Attention-driven refinement network for continuity-preserving airway segmentation in class-imbalanced CT

Tubular airway segmentation is a prerequisite for bronchoscopic intervention in treating pulmonary diseases. Training convolutional neural networks (CNNs) for airway segmentation remains a clinical challenge due to the local discontinuities and distal small airway leakages caused by low resolutions and severe data imbalances. To address these issues, we propose an attention-driven refinement network, based on the degree of feature contribution, to improve the performance of fine-grained airway segmentation. A pointwise feature recalibration (PWFR) module is first designed to implement a differential feature treatment strategy by emphasizing competitive features and continuously suppressing redundant features, highlighting the prominence of airways in the learning task. Furthermore, a novel attention-driven knowledge distillation (AttdKD) module is developed to fully integrate the spatial and channel knowledge at various stages of the network, which strengthens the focus on distal small airways under conditions of class imbalance and mitigates the local discontinuity problem. The segmentation visualization results indicate that our refinement network effectively improves the thin airway recognition rate and improves the overall continuity of the airways under the guidance of the PWFR and AttdKD modules. The branches detected (BD) and tree length detected (TD) achieved scores of 93.96 %/81.7 % and 92.71 %/79.9 % on the ATM’22 and EXACT’09 datasets, respectively, and obtained scores of 92.72 %/92.44 % and 92.16 %/91.97 % on the abnormal case test sets of COVID-19 and Fibrosis, respectively. Extensive experiments demonstrate that our proposed method exhibits excellent sensitivity to distal small airways and achieves notable overall segmentation performance compared to the state-of-the-art (SOTA) baselines.