ShapeField-lung: continuous shape embedding for early lung cancer detection via pulmonary nodule segmentation.

Abstract

Accurate segmentation of pulmonary nodules in low-dose CT (LDCT) is vital for early lung cancer detection. Existing voxel-based methods often fail to capture irregular nodule boundaries, especially under noisy, low-contrast conditions. We propose ShapeField-Nodule, a continuous shape embedding framework that models nodule geometry as a signed distance field (SDF), enabling sub-voxel precision and anatomically coherent contours. Our method integrates a lightweight MLP-based implicit head with a 3D U-Net backbone to predict dense SDF values, and introduces a shape-aware refinement loss that aligns SDF gradients with image edges. Unlike discrete masks, our representation enforces boundary smoothness, topology regularization, and robustness to perturbations. Evaluations on LIDC-IDRI, LUNA16, and Tianchi datasets show state-of-the-art Dice and surface metrics. Extensive experiments demonstrate superior generalization, robustness under noise, and inference efficiency, highlighting the potential of continuous implicit fields as a principled alternative for medical image segmentation.