Nawaf Almutairi

تخصص الذكاء الاصطناعي · دفعة 2026

ANCHOR: Anatomy-Grounded Narrative Report Generation from Maxillofacial Cone-Beam CT

A CBCT scan is a 3D X-ray of the jaw, and today a surgeon writes its report by hand, one scan at a time. We're automating that, with one deliberate constraint: the AI that writes the sentences never sees the scan. A first stage finds and labels the anatomy (each tooth, the nerve canals, the sinuses, the jawbone), a second stage turns those shapes into a structured list of facts, and only then does a language model turn that list into prose. The constraint is the whole point. A model that reads the image directly can produce fluent, confident sentences about findings that aren't there, which in a clinical report is worse than saying nothing at all. Ours can only describe what was actually labelled, so every sentence traces back to something measurable in the scan. That matters most for teeth: there are 32 nearly identical ones, and "crown on tooth 36" versus "crown on tooth 46" is the other side of the mouth and a different treatment

اقرأ التقرير النهائي (PDF)