[Paper Review] NaviAirway: a Bronchiole-sensitive Deep Learning-based Airway Segmentation Pipeline

Andong Wang, Terence Chun Tam+3 more|arXiv (Cornell University)|Mar 8, 2022

Lung Cancer Diagnosis and Treatment10 citations

TL;DR

NaviAirway is a deep learning-based airway segmentation pipeline that improves topological accuracy in chest CT scans by introducing a bronchiole-sensitive loss function and an iterative training strategy to address size imbalance across airway generations. It outperforms existing methods in detecting fine bronchioles and generalizes well across diverse CT scans, with performance gains across multiple backbone models.

ABSTRACT

Airway segmentation is essential for chest CT image analysis. Different from natural image segmentation, which pursues high pixel-wise accuracy, airway segmentation focuses on topology. The task is challenging not only because of its complex tree-like structure but also the severe pixel imbalance among airway branches of different generations. To tackle the problems, we present a NaviAirway method which consists of a bronchiole-sensitive loss function for airway topology preservation and an iterative training strategy for accurate model learning across different airway generations. To supplement the features of airway branches learned by the model, we distill the knowledge from numerous unlabeled chest CT images in a teacher-student manner. Experimental results show that NaviAirway outperforms existing methods, particularly in the identification of higher-generation bronchioles and robustness to new CT scans. Moreover, NaviAirway is general enough to be combined with different backbone models to significantly improve their performance. NaviAirway can generate an airway roadmap for Navigation Bronchoscopy and can also be applied to other scenarios when segmenting fine and long tubular structures in biomedical images. The code is publicly available on https://github.com/AntonotnaWang/NaviAirway.

Motivation & Objective

• To address the challenge of topological inaccuracy in airway segmentation, particularly the failure to detect fine bronchioles due to voxel-wise class imbalance.
• To overcome the size imbalance problem between low- and high-generation airways, where large airways dominate in volume but fine bronchioles are clinically critical.
• To improve model generalization and feature learning across airway generations using an iterative training strategy and knowledge distillation from unlabeled data.
• To develop a plug-and-play framework compatible with various backbone models to boost their segmentation performance.
• To enable accurate 3D airway roadmap generation for navigation bronchoscopy and segmentation of fine tubular structures in biomedical imaging.

Proposed method

• Proposes a bronchiole-sensitive loss function, $\mathfrak{L}_{pen}$, that penalizes moderate-confidence airway predictions and suppresses low-confidence false positives, enhancing detection of fine bronchioles.
• Introduces an iterative training strategy inspired by human visual perception, where the model alternately focuses on low- and high-generation airways to preserve learned features across generations.
• Employs a teacher-student distillation framework using numerous unlabeled CT scans to transfer knowledge of airway branch features, improving feature representation without manual annotation.
• Uses sampling based on the ratio $r_i^j \approx 2/\hat{r}$ to balance airway branch volumes during training, mitigating the impact of size imbalance.
• Applies the pipeline to multiple backbone models (3D U-Net, V-Net, VoxResNet, Wang et al.) without hyperparameter tuning, demonstrating broad compatibility and performance gain.
• Replaces standard metrics with adjusted versions—Adjusted Branch Detected (ABD) and Adjusted Tree-length Detected (ATD)—to account for missed true branches in reference labels.

Experimental results

Research questions

• RQ1 Can a loss function be designed to prioritize detection of fine bronchioles while suppressing noise, thereby improving topological accuracy in airway segmentation?
• RQ2 How can a deep learning model effectively learn features across vastly different airway generations despite severe size imbalance?
• RQ3 To what extent does knowledge distillation from unlabeled CT scans improve feature learning for airway segmentation?
• RQ4 Does the iterative training strategy enhance model performance and generalization across different CT scanners and institutions?
• RQ5 Can the proposed pipeline be effectively combined with various backbone architectures to improve their segmentation performance?

Key findings

• NaviAirway significantly improves detection of high-generation bronchioles, achieving 95.4% sensitivity and 90.2% Branch Detected (BD) on the QMH dataset, outperforming existing methods.
• The model achieves 114.2% Adjusted Branch Detected (ABD) and 113.2% Adjusted Tree-length Detected (ATD) on the BAS dataset, indicating it detects more true bronchioles than the reference labels.
• On the unseen QMH dataset, NaviAirway maintains strong performance with 132.4% ABD and 108.6% ATD, demonstrating robustness to new data and scanner variability.
• When applied to 3D U-Net, V-Net, VoxResNet, and Wang et al.'s model, NaviAirway boosts performance across all backbones without hyperparameter tuning, confirming its generalizability.
• The bronchiole-sensitive loss function effectively penalizes moderate-confidence predictions and rejects low-confidence noise, as shown by the derivative analysis of $\mathfrak{L}_{pen}$.
• The iterative training strategy prevents catastrophic forgetting of low-generation airway features during high-generation fine-tuning, preserving topological integrity.