Multiscope topology learning with conditional updating for airway segmentation

Abstract Airway segmentation is essential in computer-assisted diagnosis and screening of bronchial diseases due to the inherent difficulty in obtaining a direct and clear visualization of airway trees from raw CT images. Although medical image segmentation technology is gradually maturing and beginning to be applied in clinics, challenges like breakages and leakages remain in airway segmentation. We propose a novel framework that enhances UNet3D with large-kernel attention for improved global and local feature extraction. A multitask prediction head across voxel, neighborhood, and surface scopes is introduced to better capture airway topology, supervised by customized loss functions. Additionally, a conditional updating strategy leverages a shared encoder and dual decoders to improve segmentation of thin branches by balancing over- and under-segmentation. Specifically, one decoder is optimized with hard examples to encourage over-segmentation, and the other refines results for accurate segmentation using all samples. Our model is quantitatively evaluated on the Binary Airway Segmentation dataset, achieving a Dice score of 0.904, precision of 0.954, tree detection rate of 0.950, and branch detection rate of 0.915, outperforming several recent methods in topological accuracy. In the Airway Tree Modeling Challenge 2022 validation set, our method ranks second overall by a mean position score across all metrics. Our future work aims to enhance prediction confidence and adaptability in ambiguous regions and improve generalizability and interpretability across diverse clinical datasets.