Development of an Integrated Radiomics and Generative Artificial Intelligence Model for Differentiating Jaw Cysts and Tumors Using Cone-Beam Computed Tomography

Background: Accurate differentiation of jaw cysts and tumors using Cone-Beam Computed Tomography (CBCT) remains a diagnostic challenge due to overlapping radiographic characteristics. Recent advances in radiomics and generative artificial intelligence (AI) offer new opportunities for improving automated lesion classification and diagnostic support.

Objective: To develop and evaluate an integrated radiomics and generative AI model for differentiating jaw cysts and tumors using CBCT imaging.

Methods: A retrospective dataset comprising 100 histopathologically confirmed jaw lesions, including 50 cysts and 50 tumors, was collected from CBCT examinations. Lesion regions of interest were segmented, and radiomic features were extracted following image preprocessing and standardization. To address dataset limitations and improve model robustness, a generative AI framework was employed to generate five synthetic variations from each original lesion sample, resulting in an augmented dataset of 600 lesion volumes. Quantitative radiomic features encompassing first-order, shape-based, and texture characteristics were utilized for model development. Machine learning classification was performed using the combined original and synthetic datasets. Model performance was evaluated using accuracy, sensitivity, specificity, precision, F1-score, and area under the receiver operating characteristic curve (AUC).

Results: The integrated radiomics and generative AI model demonstrated effective differentiation between jaw cysts and tumors. Dataset augmentation substantially increased sample diversity and improved model generalizability. The developed framework showed promising classification performance, indicating the potential utility of combining radiomic feature analysis with generative AI-based data augmentation for CBCT-based lesion characterization. Texture-derived radiomic features emerged as important contributors to lesion discrimination, while generative AI enhanced dataset diversity and supported model robustness.

Conclusion: The proposed integrated radiomics and generative AI framework represents a promising approach for the CBCT-based differentiation of jaw cysts and tumors. By combining quantitative radiomic analysis with synthetic data generation, the model addresses challenges related to limited datasets and variability in lesion presentation. This approach has the potential to serve as an effective diagnostic support tool in oral and maxillofacial radiology and provides a foundation for future developments in explainable and automated AI-assisted lesion diagnosis.