Learning Tubule-Sensitive CNNs for Pulmonary Airway and Artery-Vein Segmentation in CT

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Segmenting pulmonary airways, arteries, and veins in CT scans is difficult due to class imbalance.
• Existing methods struggle with sparse supervisory signals and accurately identifying small tubular structures.

Purpose of the Study:

• To develop an accurate CNN-based method for segmenting pulmonary airways, arteries, and veins in non-contrast CT.
• To improve sensitivity to peripheral bronchioles, arterioles, and venules.

Main Methods:

• A feature recalibration module to optimize learned features and integrate spatial information.
• An attention distillation module to enhance representation learning of tubular objects via recursive attention map passing.
• Incorporation of lung context and distance transform maps for improved artery-vein differentiation.

Main Results:

• The proposed method demonstrates superior sensitivity to fine peripheral structures.
• Significant performance gains were observed due to the novel components.
• Extracted more airway and vessel branches compared to state-of-the-art methods while maintaining competitive segmentation performance.

Conclusions:

• The developed CNN method effectively addresses the challenges of pulmonary vessel and airway segmentation.
• The feature recalibration and attention distillation modules significantly enhance segmentation accuracy.
• The approach offers improved capabilities for analyzing lung vasculature and airways in medical imaging.