Active Learning and Cascaded nnU-Net v2 for Multi-structure Segmentation of Thoracic CT: A Single-Center Pilot Study

Accurate segmentation of thoracic structures on chest computed tomography (CT) is important for lung cancer surgery, but manual annotation is labor intensive and particularly challenging for small structures such as pulmonary vessels and nodules. We compared four nnU-Net v2-based segmentation configurations: single-stage full-data training, active learning alone, cascaded segmentation alone, and active learning combined with cascaded segmentation in a retrospective single-center cohort. This retrospective study included chest CT scans from 85 patients. In the cascaded configurations, pulmonary lobes were segmented in the first stage to provide global anatomical context, followed by refinement of the bronchus, pulmonary vessels, and nodules within the constrained region. All four configurations were evaluated using the same fixed patient-level internal test set of 30 patients, with segmentation performance assessed using the Dice similarity coefficient (DSC) and the 95th percentile Hausdorff distance (HD95). The full-data baseline achieved DSCs of 0.92-0.93 for the five pulmonary lobes, with lower performance for the bronchus (0.82), pulmonary vessels (0.67), and pulmonary nodules (0.78). The combined active learning and cascaded configuration improved lobar DSCs to 0.94-0.96. Compared with the baseline, DSC increased to 0.90, 0.77, and 0.86 for the bronchus, pulmonary vessels, and pulmonary nodules, respectively (all p < 0.05), while HD95 decreased from 5.4 to 2.5 mm for pulmonary vessels and from 4.8 to 2.1 mm for pulmonary nodules. In this single-center cohort, cascaded segmentation strategies showed improved internal test performance for small thoracic structures, supporting further evaluation in larger multicenter cohorts.