Abstract
Accurate delineation of pulmonary nodules in chest computed tomography (CT) is essential for early lung cancer diagnosis and treatment planning. However, voxel-wise segmentation methods often produce fragmented masks and inconsistent topology due to low contrast, anatomical variability, and imaging noise. We propose CoreFormer, a segmentation framework that models nodules through structural core anchoring and geodesic shape decoding. CoreFormer identifies the intrinsic topological core of each nodule and generates continuous boundaries guided by anatomy-aware geodesic paths. It is built upon a Swin Transformer backbone and a dual-branch decoder consisting of a Structural Core Predictor and a Context-Aware Shape Decoder, enhanced by Feature Manifold Regularization for discriminative feature learning. Extensive experiments on four public datasets-LIDC-IDRI, LNDb, Tianchi-Lung (MosMedData), and NSCLC-Radiomics-demonstrate that CoreFormer achieves state-of-the-art boundary accuracy and topological fidelity, offering robust and high-fidelity pulmonary nodule segmentation.