A neighborhood attention transformer network for enhanced 3D segmentation of the left anterior descending artery.
Authors
Affiliations (7)
Affiliations (7)
- Department of Computer Science, Wayne State University, Detroit, Michigan, USA.
- Department of Radiation Oncology, Henry Ford Health, Detroit, Michigan, USA.
- Department of Clinical Oncology, Alexandria University, Alexandria, Egypt.
- Department of Radiology, Michigan State University, East Lansing, Michigan, USA.
- College of Medicine, Michigan State University, East Lansing, Michigan, USA.
- Department of Physics, Oakland University, Rochester, Michigan, USA.
- Institute for AI and Data Science, Wayne State University, Detroit, Michigan, USA.
Abstract
Accurate segmentation of the left anterior descending (LAD) artery in 3D free-breathing, non-contrast CT is critical for cardiac dose sparing in thoracic radiotherapy. The task is inherently difficult because the LAD is extremely small, exhibits poor soft-tissue contrast, and varies substantially across patients. Even manual contours show limited inter-observer agreement on non-contrast CT, underscoring the ambiguity of the vessel boundaries. To develop a transformer-based segmentation framework that enhances LAD delineation in low-contrast, imbalanced CT using local-global context modeling and uncertainty-guided optimization. We propose NA-UNETR, a 3D transformer-based segmentation model with neighborhood attention (NA) and Dilated NA (DiNA) blocks that jointly capture fine structural detail and long-range context. To address the limited availability of annotated LAD artery data, the model is pretrained on 1000 CTA volumes representing general coronary anatomy and then fine-tuned using a parameter-efficient adaptation strategy on 20 free-breathing institutional CT scans of the LAD artery. A composite Dice-Focal and Hausdorff loss, dynamically balanced via homoscedastic uncertainty, improves overlap and boundary accuracy. Preprocessing included intensity clipping, contrast enhancement, and artery-centric sampling, and postprocessing refined morphology. NA-UNETR achieved comparable performance to SOTA with improved efficiency and boundary properties. It reached 45.64% Dice, 38.16 mm HD95, and 10.01 mm ASD, improving Dice by 3.10 percentage points over nnU-Net and reducing HD95 by 2.96 mm relative to Swin UNETR, while showing the strongest boundary accuracy among all models and demonstrating improved geometric and centerline stability. On ImageCAS, it achieved 79.49% Dice, 8.89 mm HD95, and 1.02 mm ASD. Ablations confirmed that residual blocks, variable kernels, and uncertainty-weighted loss each contributed to performance stability and boundary precision. NA-UNETR effectively balances local precision and global context, enabling accurate segmentation of thin, low-contrast LAD structures. Its integration of NA, uncertainty-weighted loss, and LoRA-based fine-tuning demonstrates a robust and computationally efficient framework for substructure-level cardiac segmentation in radiotherapy planning.