Coronary angiogram stenosis detection: annotation format mismatch broke YOLO cross-dataset

Future-Structure-296 · reddit · 2026-10-24

A developer building a high-recall stenosis candidate detector for X-ray coronary angiograms shares lessons and open questions.

YOLO worked in-domain but fell apart on a second dataset, largely due to annotation format mismatch: one set used fixed-size boxes around points, the other lesion-fitted boxes — effectively two different tasks for the same disease.
With vessel segmentation and centerlines available, the plan is a radius-profile along vessels to flag narrowing, then a network to filter candidates.
The hardest problem is evaluation: physician labels miss many narrowings, making computed recall overly optimistic — how to get honest numbers with incomplete ground truth?
Solicits papers, repos, and public datasets beyond ARCADE.

More from Research