ShapeField-Nodule: Continuous Shape Embedding for Pulmonary Nodule Segmentation
Overview
ShapeField-Nodule introduces a continuous signed distance field (SDF) framework for pulmonary nodule segmentation in low-dose CT, achieving sub-voxel precision and improved boundary delineation. Evaluations on multiple datasets demonstrate superior accuracy, boundary smoothness, and robustness compared to traditional voxel-based methods.
Background
Early detection of lung cancer via low-dose computed tomography (LDCT) is critical for improving patient outcomes. Accurate segmentation of pulmonary nodules is essential for diagnosis, malignancy risk assessment, and treatment planning. Conventional voxel-wise segmentation methods often struggle with irregular nodule shapes, noisy imaging conditions, and lack of anatomical plausibility, motivating the development of continuous shape modeling approaches.
Data Highlights
Dataset
Metric
ShapeField-Nodule
State-of-the-art Baselines
LIDC-IDRI
Dice Score
Superior (exact values not provided)
Lower
LUNA16
Surface Metrics
Improved average surface distance and Hausdorff distance
Inferior
Tianchi
Generalization & Robustness
Demonstrated superior performance under noise
Less robust
Key Findings
ShapeField-Nodule models pulmonary nodules as continuous signed distance fields, enabling sub-voxel geometric precision.
Integration of a lightweight MLP implicit head with a 3D U-Net backbone allows dense feature extraction combined with continuous shape representation.
A novel shape-aware refinement loss aligns predicted SDF gradients with image edges, enhancing boundary adherence and anatomical plausibility.
Outperforms traditional voxel-based segmentation methods on multiple public datasets in both overlap and boundary-based metrics.
Demonstrates improved boundary smoothness, topological consistency, and robustness to noise and low-contrast imaging conditions.
Clinical Implications
The continuous shape embedding approach of ShapeField-Nodule offers more accurate and anatomically coherent pulmonary nodule segmentation, which can improve early lung cancer diagnosis and treatment planning. Its robustness under noisy and low-contrast conditions may reduce diagnostic variability and support more reliable longitudinal monitoring of nodules.
Conclusion
ShapeField-Nodule represents a significant advancement in pulmonary nodule segmentation by leveraging continuous implicit shape representations, achieving superior accuracy and robustness over conventional voxel-based methods. This approach holds promise for enhancing early lung cancer detection and clinical decision-making.
