Early detection of lung cancer depends on reliable delineation of pulmonary nodules in low-dose CT (LDCT). Noisy images, low contrast and irregular nodule morphology often push voxel-based segmentation methods to their limits, producing jagged contours and unstable boundaries. A continuous shape framework called ShapeField-Nodule addresses these problems by modelling nodule geometry as a signed distance field (SDF) rather than a binary mask. Built on a 3D U-Net with a lightweight implicit decoder and a boundary-aware refinement loss, the approach aims to deliver sub-voxel precision, smooth surfaces and resilience to imaging variation. Evaluations across common datasets report improved overlap and surface metrics, qualitative gains on challenging cases and stable performance under test-time perturbations. The work positions continuous implicit fields as a credible alternative to discrete masks in LDCT nodule workflows.
Continuous Representation for Boundary Fidelity
ShapeField-Nodule replaces per-voxel classification with an SDF whose zero level set defines the nodule boundary. This continuous parametrisation regularises shape, promotes anatomical plausibility and yields contours that avoid blocky artefacts. A coordinate-conditioned multilayer perceptron predicts dense SDF values from multi-scale features extracted by a 3D U-Net, enabling sub-voxel boundary localisation without post-hoc surface clean-up.
Must Read: AI Triage Cuts Workload in Low-Dose Lung Screening
To couple geometry to image evidence, a shape-aware refinement loss aligns predicted SDF gradients with edge cues from the CT, encouraging boundaries to coincide with strong intensity transitions rather than noise. Training also concentrates learning where it matters most: a larger share of sampled points comes from a narrow band near the boundary, with the remainder drawn more broadly to preserve global shape. Ablations attribute accuracy and surface smoothness to this combination of continuous representation, edge-aligned supervision and near-boundary sampling, while positional encoding and the implicit head further improve fine-scale detail.
Performance Across Benchmarks and Perturbations
On the LIDC-IDRI benchmark, the method attains higher accuracy than strong voxel-based alternatives and an implicit SDF baseline, with gains observed on both overlap and surface distance measures. The Dice similarity coefficient (Dice) reaches 87.3% alongside improved boundary metrics, indicating better adherence to expert contours in routine and difficult cases, including juxtapleural and irregular nodules. Statistical testing supports the significance of these improvements over the strongest comparators.
Generalisation is demonstrated without fine-tuning on two external datasets. Tested directly on LUNA16 and Tianchi, ShapeField-Nodule maintains Dice scores in the mid-80s and outperforms voxel-based references as well as the implicit baseline trained under standard settings. These cross-dataset results indicate stability across acquisition protocols, contrast variations and scanner characteristics that commonly reduce transferability for mask-based approaches.
Robustness analyses add synthetic degradations at test time. Under Gaussian noise, motion blur and contrast shifts, accuracy declines only modestly and boundary quality remains stable, with limited changes in high-percentile surface distances. Qualitative examples show the continuous field preserving smooth, coherent contours under perturbation, avoiding the fragmented edges that can emerge when voxel-wise decisions are stressed by LDCT artefacts.
Efficiency and Design Choices
Beyond accuracy, the framework is designed for practical deployment. The implicit head adds minimal overhead, yielding an inference time of about 0.11 seconds per 3D patch while retaining strong accuracy. Model size remains compact compared with heavier voxel-based architectures, supporting use in screening pipelines where throughput and memory are constrained.
Ablation studies clarify contributions of core components. Replacing the SDF target with a binary mask produces the largest drop in Dice and worsens surface distances, underscoring the value of continuous shape modelling. Removing the edge-aligned refinement loss degrades boundary localisation, while omitting positional encoding reduces fine-geometry capture. Substituting the coordinate-aware implicit head with a standard convolutional classifier also harms performance, indicating that explicit coordinate conditioning is important for precise contours. Sensitivity analysis identifies balanced settings for the edge-refinement weight and near-boundary sampling ratio, stabilising training and avoiding over-sharpening or over-smoothing. Together, these choices deliver boundaries that are both smooth and anatomically plausible under LDCT noise and contrast variation.
Continuous shape embedding provides a principled route to more accurate and coherent LDCT nodule segmentation. By predicting an SDF with edge-aligned supervision and focused sampling, ShapeField-Nodule improves overlap and boundary metrics, generalises across datasets and remains robust to common degradations, all with modest computational cost. For healthcare teams building screening workflows, the approach offers reliable contours that can support volumetric tracking, radiomics-driven analysis and surface-based interpretation, while reducing susceptibility to the artefacts that challenge voxel-based masks.
Source: npj digital medicine
Image Credit: iStock
