Coronary segmentation + FP/FN adjudication on 10 training images · vision LLM via OpenRouter · our top model: super26_vesselnet_DeepLabV3Plus_efficientnet-b4
Ground truth (reference) Our model picks False positive (ours) False negative (ours) LLM segmentation / TRUE
Read-only evaluation page · GT = frozen ARCADE mask (v0) for the training split · model output generated on
· no inference or annotation happens on this page.