New AI Model Boosts Diabetic Retinopathy Diagnosis with Cross-Dataset Accuracy
This AI doesn't just diagnose eye disease—it tells doctors when it's unsure.
Researchers have developed a new uncertainty-aware deep learning model for grading diabetic retinopathy severity. The framework combines a convolutional backbone with lesion-query attention and an evidential Dirichlet-based ordinal regression head. Trained on a multi-domain setup combining APTOS, Messidor-2, and EyePACS datasets, it demonstrates strong cross-dataset generalization with competitive accuracy and high quadratic weighted kappa scores, while providing calibrated confidence estimates for low-confidence cases to improve clinical reliability.
Why It Matters
This approach could enable more trustworthy AI diagnostics in healthcare by quantifying uncertainty, helping prevent irreversible blindness.