A Neighborhood Attention Transformer Network for Enhanced 3D Segmentation of the Left Anterior Descending Artery
NA-UNETR, a 3D transformer-based segmentation model, uses Neighborhood Attention and Dilated Neighborhood Attention blocks to capture fine structural detail and long-range context. It was pretrained on 1,000 CTA volumes of general coronary anatomy and fine-tuned with LoRA-based parameter-efficient adaptation on 20 free-breathing institutional CT scans. The model achieved 45.64% Dice, 38.16 mm HD95, and 10.01 (unspecified metric) on LAD segmentation.
Researchers developed NA-UNETR, a 3D transformer-based segmentation model for the Left Anterior Descending artery in free-breathing, non-contrast CT. The model combines Neighborhood Attention and Dilated Neighborhood Attention to capture both local detail and long-range context. Due to limited annotated LAD data, it was pretrained on 1,000 CTA volumes and fine-tuned with LoRA on 20 institutional CT scans. A composite Dice-Focal and Hausdorff loss, dynamically balanced via homoscedastic uncertainty, was used. Reported results include 45.64% Dice and 38.16 mm HD95.
The use of Neighborhood Attention and Dilated Neighborhood Attention in a 3D transformer architecture enables efficient local-global context modeling for small, low-contrast anatomical structures. LoRA-based fine-tuning addresses data scarcity by adapting a pretrained model with few parameters. Uncertainty-weighted composite loss balances overlap and boundary accuracy, which is critical for ambiguous vessel boundaries.
This work demonstrates a practical approach to applying transformer-based segmentation in medical imaging with limited annotated data, leveraging pretraining on related large datasets and parameter-efficient fine-tuning. It may influence development of AI tools for cardiac radiotherapy planning, where accurate LAD delineation is needed for dose sparing.
Improved LAD segmentation could enhance cardiac dose sparing in thoracic radiotherapy, potentially reducing radiation-induced cardiac toxicity. The method's use of pretraining and LoRA may lower annotation costs and enable faster adaptation to new institutions, making it attractive for medical imaging AI vendors.
Next observable signals include validation on larger multi-institutional datasets, comparison with other state-of-the-art segmentation models, and potential integration into radiotherapy planning workflows. Further research may explore uncertainty quantification and clinical deployment feasibility.