Perilesional and Contextual Radiomic Features Improve Differentiation of Odontogenic Cysts on Panoramic Radiographs

Background: Radiographic differentiation of odontogenic cysts remains challenging due to overlapping imaging features and limited specificity of conventional assessment. Radiomics offers a quantitative approach to extract imaging biomarkers that may improve non-invasive lesion characterization. Accurate differentiation among cyst subtypes is clinically relevant, as different odontogenic cysts may warrant different management and follow-up strategies. Methods: This study included 63 patients with histopathologically confirmed odontogenic cysts (odontogenic keratocyst [OKC], n = 33; radicular cyst [RC], n = 12; dentigerous cyst [DC], n = 10; odontogenic cyst, not otherwise specified [OC-NOS], n = 8). Panoramic radiographs were manually segmented to define cyst and anatomical regions of interest. Radiomic features were extracted using three groups: agnostic texture features, morphological/intensity descriptors, and perilesional features. Classification was performed using logistic regression with leave-one-out cross-validation. Pairwise and four-class classification analyses were conducted, and feature importance was evaluated. Results: Radiomic features demonstrated significant differences between cyst and anatomical regions, with approximately 60% of features remaining significant after correction. Pairwise classification performance varied across lesion types, achieving the highest balanced accuracy for DC vs. OC-NOS (0.838) and OKC vs. OC-NOS (0.752), while comparisons involving OKC vs. RC (0.602) and OKC vs. DC (0.644) were less discriminative. Texture-based features, particularly NGTDM-derived metrics, were among the most informative. The integration of morphological and perilesional features improved performance: four-class balanced accuracy increased from 0.416 with texture-only features to 0.487 with the combined feature set (p = 0.002). Conclusions: Radiomics-based analysis of panoramic radiographs enables moderate but preliminary differentiation of odontogenic cysts, with performance comparable to prior studies. The inclusion of perilesional and spatial features enhances classification, highlighting the importance of contextual information. Despite these advances, substantial overlap between cyst types persists, underscoring the need for multimodal diagnostic approaches and further validation in larger cohorts.