Segmentation baselines exceeded on ARCADE

Vessel segmentation reached 0.799 Dice and 0.807 clDice on the ARCADE test split. Stenosis segmentation reached 0.61 Dice, against a published best of 0.40 for detection.

The vessel figure needs reading carefully. ARCADE's own inter-observer Dice averages 0.81, so the model now agrees with the annotation about as well as two annotators agree with each other. That is a ceiling, not a headline — three independently trained models all land in the same place, and no loss function or architecture change we tried moved it.

We report ARCADE for comparability with the published literature, not as an achievement. The number that decides whether any of this works is the agreement study, not this one.