Author: Blessing Akinro, Soumyanil Banerjee, Ming Dong, Carri K. Glide-Hurst, Prashant Nagpal, Chase Ruff, Nicholas R. Summerfield, Timothy P. Szczykutowicz 👨🔬
Affiliation: Departments of Human Oncology and Medical Physics, University of Wisconsin-Madison, Departments of Radiology and Medical Physics, University Wisconsin-Madison, Department of Radiology, University of Wisconsin-Madison, Department of Computer Science, Wayne State University, Department of Human Oncology 🌍
Purpose: Radiation dose to coronary arteries (CAs) during thoracic radiotherapy (RT) is linked to cardiotoxicity. However, precise CA delineation for avoidance is limited by image quality and CA complexity. We propose a novel, deep learning-based CA habitat model for routine treatment planning to enable cardiac sparing for any patient at risk for cardiotoxicity.
Methods: A state-of-the-art nnU-Net pipeline with self-distillation (nnU-NetSD) was trained for CA segmentation on 100 (training/validation/testing=80/10/10) diagnostic CT coronary angiography (CCTA) cases (0.49x0.49x0.63mm3, reconstructed with a deep learning algorithm (SnapShot Freeze-2, GE Healthcare) to minimize cardiac motion). Data were augmented with 82 external CCTA. To translate high-quality CA delineations on CCTA to CT simulation (CT-SIM) data, 40 best-fit CCTA cases were selected based on template-matched whole-heart agreement and deformably registered to each CT-SIM using uniGradICON registration. Final patient-specific CA habitats were defined via Euclidean clustering. To validate final habitats, eleven lung cancer patients with mean heart dose (MHD) >15 Gy were retrospectively evaluated, and main-branch CAs (proximal/medial/distal segments) were manually delineated, verified by a cardiovascular radiologist, and assessed for agreement with habitats. Habitat-spared treatment plans were created for three patients with high CA dose.
Results: nnU-NetSD achieved a Dice similarity (DSC) of 0.82±0.07 on CCTA. Across CT-SIM, registered whole-heart DSC averaged 0.95±0.01 and predicted main/full-branch habitats contained 90.9±11.3%/94.9±8.8% of CAs, respectively. Main/full-branch habitats accounted for 2.8-8.6%/2.8-20.6% of whole-heart volume, respectively. Habitat-spared plans reduced CA/spinal cord Dmax, and MHD by 13.0-15.6, 8.0, and 3.5 Gy with comparable plan quality (target, lung, esophageal endpoints <1 Gy/<1%). Right-CA V25Gy and left-anterior-descending-CA V15Gy were reduced by 25.0% and 22.1% across habitat-spared plans.
Conclusion: Patient-specific habitats, derived from translation of CCTA to CT-SIM, contained >90% of CA volumes. Including habitats in treatment planning reduced CA dose while maintaining plan quality. Future work includes applying habitats to additional RT cohorts for cardiac sparing.