Zero-shot segmentation of coronary arteries from conventional angiograms · 10 random ARCADE train images (seed 42) · same image + same neutral prompt for every model · models (meta/muse-spark-1.2 via Meta Spark, the rest via OpenRouter)
⚠️ Pilot (n=10) — per reviewers: 10 paired images cannot rank models; CIs overlap. Means with 95% paired-bootstrap CIs shown; treat as an orientation, not a verdict. "Zero-shot" = no in-context help, no fine-tuning; ARCADE is public since 2022, so flagship models may have seen similar images during pretraining.
🟢 green = ground truth · 🟡 amber = model prediction — translucent fill (α≈0.35) with a solid outline, so the vessels stay visible underneath