Direction-Aware convolution for airway tubular feature enhancement network

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Manual airway annotation in CT scans is labor-intensive and requires expertise.
• Accurate airway segmentation is vital for bronchoscopic navigation and robotic systems.
• Existing deep learning methods struggle with the intricate, tree-like airway structures, leading to segmentation errors.

Purpose of the Study:

• To develop a novel deep learning network, TfeNet, for improved automatic airway segmentation.
• To address the limitations of conventional convolutions in capturing fine, tubular airway features.
• To enhance the accuracy, continuity, and robustness of airway segmentation in medical images.

Main Methods:

• Proposed TfeNet, featuring a direction-aware convolution operator for adaptive alignment with tubular airway structures.
• Introduced a tubular feature fusion module (TFFM) using asymmetric convolutions and residual connections.
• Validated TfeNet on public (BAS) and challenge datasets (ATM22, AIIB23).

Main Results:

• TfeNet demonstrated superior accuracy and continuity on the BAS dataset.
• Achieved a top score of 94.95% on the ATM22 dataset, balancing accuracy and continuity.
• Showcased excellent leakage control and precision on the challenging AIIB23 dataset.

Conclusions:

• TfeNet effectively segments complex airway structures, overcoming limitations of existing methods.
• The proposed direction-aware convolution and TFFM significantly enhance airway segmentation performance.
• TfeNet shows strong potential for clinical applications in bronchoscopy and related fields.