Progressive curriculum learning with Scale-Enhanced U-Net for continuous airway segmentation.
Authors
Affiliations (7)
Affiliations (7)
- State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, 100190, China.
- School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, 100049, China.
- Department of Pulmonary and Critical Care Medicine, The First Affiliated Hospital of Sun Yat-Sen University, Guangzhou, 510080, China.
- Centre of AI and Robotics (CAIR), Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences, Hong Kong, China.
- State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, 100190, China. [email protected].
- Centre of AI and Robotics (CAIR), Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences, Hong Kong, China. [email protected].
- School of Biomedical Engineering and Imaging Sciences, King's College London, London, UK. [email protected].
Abstract
Continuous and accurate segmentation of airways in chest CT images is essential for preoperative planning and real-time bronchoscopy navigation. Despite advances in deep learning for medical image segmentation, maintaining airway continuity remains a challenge, particularly due to intra-class imbalance between large and small branches and blurred CT details. To address these challenges, we propose a progressive curriculum learning pipeline and a Scale-Enhanced U-Net (SE-UNet) to improve detail extraction, thereby enhancing segmentation continuity. Compared with previous connectivity-aware methods, our framework directly tackles the imbalance between large and small branches through end-to-end progressive learning, while balancing airway tree completeness and accuracy. Specifically, our curriculum learning pipeline comprises three stages. Stage 1 performs coarse learning to extract main airways. Stage 2 introduces a General Union Loss (GUL) to improve the identification of smaller airways. In Stage 3, we propose an Adaptive Topology-Responsive Loss (ATRL), which encourages the network to focus on preserving airway continuity. Throughout all stages, a crop sampling strategy is employed to reduce feature interference between airways of varying scales, effectively addressing the intra-class imbalance. The progressive training pipeline shares the same SE-UNet, integrating multi-scale inputs and Detail Information Enhancers (DIEs) to enhance information flow and effectively capture the intricate details of small airways. Additionally, we propose a robust airway tree parsing method and hierarchical evaluation metrics to provide more clinically relevant and precise analysis. Extensive experiments demonstrate that our method outperforms existing approaches on the ATM'22 challenge dataset and achieves a 9.631% improvement in the Tree length Detection rate (TD) of small airways and a 4.622% improvement in the Branch Detection rate (BD) of the overall airway tree on our in-house dataset, thereby significantly enhancing small-airway accuracy and overall airway tree completeness.