An Opportunistic Screening Framework for Coronary Artery Calcium Assessment on Radiotherapy Datasets
Abstract
Purpose
Coronary artery (CA) calcifications (CACs) connect upstream cardiovascular risks with downstream radiotherapy planning decisions, yet these powerful imaging biomarkers are underutilized in radiation oncology. We developed and validated an opportunistic screening deep learning (DL) pipeline that extracts total and vessel-specific CACs to localize high-risk areas from routine CT scans, enabling comprehensive cardiac risk assessment and mitigation in radiotherapy throughout the cancer care continuum.
Methods
A cohort of 365 thoracic CTs (154 diagnostic calcium scoring, 67 lung cancer screening, 101 radiotherapy planning, 43 PET-CT) were segmented with CACs, whole-heart, and descending/ascending aorta. CA “habitats” (i.e., probabilistic high-risk regions) were defined and used to post-process model predictions. A novel nnU-Net with self-distillation was trained for heart localization, whole-heart/aorta segmentation, and vessel-specific CAC segmentation, with a 70%/10%/20% split for training/validation/testing. Twenty scans from each CT modality were tested. Model performance was evaluated with Dice Similarity Coefficient (DSC) and detection accuracy.
Results
Across modalities, for detection of total CACs, average DSC and detection accuracy were 89.7%±19.7% and 96.3%, respectively. Vessel-specific performance varied, ranging from 60.4-87.5% for DSC and 83.8-98.8% for detection accuracy across individual CAs. DSC and accuracy were lowest for the LMCA and RCA, respectively, while DSC/accuracy were highest for the LADA. For modality-specific performance for total CACs, DSC was lowest for lung cancer screening CT (85.2%) and highest for diagnostic calcium scoring CT (96.9%). For vessel-specific CACs, the average DSC was lowest for PET-CT (75.7%) while simulation CT performed best (88.7%). For accuracy, the lung cancer screening CT (87.5%) had the worst performance compared to PET-CT (97.5%).
Conclusion
Our pipeline successfully segments CACs on diagnostic and routine radiotherapy datasets. Agreement between predicted and ground-truth CACs was strong, achieving high DSC and accuracy for total and vessel-specific CACs across multiple CT modalities, offering promise for use in cardiovascular risk assessment in radiotherapy.