AL-CAF: An AI Framework for Adaptive Lesion-Aware Contrast Enhancement and Confidence-Calibrated Attention Fusion in Diabetic Retinopathy Detection from Fundus Images

Diabetic retinopathy (DR) remains a leading cause of preventable blindness, and automated five-class severity grading from retinal fundus images is central to scalable screening programs. Two practical gaps limit current deep-learning graders. First, fundus images collected in community screening settings vary widely in illumination, contrast, and focus, yet most pipelines apply a single fixed-strength contrast-enhancement step to every image regardless of its actual quality, which under-corrects poor-quality images and over-processes already-clear ones. Second, fusing features from multiple spatial scales — fine microaneurysm-level detail and coarse exudate/hemorrhage-level detail — typically yields a single softmax probability with no calibration, so the grader is frequently overconfident on the visually similar adjacent grades (e.g., Mild vs. Moderate) where clinical risk is highest. This paper proposes AL-CAF, a framework built on two lightweight mechanisms: Adaptive Lesion-Contrast Enhancement (ALCE), which estimates a per-image contrast-quality score from the green channel and gates CLAHE strength and gamma correction accordingly, and Confidence-Calibrated Multi-Scale Attention Fusion (CC-MSAF), which combines fine- and coarse-scale features through a learned softmax attention weighting and applies validation-fitted temperature scaling to the final grade probabilities. Evaluated on APTOS 2019, IDRiD, and Messidor-2, AL-CAF improves five-class grading accuracy over a plain CNN baseline and an attention-only baseline on every dataset (93.78%, 91.02%, and 92.61%, respectively) and, under simulated acquisition-quality degradation on APTOS 2019, retains 95.10% AUC under Gaussian blur and 94.30% AUC under low illumination — compared to 88.10%/85.40% for the plain CNN baseline and 90.05%/87.90% for the attention-only baseline. These results indicate that quality-adaptive preprocessing combined with calibrated multi-scale attention fusion directly addresses the robustness and overconfidence gaps left open by fixed-preprocessing, uncalibrated DR graders.