Multi-label classification of temporomandibular joint osteoarthritic changes on CBCT: An anatomically guided artificial intelligence framework

OBJECTIVE: To develop and evaluate an anatomically guided artificial intelligence framework for joint-level detection and multi-label characterization of temporomandibular joint (TMJ) osteoarthritic changes on cone-beam computed tomography (CBCT). METHODS: CBCT examinations from 71 patients (142 TMJs) were assessed independently by two expert readers for erosion, osteophyte formation, subchondral sclerosis, subchondral cysts, joint-space narrowing, and condylar flattening. Discordant cases underwent third-reader adjudication to establish the reference standard. A 3-dimensional ResNet18 model was trained using patient-level training, validation, and independent test partitions. Joint-level OA-associated features were evaluated. Complete TMJ region-of-interest (ROI) and segmentation-masked condyle inputs were compared. Condylar flattening was analyzed separately. RESULTS: Inter-reader agreement varied by feature, ranging from slight agreement for flattening (κ = -0.01) and joint-space narrowing (κ = 0.15) to substantial agreement for osteophyte formation (κ = 0.63). Automated condyle segmentation achieved a mean Dice similarity coefficient of 0.91 ± 0.042. The joint-level model achieved an area under the receiver operating characteristic curve (AUC) of 0.9325, with sensitivity of 0.8470 and specificity of 0.9081. Feature-level AUCs ranged from 0.6756 to 0.9515. Segmentation-masked condyle input yielded higher classification of erosion (AUC, 0.8685), osteophyte formation (0.9090), subchondral sclerosis (0.9134), and subchondral cysts (0.9515), whereas complete TMJ ROI input performed better for joint-space narrowing (0.8132). CONCLUSIONS: This framework extends CBCT assessment beyond binary detection of TMJ OA-associated changes by providing feature-level characterization. Anatomically guided masking showed higher AUCs for abnormalities intrinsic to the mandibular condyle, whereas joint-space narrowing performed better with broader joint context. These observations require confirmation in larger datasets.