Alleviating Class-Wise Gradient Imbalance for Pulmonary Airway Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Pulmonary Medicine

Background:

• Automated airway segmentation is crucial for pulmonary interventions.
• Severe class imbalance in peripheral bronchi hinders CNN-based segmentation accuracy.
• Gradient erosion and dilation during back-propagation limit learning of small airway structures.

Purpose of the Study:

• To address gradient issues and class imbalance in automated airway segmentation.
• To improve the accuracy and morphological completeness of distal airway segmentation.
• To enhance the training of shallow layers in CNNs for better small structure learning.

Main Methods:

• Utilized group supervision and WingsNet to provide complementary gradient flows.
• Introduced a General Union loss function to mitigate intra-class imbalance between large and small airways.
• Employed distance-based weights and adaptive gradient ratio tuning within the loss function.

Main Results:

• The proposed method demonstrated improved accuracy in predicting airway structures.
• Enhanced morphological completeness of segmented airways was observed.
• The approach outperformed existing baseline methods on public datasets.

Conclusions:

• The developed techniques effectively overcome gradient erosion and class imbalance challenges in airway segmentation.
• The General Union loss function and group supervision significantly improve the segmentation of small and peripheral airways.
• This work advances automated airway segmentation for improved pulmonary diagnostics and navigation.