254 - A Novel Coronary Artery Habitat and Calcium Detection Model for Cardiac Sparing in Radiotherapy

Purpose/Objective(s): Emerging evidence suggests that the risk of cardiotoxicity increases with increased radiation dose to coronary artery (CA) and the presence of CA calcifications (CACs). However, CA delineation and CAC assessment are challenged by radiation therapy imaging limitations. We have developed a novel end-to-end pipeline that defines high risk CA regions (“habitats”) coupled with CAC segmentation by leveraging high-resolution coronary CT angiography (CCTA) and deep learning (DL) to support enhanced cardiac sparing.

Materials/Methods: 411 patients with paired CCTA and non-contrast CT were retrospectively reviewed. CAs were labeled for 182 contrast-enhanced CCTAs. CACs were segmented on 128 non-contrast CTs. For a given image, CA habitats—probabilistic regions where CAs are expected—are generated via a three-step process: (1) identifying 30 best-fit CCTA via whole heart template matching, (2) deformably warping CA labels from each CCTA using uniGradICON DL registration and (3) defining final habitats by Euclidean clustering after outlier removal (>1 nearest neighbor distance). We then trained a DL pipeline, incorporating an nnU-Net model with self-distillation, to perform CAC segmentation on non-contrast CT (training/validation/testing = 107/5/16), including predicted CA habitats as model inputs. Our nnU-Net was evaluated via the Dice similarity coefficient (DSC) and Agatston score risk accuracy. To determine generalizability to radiation therapy, our pipeline was applied to simulation CT (SIM-CT) data from a cohort of 31 lung cancer patients (12 for model fine-tuning and 19 for evaluation). CA habitats and CACs were predicted on each SIM-CT and evaluated via CA-habitat inclusion (i.e. percentage of each CA contained in its habitat) and CAC detection accuracy compared to annotations by a cardiovascular radiologist.

Results: End-to-end prediction time of CACs and CA habitats was ~8 minutes per patient. Our nnU-Net achieved an artery-specific CAC DSC of 78.0±31.5% to 96.1±5.3% on non-contrast CTs. Predicted cardiovascular risk accuracy was 93.8%. On SIM-CT, predicted CA habitats contained 88.6±21.4% to 95.5±11.4% of CAs, averaging 92.8%, and CAC detection accuracy was 63.2% to 78.9% across CAs, averaging 71.1%. Average CA habitat size ranged from 1.8±0.2% to 20.2±2.7% of the heart volume.

Conclusion: We successfully trained our flexible DL pipeline on diagnostic and RT cohorts. We achieved excellent CAC risk accuracy on non-contrast CT and promising CAC detection and habitat identification on SIM-CT. Future work includes expanding cohorts for improving model performance and incorporating CA habitats and CACs for enhanced cardiac-spared treatment planning.