PRISM-DR: Per-lesion Retinal Inference with Specialist Models for Diabetic Retinopathy

2026-07-23 · Source: cs.CV updates on arXiv.org · Field: Science & Research — Health & Medical Research, Mathematics & Computational Sciences, Engineering & Applied Sciences · Depth: Expert, extended

Summary

PRISM-DR, a lesion-specific pipeline, addresses the challenge of detecting four non-proliferative diabetic retinopathy (DR) lesions—microaneurysms (MA), hemorrhages (HE), hard exudates (EX), and soft exudates (SE)—which often differ sharply in size, color, morphology, and prevalence. Unlike conventional multi-class models that favor common lesions, PRISM-DR employs a separate single-class YOLO detector for each lesion, allowing for individual configuration optimization. The system integrates region of interest (ROI) cropping, fundus-specific preprocessing, parallel YOLO detectors, tiling (for MA, HE, EX), per-lesion ensembling of five cross-validation folds, and an inter-lesion suppression step based on physical size and clinical priority. Trained on IDRiD with stratified five-fold cross-validation, PRISM-DR achieved a test mAP50 of 0.527 and an F1 score of 0.529, with hard exudates showing the highest AP50 at 0.561. The models demonstrated good transferability to e-ophtha (0.483 mAP50) without fine-tuning when imaging scales were similar to IDRiD, but performance degraded on DDR (0.159 mAP50) and TJDR (0.267 mAP50) due to greater variations in field of view and resolution.

Key takeaway

For Machine Learning Engineers developing medical image analysis systems, if you are building detectors for multiple, diverse lesion types, consider adopting a per-lesion model approach. This strategy, demonstrated by PRISM-DR, allows for tailored optimization of architecture, augmentation, and tiling for each lesion, improving overall detection accuracy, especially for small, hard-to-detect lesions like microaneurysms. You should also integrate clinical knowledge into post-processing steps, such as inter-lesion suppression, to resolve overlaps effectively.

Key insights

Treating each diabetic retinopathy lesion as an independent detection problem significantly improves accuracy over single multi-class models.

Principles

• Lesion-specific models outperform shared models for diverse targets.
• Optimize architecture and augmentation per lesion characteristics.
• Clinical knowledge improves model fusion and conflict resolution.

Method

PRISM-DR uses parallel YOLO detectors, each optimized for a specific DR lesion, with ROI cropping, preprocessing, tiling, Bayesian-tuned augmentation, five-fold ensembling, and clinical-priority-based inter-lesion suppression.

In practice

• Implement separate models for distinct object classes.
• Tune data augmentation based on object morphology.
• Use physical/clinical rules for resolving overlapping detections.

Topics

• Diabetic Retinopathy
• Object Detection
• YOLO Models
• Medical Image Analysis
• Lesion Detection
• Fundus Imaging

Code references

• zubeyrozeren/PRISM-DR

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• PRISM-DR: Per-lesion Retinal Inference with Specialist Models for Diabetic Retinopathy — Specialized single-class detectors for each diabetic retinopathy lesion type outperform general multi-class models by addressing lesion diversity.
• A Dual Edge Spatial Jacobian Image Graph for Interpretable Diabetic Retinopathy Grading — Interpretable DR grading benefits from graph representations linking spatial lesion-vessel geometry and embedding-biomarker sensitivity.
• Region-Grounded Vision-Language Learning for Detection-Guided Mammographic Lesion Classification — Region-grounded vision-language learning improves mammographic lesion classification by mirroring radiologists' localized diagnostic approach.
• MIRAGE: Multi-scale Lesion-Informed Representation with Auxiliary Guidance for MRI Contrast Enhancement — MIRAGE enhances MRI contrast inference by integrating lesion-aware supervision with global reconstruction, improving tumor localization.
• HSQ-VLM: A Novel Spatially-Constrained Quadrant Segmentation VLM Model for Explainability in Diabetic Retinopathy — HSQ-VLM provides explainable DR diagnostics by segmenting fundus images with fovea-centered anatomical precision.
• Towards Interpretable Foundation Models for Retinal Fundus Images — Foundation models for medical imaging can achieve high performance and inherent interpretability simultaneously.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.