Dual-Scale Transformer-Guided Attention Network for Efficient Multi-OAR Segmentation in Head and Neck Radiotherapy

Accurate segmentation of organ-at-risk (OARs) in head and neck CT images is crucial for radiotherapy planning, but it remains a challenging task due to anatomical complexity, low soft-tissue contrast, and the presence of small, variable structures. We propose DSTANet, a novel dual-scale transformer-guided attention network that integrates multi-resolution encoding, transformer-based global context fusion, and anatomically guided attention refinement to deliver precise multi-OAR segmentation. Unlike traditional CNN-based methods, DSTANet effectively models long-range spatial dependencies while preserving high-resolution boundary detail. On the HNSCC-3DCT-RT dataset, DSTANet achieved a mean Dice Score of 97.5% and a mean 95th percentile Hausdorff Distance (HD95) of 2.32 mm, while on the MICCAI 2015 benchmark dataset, it achieved 90.0% Dice, which surpasses several state-of-the-art approaches both in terms of overlap and geometric accuracy. These results, combined with a sub-20-second inference time, establish DSTANet as a robust and clinically viable solution for automated head and neck OAR segmentation.