated for diabetic maculopathy with varying degrees of severity. The model analyzes these images to extract relevant features and accurately classify them according to the corresponding stages of maculopathy. Achieving a training accuracy of 96% and a validation accuracy of 89.95% (five-fold cross-validation repeated twice), this study underscores the potential of this method for enhancing clinical applications. Furthermore, it represents a significant advancement in the automated assessment of diabetic eye diseases using deep learning. Future work will involve evaluating the model’s effectiveness in real-world clinical settings and exploring methods to improve its transparency and reliability.