Cone-beam computed tomography radiomics for differentiating clinicoradiologic labels of osteosclerosis, cemento-osseous dysplasia and odontoma: an internal model-development study
Dentomaxillofacial Radiology · 2026-10-24
DOI: 10.1093/dmfr/twag067 ↗Abstract
Abstract Objectives To develop and internally validate cone-beam CT radiomics for recorded clinicoradiologic classification of osteosclerosis, cemento-osseous dysplasia, and odontoma, with secondary same-case 3D convolutional neural-network benchmarking. Methods This retrospective single-center study included 126 lesions from 126 patients. From manual 3D segmentations, 107 original radiomics features were extracted. A class-weighted least absolute shrinkage and selection operator one-vs-rest logistic-regression model was evaluated using 100 repeats of stratified 5-fold held-out resampling. No independent histopathology report was retrievable in the retained study archive. A secondary strict 124-lesion 3D analysis used 0.30-mm isotropic variable-size lesion-centerd inputs, at least 5 mm of physical context and nested architecture/epoch selection; locked radiomics was evaluated on the same cases and outer folds. Results Primary radiomics achieved accuracy 0.825, balanced accuracy 0.842, macro F1-score 0.834, and macro area under the curve 0.932. In the matched 124-lesion analysis, macro area under the curve was 0.935 (95% bias-corrected and accelerated CI 0.889-0.963) for radiomics and 0.601 (0.516-0.682) for the nested 3D network. The network showed marked odontoma-class collapse (sensitivity 1/22, 0.045). Conclusions Radiomics showed promising internal discrimination of the recorded labels, whereas the secondary 3D benchmark was unstable for odontoma in this small cohort. These findings do not establish diagnostic accuracy against histopathology or general algorithmic superiority; external validation, standardized imaging, and stronger reference standards are required. Advances in knowledge This study extends cone-beam CT radiomics to radiopaque/mineralized jaw-lesion recorded labels and reports a same-case nested 3D benchmark with explicit disclosure of class-specific failure.
EUVIMED