Detection v2.0.0

YOLOv26s end2end NMS, 6 clases. Input 640x640.

DIRD Models v1.0.1

DIRD Models v1.0.1

First tagged release of pretrained YOLO models for diabetic retinopathy detection.

Included

• detection-v1.0.1.onnx — YOLOv11 detection weights (opset 14, browser-ready)
• detection-v1.0.1.json — model metadata (11 classes, 7 active)
• MODEL_CONVERTER.py — ONNX → browser-compatible converter

Detection model

• Architecture: YOLOv11
• Input: 640 × 640
• Classes: 11 (7 currently detected, 4 reserved)
• Global mAP50: 0.826 · mAP50-95: 0.450
• Status: REQUIRES_IMPROVEMENT

Notes

• Segmentation model (blood vessels, exudates, hemorrhages, neovascularization) still in development — not included in this release.
• Licensed under MIT.
• Research and assistive tooling only. Not a medical device.