Contribution of CBCT-Derived Radiomics and Machine Learning to the Characterization of Osteolytic Lesions of the Jaws.
TL;DR
A multivariate machine learning approach may identify complex signatures with moderate discriminative performance in osteolytic jaw lesions to aid differential diagnosis between developmental and inflammatory Odontogenic Cysts.
Abstract
Objective
Osteolytic jaw lesions are heterogeneous and share radiographic features. Radiomics from CBCT may provide quantitative biomarkers to aid differential diagnosis between developmental and inflammatory Odontogenic Cysts.
Methods
A retrospective radiomic analysis was performed on CBCT scans of osteolytic jaw lesions from 52 patients; after quality control, 44 lesions were included (28 developmental and 16 inflammatory cysts). Lesions were segmented in 3D Slicer, and radiomic features were extracted with PyRadiomics. Unsupervised analyses used K-means and principal component analysis (PCA). Group differences were tested with ANOVA. For supervised classification, Logistic Regression, Random Forest, SVM, Multilayer Perceptron, and a stacking model were trained using PCA-based dimensionality reduction, mutual-information feature selection, SMOTE oversampling, and repeated stratified 5-fold cross-validation.
Results
One thousand four hundred and thirty-seven features per lesion were extracted; univariate analysis identified 11 radiomic features showing differences between the two diagnostic groups (p < 0.05). However, none remained statistically significant after false discovery rate (FDR) correction, indicating the exploratory nature of these findings. Random Forest showed the best performance (AUC 0.76, accuracy 0.74, specificity 0.81, sensitivity 0.61). Unsupervised clustering did not reveal separated groups.
Conclusions
While individual radiomic features did not demonstrate significance, a multivariate machine learning approach may identify complex signatures with moderate discriminative performance.