A Regulation-Aware Explainable CNN-Based AI Platform for Early Diabetic Retinopathy Detection with Confidence-Calibrated Clinical Decision Support

Abstract

Diabetic retinopathy (DR) is a leading cause of preventable blindness worldwide, where early detection is critical for timely intervention and vision preservation. This study proposes a Regulation-Aware Explainable Convolutional Neural Network (CNN)-based AI platform for the early detection of diabetic retinopathy from retinal fundus images, integrated with confidence-calibrated clinical decision support. The proposed system is designed to align with key medical AI regulatory principles, including transparency, reliability, and clinical safety. A deep CNN architecture is trained on publicly available retinal imaging datasets to classify early-stage DR with high sensitivity. To enhance clinical trust and interpretability, explainable AI techniques such as heatmap-based visual explanations are incorporated, enabling clinicians to identify image regions contributing to model predictions. Additionally, uncertainty estimation and confidence calibration methods are employed to quantify prediction reliability, allowing the system to flag low-confidence cases for expert review. The platform is deployed as a web-based clinical decision support tool, facilitating seamless integration into ophthalmology workflows. Experimental results demonstrate robust diagnostic performance, effective confidence calibration, and improved interpretability without compromising accuracy. This regulation-aware, explainable, and confidence-driven AI framework has the potential to support ophthalmologists in early DR screening, reduce diagnostic workload, and improve patient outcomes while adhering to emerging medical AI regulatory standards.