PRISM-DR uses per-lesion specialist models to detect diabetic retinopathy
Specialist models outperform single multi-class detectors on rare lesions by 0.561 AP50
Researchers introduce PRISM-DR, a pipeline that trains separate YOLO detectors for each of four diabetic retinopathy lesions, using Bayesian-optimized augmentation and per-lesion ensembling of five cross-validation folds. Trained on the IDRiD dataset, the system achieves a test mAP50 of 0.527 and F1 of 0.529, with the highest AP50 on hard exudates at 0.561. The authors note that these modest absolute results reflect a small single-source training set and a difficult task, but treating each lesion as a separate detection problem offers a practical alternative to a shared multi-class model.
- PRISM-DR trains four independent YOLO detectors, one per lesion type, each with custom augmentation and config.
- Achieves mAP50 0.527 and F1 0.529 on IDRiD; hard exudates reach 0.561 AP50.
- Inter-lesion suppression resolves overlaps using physical size and clinical priority instead of confidence scores.
Why It Matters
Per-lesion specialist models can catch rare early signs better, potentially reducing preventable blindness rates.