Hierarchical deep learning for pulmonary structures segmentation in surgical planning

Objective: Precise segmentation of pulmonary segments from chest CT is essential for the diagnosis and treatment of lung diseases, especially for procedures such as segmentectomy. However, pulmonary segments are abstract anatomical units whose accurate delineation depends on the reliable segmentation of tubular structures including bronchi and arteries. This study aims to develop a robust automated framework to address this clinical challenge. Methods: Given chest CT images as input, we propose a two-stage automated segmentation methodology. In the first stage, a UMaxLSTM network with a union loss function is developed to perform efficient segmentation of pulmonary tubular structures (bronchi and arteries) as well as lung lobes. These structures provide anatomical boundaries critical for distinguishing pulmonary segments. In the second stage, we introduce the MultiFuseUNet network, which incorporates the tubular structure segmentation from stage one as guiding information to achieve precise automatic segmentation of pulmonary segments. Results: Segmentation performance of bronchi and arteries was evaluated on two public datasets, while pulmonary segment segmentation was validated on a private dataset with high-quality expert annotations. Experimental results demonstrate that the proposed two-stage methodology effectively segments pulmonary anatomical structures, achieving approximately 85% Dice score in pulmonary segment segmentation. Conclusion: The proposed approach provides thoracic surgeons and pulmonologists with an accurate and reliable tool for understanding pulmonary segment anatomy, showing significant clinical value. It is particularly suitable for preoperative planning of segmentectomy, helping clinicians avoid damage to adjacent healthy tissues.