Panoramic dental X-rays are common in screening, but automated detection models often struggle when moved from one dataset or software environment to another because class definitions differ. A new arXiv preprint introduces PanoDent-KD, a pipeline that uses knowledge distillation to turn a generalist detector into a specialist model for panoramic radiographs.
The system is built around a 1280-pixel YOLO26m model and links a 12-class source detector to a downstream specialist detector. Knowledge distillation transfers what the generalist has learned, allowing the specialist to keep consistent performance even when the target dataset uses different categories or labels. This cross-dataset behaviour is one of the paper's main concerns.
The authors also emphasise auditability for embedded deployment. The goal is a model that can run locally within clinical software and produce detections that clinicians can trace. PanoDent-KD is therefore designed as a pipeline that connects model outputs to deployment tools, rather than as a standalone research model.