Multi-Level Automated Classification of Temporomandibular Joint Degenerative Joint Disease on CBCT Using 3D Volumetric and 2D Slice-Based Approaches.

BACKGROUND

Temporomandibular joint degenerative joint disease (TMJ DJD) causes progressive osseous changes of the mandibular condyle and articular fossa. Although cone-beam computed tomography (CBCT) enables detailed assessment, manual interpretation is time-consuming and subject to inter-observer variability.

OBJECTIVES

To develop and validate a multilevel AI framework for CBCT-based TMJ DJD assessment using three-dimensional classification, quantitative volumetric analysis, and two-dimensional slice-based detection.

METHODS

Full-field-of-view CBCT scans from 222 patients were used for three-dimensional analysis. A 3D V-Net classified condyles into Normal, Osteophyte, Erosion, Flattening, and Severe Degeneration categories. Condylar volume was quantitatively compared between TMJ DJD and non-TMJ DJD groups and among degenerative subgroups. For slice-based analysis, 727 temporomandibular joints from 416 patients were retrospectively evaluated. Automated sagittal slice extraction based on the condylion landmark generated 5787 condylar and 666 articular fossa slices. YOLOv10 was used for multi-lesion detection of six condylar and three fossa categories.

RESULTS

The 3D V-Net achieved an overall sensitivity of 79.47%. Condylar volume differed significantly between TMJ DJD and non-TMJ DJD groups (ANOVA, p = 0.025), whereas differences among the TMJ DJD subgroups were not statistically significant. For condylar slice-based detection, YOLOv10 achieved 94.0% sensitivity, 82.8% precision, 88.0% F1-score, and 91.3% AP@0.50. For articular fossa detection, YOLOv10 achieved 92.41% sensitivity, 83.91% precision, 87.95% F1-score, and 82.01% AP@0.50.

CONCLUSIONS

Three-dimensional AI enabled overall TMJ DJD classification but showed limited discrimination of specific subgroups, particularly erosion. Volumetric analysis differentiated DJD from non-DJD cases but not among DJD subgroups. In contrast, slice-based YOLOv10 effectively detected and labelled specific condylar and fossa degenerative changes, providing more detailed lesion-level characterization.