Starting a project to detect candidate stenoses in x-ray coronary angiograms. The goal is high recall, false positives are fine as long as it doesn't flag half the image.

I already tried YOLO on a couple of in-house datasets earlier. It looked fine in-domain and fell apart on the other dataset, and a big chunk of that turned out to be annotation format. One set had fixed-size boxes around a point, the other had boxes fitted to the lesion. Same disease, basically two different tasks.

I also have vessel segmentation and centerlines, so my first instinct is a radius profile along the vessel to flag narrowing, then maybe a network to filter the candidates. Has anyone actually done something like that, or is there a smarter way in?

The part that bugs me most is evaluation. The labels come from doctors and they obviously don't mark every narrowing, so any recall I compute against them is probably too optimistic.…