SCPM-Net

SCPM-Net detects lung nodules in 3D Computed Tomography (CT) scans using an anchor-free sphere representation and center-points matching to improve detection accuracy for lung cancer screening.

Key Features:

• Anchor-free detection: Eliminates predefined anchor boxes and associated parameters required by anchor-based detectors that rely on Convolutional Neural Networks (CNNs).
• Sphere representation: Replaces bounding boxes with a bounding sphere defined by centroid, radius, and local offset in 3D space.
• Sphere-based IoU loss: Uses a sphere-based intersection-over-union loss function to stabilize and efficiently train the detection network.
• Center points matching (CPM): Implements positive center-points selection and matching to assign detection points without anchors.
• Online hard example mining and re-focal loss: Applies online hard example mining and re-focal loss to improve point assignment accuracy and mitigate class imbalance.
• Spatial coordinate fusion: Incorporates multi-level spatial coordinate maps fused with the feature extractor to capture spatial information.
• 3D attention modules: Integrates 3D squeeze-and-excitation attention modules to capture 3D contextual information.
• Validation and performance: Validated on the LUNA16 dataset with an average sensitivity of 89.2% at seven predefined false positives per scan and shown to outperform existing anchor-based and other anchor-free methods, with sphere representation yielding higher detection accuracy than bounding-box representation.

Scientific Applications:

• Lung cancer screening and diagnosis: Detects and localizes lung nodules in 3D CT scans to support screening and diagnostic workflows.
• Benchmarking detection methods: Serves as a framework for evaluating nodule detection performance on datasets such as LUNA16.
• Robust detection across nodule sizes: Improves robustness for detecting lung nodules of diverse sizes without manual anchor parameter tuning.

Methodology:

SCPM-Net implements an anchor-free detection pipeline using sphere representation (centroid, radius, local offset), a sphere-based IoU loss, positive center-points selection and matching (CPM), online hard example mining, re-focal loss, multi-level spatial coordinate map fusion with a feature extractor, and 3D squeeze-and-excitation attention modules.