Dentin-Normalized CBCT Radiomics within a Periapical Lesion Differential Diagnosis System for Differentiating Granulomas and Radicular Cysts.
Authors
Affiliations (5)
Affiliations (5)
- H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology, Atlanta, GA, USA.
- School of Computing and Augmented Intelligence Arizona State University, Tempe, AZ.
- Department of Oral Medicine, University of Pennsylvania, Philadelphia, PA.
- MS Technologies Corporation, Rockville, MD.
- Department of Endodontics, University of Pennsylvania, Philadelphia, PA. Electronic address: [email protected].
Abstract
Periapical granulomas and radicular cysts cannot be reliably differentiated by clinical examination or conventional radiography. Histopathological examination remains the gold standard for definitive diagnosis. Because biopsy is invasive, cone-beam computed tomography (CBCT)-based radiomics approaches have been explored as noninvasive alternatives. However, their performance is affected by intensity variability inherent to CBCT imaging. The purpose of this study was to evaluate a dentin-based intensity normalization approach to improve the robustness of radiomics-based classification of periapical lesions. CBCT scans from 140 patients with histopathologically confirmed periapical lesions (100 granulomas, 40 cysts) were retrospectively analyzed. Lesion-adjacent root dentin was approximated as an internal reference region and used for affine intensity normalization. Radiomic features were extracted from segmented lesion volumes before and after normalization within the Periapical Lesion Differential Diagnosis System (PLDDS). Models were evaluated using morphology-only, first-order/texture-only, and full-radiomics feature settings with eXtreme Gradient Boosting and logistic regression classifiers. Performance was assessed using stratified 5-fold cross-validation, AUC (area under the curve), complementary classification metrics, confidence intervals, and statistical comparison of AUCs. Dentin-normalized full radiomics with XGBoost achieved the highest overall performance, with an AUC of 0.78. Dentin normalization significantly improved AUC compared with raw full radiomics in the XGBoost model (0.68 vs 0.78; P<.05). AUC also improved after normalization in the first-order/texture setting for both classifiers, suggesting that normalization primarily benefited gray-value-dependent features. Dentin-based intensity normalization improved the performance and stability of CBCT radiomics-based models for differentiating periapical granulomas and radicular cysts. The higher AUC of dentin-normalized full radiomics compared with morphology-only features suggest that lesion-adjacent dentin may serve as a practical internal reference for stabilizing gray-value-dependent radiomic features and that first-order and texture features provided complementary information beyond lesion morphology. Further validation across different CBCT systems and acquisition protocols is required before this approach may be considered for clinical implementation.