Deep learning diagnosis of TMJ osteoarthritis using MRI ensembles and CBCT–MRI fusion

Abstract

We compared single-modality models, intra-modality ensembles, and selective cone-beam computed tomography (CBCT)–magnetic resonance imaging (MRI) fusion for diagnosing temporomandibular joint osteoarthritis (TMJ-OA). This single-center retrospective cohort included 1598 patients (1118 females and 480 males; mean age ± SD, 37.88 ± 19.19 years), with patient-level splitting into training, validation, and test sets at a 60:20:20 ratio. Joint-level labels were defined using clinical and imaging criteria. Single models were trained using right/left CBCT views (sagittal and coronal) and MRI sequences (T2-weighted and proton density [PD]). We evaluated intra-modality ensembles (CBCT sagittal + coronal and MRI T2 + PD) and selective CBCT–MRI late fusion. DeLong tests were used to compare model performance, and Grad-CAM was used to assess explainability. In the full dataset, single CBCT models achieved area under the receiver operating characteristic curve (AUROC) values of 0.771 for sagittal views (95% confidence interval [CI], 0.732–0.807) and 0.787 for coronal views (95% CI 0.748–0.821). Single MRI models achieved AUROCs of 0.835 for T2-weighted images (95% CI 0.768–0.893) and 0.827 for PD images (95% CI 0.759–0.888), while the single combined-sequence MRI model achieved an AUROC of 0.881 (95% CI 0.836–0.913). Among the ensemble models, the MRI T2 + PD ensemble showed the highest performance (AUROC, 0.916; 95% CI 0.823–0.977), followed by the CBCT sagittal + coronal ensemble (AUROC, 0.832; 95% CI 0.781–0.881). Selective CBCT–MRI fusion improved performance compared with CBCT alone but did not surpass the best MRI ensemble. Findings were consistent across sex and age strata. Overall, MRI sequence aggregation provides a strong benchmark for TMJ-OA discrimination, whereas selective CBCT–MRI fusion may offer incremental value over CBCT-based models in selected pairings.