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

Area of Science:

• Medical Imaging
• Computer Vision
• Radiology

Background:

• Accurate pulmonary nodule segmentation in low-dose CT (LDCT) is crucial for early lung cancer detection.
• Voxel-based segmentation methods struggle with irregular nodule shapes and low-contrast imaging.

Purpose of the Study:

• To introduce ShapeField-Nodule, a novel continuous shape embedding framework for precise pulmonary nodule segmentation.
• To overcome limitations of existing methods in capturing nodule geometry and boundary details.

Main Methods:

• Developed ShapeField-Nodule, modeling nodule geometry as a signed distance field (SDF) for sub-voxel accuracy.
• Integrated a lightweight MLP-based implicit head with a 3D U-Net backbone to predict SDF values.
• Introduced a shape-aware refinement loss to align SDF gradients with image edges.

Main Results:

• Achieved state-of-the-art Dice and surface metrics on LIDC-IDRI, LUNA16, and Tianchi datasets.
• Demonstrated superior generalization, robustness to noise, and inference efficiency compared to existing methods.
• Validated the effectiveness of continuous implicit fields for medical image segmentation.

Conclusions:

• ShapeField-Nodule offers a principled approach for medical image segmentation, providing anatomically coherent contours.
• The continuous SDF representation enforces boundary smoothness and topology regularization, enhancing segmentation accuracy.
• This framework shows significant potential for improving early lung cancer detection through advanced image analysis.