Topology-Constrained Deep Learning for Automated Segmentation of the Left Anterior Descending Coronary Artery for Thoracic Radiotherapy
Abstract
Purpose
Accurate segmentation of the left anterior descending (LAD) coronary artery is essential for cardiac sparing in thoracic radiotherapy, particularly for lung cancer patients receiving high-dose radiation near the heart. However, the LAD is thin, low-contrast, and often partially obscured on non-contrast planning CT, making fully automated segmentation challenging for standard deep learning models. This study proposes a topology-constrained 3D deep learning framework that integrates anatomical continuity and geometric priors to improve LAD segmentation performance.
Methods
A 3D U-Net convolutional neural network was trained to segment the LAD using thoracic RT planning CT scans from 46 lung cancer patients. In addition to conventional binary cross-entropy (BCE) and Dice losses, we introduced a composite topology-constrained loss function incorporating soft centerline Dice (clDice), a differentiable skeleton length prior, endpoint probability enforcement with distal emphasis, and surface-based regularization. Model performance was evaluated on a held-out cohort using both soft (probabilistic overlap) Dice and hard (thresholded) Dice, reported per case and as cohort means.
Results
The proposed topology-constrained framework achieved substantial improvements in segmentation accuracy and anatomical completeness compared to Dice-only training. Mean hard Dice increased from 0.46 ± 0.07 in early training to 0.79 ± 0.05 after incorporating topological constraints, with corresponding mean soft Dice improving from 0.54 ± 0.08 to 0.85 ± 0.09. The inclusion of centerline and endpoint priors significantly reduced vessel discontinuities and false truncations, producing anatomically consistent and topologically connected LAD segmentations spanning from the proximal coronary origin to distal myocardial segments.
Conclusion
A topology-constrained 3D deep learning framework improved automated LAD segmentation performance on lung RT planning CT and produced more anatomically continuous predictions. This approach provides a practical foundation for routine LAD contour generation and subsequent LAD dosimetry extraction, supporting future investigations of LAD dose-response relationships and associations with post-RT cardiac events.