NA-UNETR: Transformer-Based 3D Segmentation of Left Anterior Descending Artery
A new deep learning model called NA-UNETR has been developed to improve the 3D segmentation of the Left Anterior Descending (LAD) artery in non-contrast CT scans taken during free breathing. This is really important because it helps reduce radiation exposure to the heart during thoracic radiotherapy. The LAD artery is quite small, has poor soft-tissue contrast, and varies a lot among different patients, making it tough to define its boundaries, especially with inconsistent manual contours. The model uses specialized Neighborhood Attention (NA) and Dilated NA (DiNA) blocks to capture detailed structures and broader context. Since there aren’t many labeled LAD datasets, it was pretrained on 1,000 CTA volumes and fine-tuned for better accuracy. You can find the study on arXiv with the ID 2608.12274v1.
Key facts
- NA-UNETR is a 3D transformer-based segmentation model.
- It uses Neighborhood Attention (NA) and Dilated NA (DiNA) blocks.
- The model is pretrained on 1,000 CTA volumes of general coronary anatomy.
- Fine-tuning uses LoRA-based parameter-efficient adaptation.
- The LAD artery is extremely small and has poor soft-tissue contrast.
- Manual contours show limited inter-observer agreement.
- The research is published on arXiv with identifier 2608.12274v1.
- The goal is to improve cardiac dose sparing in thoracic radiotherapy.
Entities
Institutions
- arXiv