AirSeg: Learnable Interconnected Attention Framework for Robust Airway Segmentation

Chetana Krishnan; Shah Hussain; Denise Stanford; Venkata Sthanam; Sandeep Bodduluri; S Vamsee Raju; Steven M Rowe; Harrison Kim

doi:10.1007/s10278-025-01545-z

AirSeg: Learnable Interconnected Attention Framework for Robust Airway Segmentation

J Imaging Inform Med. 2025 May 22. doi: 10.1007/s10278-025-01545-z. Online ahead of print.

Authors

Chetana Krishnan¹, Shah Hussain², Denise Stanford², Venkata Sthanam³, Sandeep Bodduluri^{2

3}, S Vamsee Raju², Steven M Rowe², Harrison Kim⁴

Affiliations

¹ Department of Biomedical Engineering, The University of Alabama at Birmingham, Birmingham, AL, 35294, USA.
² Department of Pulmonary, Allergy, and Critical Care, The University of Alabama at Birmingham, Birmingham, AL, 35294, USA.
³ Department of Electrical Engineering, The University of Alabama at Birmingham, Birmingham, AL, 35294, USA.
⁴ Department of Radiology, The University of Alabama at Birmingham, VH G082, 1720 2nd Avenue South, Birmingham, AL 35294, USA. hyunkiuab@gmail.com.

PMID: 40404874
DOI: 10.1007/s10278-025-01545-z

Abstract

Accurate airway segmentation is vital for diagnosing and managing lung diseases, yet it remains challenging due to data imbalance and difficulty detecting small airway branches. This study proposes AirSeg, a learnable interconnected attention framework incorporating advanced attention mechanisms and a learnable embedding module, to enhance airway segmentation accuracy in computed tomography (CT) images. The proposed framework integrates multiple attention mechanisms, including image, positional, semantic, self-channel, and cross-spatial attention, to refine feature representations at various network and data levels. Additionally, a learnable variance-based embedding module dynamically adjusts input features, improving robustness against spatial inconsistencies and noise. This improves the model's robustness to spatial inconsistencies and noise, leading to more reliable segmentation results, especially in clinically challenging regions. AirSeg can be integrated with any UNet-like network with flexibility. The framework was evaluated on two datasets (in vivo and in situ) using several UNet-based architectures, comparing performance with and without AirSeg integration. Training employed data augmentation, a hybrid loss function combining Dice Similarity Coefficient and Intersection over Union losses, and statistical analysis to assess accuracy improvements. Integrating AirSeg into segmentation models led to statistically significant improvements in accuracy. Specifically, accuracy increased by 16.18% (p = 0.0035) for in vivo datasets and by 10.32% (p = 0.0097) for in situ datasets. These enhancements enable more precise identification of airway structures, including small branches, critical for early diagnosis and treatment planning in pulmonary care. The proposed model achieved a weighted average accuracy improvement of 12.43% (p = 0.0004) over other conventional models. AirSeg demonstrated superior performance in capturing both global structures and fine details, effectively segmenting large airways and intricate branches. Ablation studies validated the contributions and impact of individual attention mechanisms and the embedding module. The improvement in accuracy translates to more precise airway segmentation, enhancing the detection of small branches crucial for early diagnosis and treatment planning. The statistically significant p-values confirm that these gains are reliable, reducing manual correction efforts and improving the efficiency of automated airway analysis in clinical settings.

Keywords: Airways; Attention; Embedding; Learnable; Medical image segmentation; Multi-variant; Transformers.

Abstract

Grants and funding