A modality-agnostic coronary artery habitat model for cardiac sparing in radiotherapy.
Authors
Affiliations (4)
Affiliations (4)
- Department of Medical Physics, University of Wisconsin-Madison, Madison, Wisconsin, USA.
- Department of Radiation Medicine, University of Wisconsin-Madison, Madison, Wisconsin, USA.
- Department of Computer Science, Wayne State University, Detroit, Michigan, USA.
- Department of Radiology, University of Wisconsin-Madison, Madison, Wisconsin, USA.
Abstract
Emerging evidence suggests that the risk of cardiotoxicity increases with increased radiation dose to coronary arteries (CAs). However, robust tools to evaluate this increased burden for cancer patients are not available due to current limitations in imaging for radiotherapy treatment planning. We have developed a novel, statistical methodology to define coronary artery "habitats", or probabilistic, high-risk regions to facilitate enhanced cardiac sparing in radiation therapy. Our modality-agnostic approach leverages state-of-the-art deep learning models and high-resolution coronary CT angiography images for fully automatic prediction of CA habitats on varied radiation therapy datasets. Incorporating imaging data from two separate institutions and two cardiac segmentation challenge datasets, 182 coronary CT angiography (CCTA) and 60 thoracic radiotherapy volumes (simulation CT, 0.35T MRI) were evaluated. Images were labeled using nnU-Net for CAs and whole-heart and verified by a cardiovascular radiologist. Habitats were predicted by (1) template matching to select best-fit CCTA, (2) deformably warping CAs from CCTA to each input image via deep learning models, (3) estimating each CA probability density function, and 4) final habitat derivation. Model parameters were optimized using a subset of 30 CCTA. Full-branch (i.e., derived from all CA segments on CCTA) and main-branch (i.e., derived from main CA branches only) habitats were predicted. Final habitats were evaluated on 60 CCTA, 34 simulation CT, and 26 MR-Linac volumes via inclusion ratio and Hausdorff distance. Additional quantitative evaluation was performed with CAs with added planning organ-at-risk volumes (PRVs). Dosimetric correlation was evaluated between predicted habitats and CAs with added PRVs for 51 clinical treatment plans. To demonstrate applicability of our model to radiation treatment planning, habitat-spared treatment plans were retrospectively re-optimized for three patients and compared to re-optimized plans sparing the whole-heart. Predicted full-branch and main-branch habitats contained 92.4 ± 8.9%-97.6 ± 4.5% and 89.7 ± 11.7%-98.2 ± 2.6% of CAs, respectively. Considering CAs with added PRVs, inclusion within habitats was lower compared to CAs alone, ranging from 83.1 ± 11.1%-92.3 ± 8.0% and 75.3 ± 15.0%-87.6 ± 9.4% for the CT-SIM and MR-linac datasets, respectively. When CAs were not contained within habitats, Hausdorff distance between main-branch habitats and CAs ranged from 0.7 ± 1.6-4.4 ± 4.5 mm across all modalities. Predicted habitat volume was 2.1%-17.4% of the whole heart, on average. End-to-end prediction time was <4 min across all modalities. D<sub>mean</sub> and D<sub>0.03cc</sub> for habitats and CAs with an added PRV were strongly correlated (r >0.90) and in agreement (ICC >0.86) across 51 clinical plans for all CAs. CA D<sub>0.03cc</sub> was reduced by up to 21.1 Gy in habitat-spared plans, while D<sub>mean</sub> was reduced by up to 18.5 Gy. Habitats contained on average >90% of CAs and 75%-92% of CAs with PRVs across evaluated modalities with strong agreement in dosimetric indices between manually delineated CAs and predicted main-branch habitats. Substantial reductions in CA dosimetric endpoints were possible using habitat-spared treatment plans while maintaining plan quality. Habitats serve as a clinically feasible alternative for CA delineation and enable treatment planning strategies to offer improved cardioprotection.