ENTERPRISE AI ANALYSIS

The study utilized 56 matched pairs of RVG and CBCT scans, analyzing 109 regions of interest (ROIs) categorized as onlyCBCT (lesions present on CBCT but not RVG) or noLesion (true negative). Radiomic texture features were extracted from circular ROIs using PyRadiomics. A logistic regression classifier was trained using nested cross-validation, with grouping at the examination level to prevent data leakage. Critically, the study found that smaller ROIs (20–40 pixels), particularly a 40-pixel radius, yielded higher AUCs and accuracy, indicating optimal capture of local texture information without dilution from irrelevant structures.

For a 40-pixel ROI radius, 44 radiomic texture features showed statistically significant differences between onlyCBCT and noLesion ROIs after FDR correction, primarily exhibiting small to medium effect sizes. These features were dominated by gray-level texture measures from the GLCM, GLDM, GLRLM, GLSZM, and NGTDM. Notably, several second-order GLCM features (e.g., Imc2, Cluster Prominence, Sum Squares) showed higher median values in the onlyCBCT group, while gray-level non-uniformity metrics tended to be lower. These patterns suggest systematic, subtle differences in gray-level texture reflecting early pathological changes.

This research highlights radiomics' potential to address the significant clinical challenge of detecting CBCT-confirmed, visually occult periapical lesions—a scenario where conventional visual interpretation falls short. The moderate performance reflects the inherent difficulty of the task, focusing on sub-visual pathology rather than overtly visible lesions. While the single-center, retrospective design and relatively small sample size are limitations, the study provides a strong rationale for larger, multicenter prospective studies to validate and further develop radiomics for early periapical lesion detection, particularly when CBCT is not routinely available.