Full-Volume Lung CT Segmentation and Surface Export

summary

LungVolSeg is a reproducible engineering pipeline for full-volume chest CT lung segmentation, surface extraction, and metric reporting. The goal was to build something closer to a navigation-prep system than a one-off segmentation notebook.

problem

Robotic bronchoscopy and related lung navigation workflows rely on CT-derived anatomical understanding. Before navigation, a CT scan is commonly used for planning, segmentation, registration, and 3D visualization. A practical navigation-prep pipeline therefore has to manage spatial metadata correctly, segment anatomy in 3D, export geometric surfaces, and report quantitative validation metrics.

approach

The pipeline focuses on the engineering substrate of that workflow rather than claiming clinical deployment. It is built from standard open-source imaging tools:

• SimpleITK for medical image IO and spatial preprocessing
• MONAI for 3D medical image segmentation
• VTK for surface extraction and mesh export
• PyTorch for model optimization

The first real-data target is binary lung segmentation from full-volume CT scans because it provides a stable organ-level surface for navigation-prep experiments.

dataset

The pipeline uses the COVID-19 CT Lung and Infection Segmentation Dataset hosted on Zenodo.

• DOI: 10.5281/zenodo.3757476
• License: CC-BY-4.0
• Cases: 20 labeled full CT scans
• Downloaded archives: COVID-19-CT-Seg_20cases.zip and Lung_Mask.zip

The original dataset includes left lung, right lung, and infection annotations. This implementation converts the lung masks into a binary target:

• 0: background
• 1: lung

pipeline

The workflow performs the following steps:

1. Download CT and lung-mask archives from Zenodo
2. Verify archive MD5 checksums
3. Extract NIfTI volumes
4. Pair CT volumes with lung masks
5. Crop around the lung-mask bounding box
6. Clip CT intensities to [-1000, 400] HU and normalize to [0, 1]
7. Resize each case to a compact 3D target shape
8. Train a MONAI 3D UNet
9. Run inference on prepared cases
10. Export predicted lung surfaces as .stl and .vtp
11. Compute Dice and HD95 metrics
12. Generate a model card

The default target shape is 96 x 128 x 128. A smaller validation run uses 64 x 96 x 96.

model and export

The segmentation model is a MONAI 3D UNet with in_channels=1, out_channels=2, channel schedule (16, 32, 64, 128), residual blocks, and 0.1 dropout. Training uses DiceCELoss with Adam at learning rate 1e-3.

Inference uses sliding-window prediction. VTK then extracts and smooths surfaces and exports:

• lung.stl
• lung.vtp

result

A full run over all 20 CT cases trained for 25 epochs produced:

• mean lung Dice: 0.9243
• mean lung HD95: 10.5909
• best validation Dice: 0.9539

The first half of the dataset is near-perfect, while a weaker subgroup appears in the second half. That points to structured case difficulty rather than random optimization noise.

limitations

• The reported top-line result averages predictions over all 20 prepared cases, not a strict external test set
• Compact resizing sacrifices native scanner resolution
• The current split protocol should be hardened into an explicit held-out evaluation
• The target is binary lung, not airway, vessel, lobe, or lesion anatomy
• The generated meshes are suitable for engineering demonstration but are not validated for procedural planning

next steps

• harden the held-out evaluation protocol
• inspect the weak subgroup of CT cases directly
• extend the pipeline toward navigation-specific anatomy such as airways or lobes

reproducibility

Install dependencies:

python3 -m pip install -r requirements.txt
python3 -m pip install -e .

Run the documented full pipeline:

python3 scripts/run_zenodo_lung_pipeline.py \
  --workspace outputs/zenodo_lung_full_e25 \
  --epochs 25 \
  --target-depth 96 \
  --target-height 128 \
  --target-width 128