Adaptive Patch-Based Attention 3D U-Net with Topology-Aware Refinement for Robust Airway Segmentation from Chest CT

Abstract

The precision of the airway segmentation obtained from chest CT is critical for examination of pulmonary disease, surgical planning, and quantitative airway assessment. Yet, segmentation of peripheral airway branches is complicated by class imbalance, small size of structures, and topological discontinuities. This work offers an adaptive patch-based Attention 3D U-Net, which incorporates multi-scale attention and dynamic patch sampling to improve feature learning in small airway regions. Adaptive patching intelligently selects informative sub-volumes based on coarse predictions generated by the network, ensuring meaningful supervision. In addition, the network is combined with a topology-aware post-processing pipeline that maintains airway connectivity by selecting the largest component and refining it with support from skeletons. Experiments on the AeroPath airway dataset exhibit improved segmentation completeness and structural preservation relative to classic 3D U-Net and Attention 3D U-Net architectures. The proposed method obtains greater Dice similarity, higher Branch Detection Rate (BDR), and enhanced tree-length recovery while maintaining computational efficiency. Skeleton-based analyses establish that segmented airway trees are more structurally consistent.