148 - Branching Out: A Deep Learning Framework for Precise Coronary Artery Delineation in Radiation Therapy
Presenter(s)
C. Ruff1, N. Summerfield1, M. Dong2, P. Nagpal3, A. M. Baschnagel4, and C. Glide-Hurst1; 1Department of Human Oncology and Medical Physics, University of Wisconsin–Madison, Madison, WI, 2Department of Computer Science, Wayne State University, Detroit, MI, 3Department of Radiology, University of Wisconsin-Madison, Madison, WI, 4Department of Human Oncology, University of Wisconsin–Madison, Madison, WI
Purpose/Objective(s): Recent studies link increased radiotherapy (RT) dose to the major coronary artery (CA) branches with elevated cardiotoxicity risk. Yet, smaller CA branches with essential cardiac functions are not visible on standard RT imaging, preventing their delineation, sparing, and dosimetric assessment during treatment planning. To overcome this limitation and challenge the status quo of contouring only major CAs, we present a novel probabilistic CA model, integrating advanced deep learning (DL) and high-resolution contrast-enhanced coronary CT angiography (CCTA), to enable the estimation of otherwise unresolvable CA branches.
Materials/Methods: To develop a high-resolution segmentation framework, detectable main-branch CAs and sub-branches (“full branch”) were segmented on 182 CCTA images using nnU-Net with self-distillation and verified. Segmentation performance (random 70-10-20% split for training, validation, and testing) was evaluated using the Dice Similarity Coefficient (DSC). For a given RT image, full-branch probabilistic CA regions were predicted by 1) template matching and registering best-fit CCTA to the input via DL-based registration (multiGradICON), 2) estimating a probability density function (PDF) from the registered full-branch CA labels, incorporating random (e.g., cardiac motion) and systematic (e.g., delineation error) uncertainties via Monte Carlo, and 3) defining final regions as 90% of the PDF. To evaluate our full-branch model against the conventional main-branch model, 10 patients with cancer who underwent RT and CCTA were identified. Full-branch CAs were contoured on CCTA and compared to standard RT and probabilistic regions using inclusion metrics and 95% Hausdorff distance (HD) between each delineation and excluded full-branch CAs. A dosimetric comparison was shown for a patient with left-sided breast cancer.
Results: Segmentation of full-branch CAs yielded DSC of 0.79±0.07 for the test cohort. End-to-end prediction time of our full-branch probabilistic segmentation framework for 10 patients was 177±14 seconds, while manual delineation of main-branch RT definitions took 318±90 seconds. Inclusion of full-branch CAs within main-branch RT delineations ranged from 45.1-83.9%, while our probabilistic method resulted in ~94% inclusion. Similarly, HD ranged from 19.1-40.6 mm for manual RT delineations yet was <6.0 mm for our probabilistic, full-branch model. When applied to a patient with breast cancer, 19.4% of the full left anterior descending CA branch was excluded by standard RT definitions and received a Dmean of 20.1 Gy, yet was included by our probabilistic model.
Conclusion: Our DL full-branch CA segmentation model yielded acceptable accuracy. Predicted probabilistic regions yielded higher inclusion and lower HD for full-branch CAs compared to the standard main-branch model. Future work includes full-branch dose assessments in thoracic cancer to lay the groundwork for deeper insights into RT-related cardiotoxicities.