Deep learning and radiomics-based vascular calcification characterization in dental cone beam computed tomography as a predictive tool for cardiovascular disease: a proof-of-concept study

Objectives: This study evaluated an automated deep learning method for detecting calcifications in the extracranial and intracranial carotid arteries and vertebral arteries in cone beam computed tomography (CBCT) scans. Additionally, a model utilizing CBCT-derived radiomics imaging biomarkers was evaluated to predict the cardiovascular diseases (CVD) of stroke and heart attack. Methods: Models were trained using the nn-UNet architecture to identify three locations of arterial calcifications: extracranial carotid calcification (ECC), intracranial carotid calcification (ICC), and vertebral artery calcification (VAC). In total, 148 scans were used for model training and validation. Radiomics features extracted from 135 calcification regions were used to characterize arterial calcifications for predicting CVD. Results: The models demonstrated acceptable performance for detecting regions of calcification for ECC and ICC with bounding box accuracies of 0.71 ± 0.06 and 0.78 ± 0.12 respectively, although VAC performance was lower at 0.53 ± 0.17. Combining clinical data with radiomics for ICC improved stroke predictions, yielding an area under the curve derived from receiver operating characteristic analysis (AUC-ROC) of 0.94 ± 0.09, and combining data for ECC and ICC improved heart attack predictions, with AUC-ROC values of 0.88 ± 0.04 and 0.84 ± 0.16, respectively. Conclusion: Automated, quantifiable methods have potential for detecting ECC and ICC and predicting the incidence of cardiovascular disease based on arterial calcification detection in dental CBCT scans.