An end-to-end multi-scale airway segmentation framework based on pulmonary CT image

Area of Science:

• Medical Imaging
• Computer Vision
• Radiology

Background:

• Accurate airway segmentation is crucial for diagnosing lung diseases.
• The multi-scale, tree-like structure of airways presents a significant segmentation challenge.
• Convolutional Neural Networks (CNNs) have shown promise in medical image segmentation.

Purpose of the Study:

• To propose a novel two-stage 2D + 3D deep learning framework for multi-scale airway tree segmentation.
• To improve the accuracy of airway segmentation, particularly for intrapulmonary branches.
• To enhance the diagnostic capabilities for lung diseases through precise airway analysis.

Main Methods:

• A two-stage framework combining a 2D full airway SegNet (2D FA-SegNet) and a 3D airway RefineNet (3D ARNet).
• Incorporation of multi-scale spatial pyramid and residual skip connection modules in 2D FA-SegNet.
• Implementation of a hard sample selection strategy and false positive/negative losses in 3D ARNet for refined segmentation.

Main Results:

• Achieved a Dice Similarity Coefficient (DSC) of 0.931 and IoU of 0.871 for the whole airway tree.
• Obtained DSC of 0.699 and IoU of 0.543 for the intrapulmonary bronchi tree.
• Cascaded 3D ARNet with other methods increased detected tree length and branch rates by up to 46.33% and 42.97%, respectively.

Conclusions:

• The proposed 2D + 3D framework effectively addresses the multi-scale nature of airway segmentation.
• The method demonstrates superior performance in segmenting both overall and fine intrapulmonary airways.
• This approach holds significant potential for advancing lung disease diagnosis and analysis.