ARC-CT: Anatomy-Routed Contrastive Vision-Language Learning for 3D Chest CT
ARC-CT reports a 0.86 macro AUC across 18 chest CT abnormalities without manual labels or boxes.
The method routes learning through organ masks so small, localized findings are less likely to vanish in a whole-volume embedding. It also softens contrastive negatives when scans share report-derived labels, reducing penalties between clinically similar cases. An organ-level loss aligns mask-pooled CT features with organ-specific report text extracted offline by an LLM. The paper says ARC-CT beats comparable efficient baselines and several larger transformer models. ArXiv · AI/CL/LG's note
The method routes learning through organ masks so small, localized findings are less likely to vanish in a whole-volume embedding. It also softens contrastive negatives when scans share report-derived labels, reducing penalties between clinically similar cases. An organ-level loss aligns mask-pooled CT features with organ-specific report text extracted offline by an LLM. The paper says ARC-CT beats comparable efficient baselines and several larger transformer models. ArXiv · AI/CL/LG's note
score 4