Vogt AI: A Deep Learning Framework for Early Detection and Severity Grading of Diabetic Macular Edema Through Retinal Imaging and Clinical Insights

This study introduces Vogt AI, an advanced system for detecting diabetic macular edema (DME) through retinal fundus images and clinical data integration. The system incorporates several specialized components MaculaSegmentationNet (U-Net), designed to accurately isolate the macula region in retinal images, enhancing segmentation precision, DME Feature Extractor Net (CNN), which extracts critical features indicative of DME, such as exudates and retinal thickening, and an SVM classifier, which classifies retinal images as either "DME" or "Healthy Eye," with optional severity assessment. For explainability, DME Grad Cam (Grad-CAM) and DME Feature Shapley (SHAP) provide interpretability by visualizing the model’s decision-making process. Additionally, DME Risk Factor Predictor analyzes clinical data to refine DME risk predictions, further enhancing model accuracy. Experimental results show a classification accuracy of 91.8%, with strong sensitivity, specificity, and AUCROC metrics, confirming the model’s potential for early DME detection and personalized management. These findings position Vogt AI as a valuable tool in both clinical and research settings for detecting and managing DME.