TMJ-OD3D: A cone-beam CT Dataset for 3D Detection of Temporomandibular Joint Osseous Abnormalities

The dataset contains 1,043 de-identified CBCT examinations from 1,043 patients, corresponding to 2,086 left and right TMJs. For each patient, the released data include a de-identified DICOM image series containing TMJ-related axial slices, side-specific JSON annotation files for the left and right joints, and a patient-level metadata record including age, sex, and labels.

The annotations provide joint-side-specific labels for TMJ osseous abnormalities and lesion localization using a structured 3D bounding-box representation. Each 3D box is encoded by a two-dimensional rectangular box drawn on a representative DICOM slice and the corresponding axial slice range over which the abnormality is visible. All annotations were generated and reviewed by oral and maxillofacial surgeons through an expert-consensus workflow.