AirSeg: Learnable Interconnected Attention Framework for Robust Airway Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate airway segmentation in computed tomography (CT) is essential for diagnosing and managing lung diseases.
• Challenges include data imbalance and difficulty in segmenting small airway branches.

Purpose of the Study:

• To propose AirSeg, a learnable interconnected attention framework, to enhance airway segmentation accuracy in CT images.
• To improve robustness against spatial inconsistencies and noise for more reliable segmentation.

Main Methods:

• Developed AirSeg, integrating multiple attention mechanisms (image, positional, semantic, self-channel, cross-spatial) and a learnable variance-based embedding module.
• Evaluated AirSeg on in vivo and in situ datasets using UNet-based architectures with data augmentation and a hybrid loss function.
• Performed statistical analysis to assess accuracy improvements and conducted ablation studies.

Main Results:

• AirSeg integration led to statistically significant accuracy improvements: 16.18% for in vivo and 10.32% for in situ datasets.
• Achieved a weighted average accuracy improvement of 12.43% over conventional models.
• Demonstrated superior performance in segmenting both large airways and intricate small branches.

Conclusions:

• AirSeg significantly enhances airway segmentation accuracy and robustness in CT images.
• The framework enables more precise identification of airway structures, critical for early diagnosis and treatment planning.
• AirSeg improves the efficiency of automated airway analysis in clinical settings.