An automated computational geometry pipeline that reconstructs watertight 3D triangle meshes (STL) from volumetric CT segmentation masks using Marching Cubes, Laplacian smoothing, and topological repair.
Reconstructed thoracolumbar spine (T10-L5) validated in 3D Slicer (v5.12.4), demonstrating surface continuity, anatomical curvature, and watertight geometry.
Medical deep learning models typically output volumetric voxel masks. However, surgical navigation systems, patient-specific implant CAD design, and additive manufacturing (3D printing) require continuous, watertight boundary representations.
BoneForge3D bridges this gap by converting multi-class vertebral CT segmentation masks into production-ready STL surface models. The pipeline accounts for physical voxel anisotropy, performs isosurface extraction via Marching Cubes, applies Laplacian smoothing to eliminate voxel discretization staircases, computes clinical morphometric properties (volume, surface area, compactness), and validates reconstructed meshes in 3D Slicer.
- Anisotropic voxel scaling: scales vertex coordinates by physical slice zooms (e.g. 0.68 x 0.68 x 0.90 mm) to maintain anatomically accurate aspect ratios.
- Isosurface extraction: runs Marching Cubes (Lorensen & Cline) to generate triangular mesh surfaces from discrete segmentation labels.
- Surface cleanup: applies Laplacian smoothing, removes duplicate or degenerate faces, and closes boundary holes to produce watertight 2-manifold STL models.
- Morphometric measurements: calculates physical volume (mm³) with the Divergence Theorem, surface area (mm²), and compactness.
- Clinical software export: writes binary STL files formatted for direct import into 3D Slicer and surgical CAD packages.
Sagittal, Coronal, and Axial cross-sections of patient sub-verse500 from the VerSe2020 dataset:
Reconstructed Lumbar 3 vertebra showing the anterior load-bearing vertebral body, transverse and spinous processes, and central spinal canal:
+------------------------------+
| Patient CT Scan (NIfTI) |
+--------------+---------------+
|
v
+------------------------------+
| Volumetric Segmentation Mask | <- Multi-class labels (T10 to L5)
+--------------+---------------+
|
v
+------------------------------+
| Physical Coordinate Spacing | <- Voxel zooms: 0.68 x 0.68 x 0.90 mm
+--------------+---------------+
|
v
+------------------------------+
| Marching Cubes Extraction | <- Discrete voxels to triangle mesh
+--------------+---------------+
|
v
+------------------------------+
| Topological Mesh Repair | <- Laplacian smoothing & hole filling
+--------------+---------------+
|
+------+------+
| |
v v
+--------------+ +------------------------------+
| STL Export | | Computational Morphometry |
| (T10 to L5) | | Volume (mm³), Area, Sphericity
+-------+------+ +--------------+---------------+
| |
v v
+--------------+ +------------------------------+
| 3D Slicer | | vertebra_geometry.csv |
+--------------+ +------------------------------+
Metrics computed from VerSe2020 patient sub-verse500:
| Vertebra | Spinal Region | Vertices | Faces | Volume (mm³) | Surface Area (mm²) | Compactness | Watertight |
|---|---|---|---|---|---|---|---|
| T10 | Lower Thoracic | 30,394 | 60,788 | 45,879.9 | 12,131.5 | 0.5109 | Yes |
| T11 | Lower Thoracic | 30,747 | 61,498 | 48,673.4 | 12,337.5 | 0.5225 | Yes |
| T12 | Thoracolumbar Junction | 34,135 | 68,274 | 57,100.5 | 13,624.3 | 0.5263 | Yes |
| L1 | Upper Lumbar | 39,393 | 78,798 | 66,201.5 | 16,016.3 | 0.4941 | Yes |
| L2 | Mid Lumbar | 41,835 | 83,674 | 70,434.8 | 17,159.7 | 0.4807 | Yes |
| L3 | Mid Lumbar | 44,716 | 89,432 | 73,955.5 | 18,198.2 | 0.4682 | Yes |
| L4 | Lower Lumbar | 42,878 | 85,756 | 72,177.2 | 16,927.1 | 0.4953 | Yes |
| L5 | Lumbosacral Base | 51,125 | 102,286 | 83,091.1 | 19,046.0 | 0.4835 | Yes |
- Load-Bearing Volume Gradient: Bone volume scales from 45,880 mm³ (T10) to 83,091 mm³ (L5), representing an 81.1% increase in bone volume along the inferior spinal axis to accommodate cumulative body weight.
- Compactness Ratio: Sphericity values cluster tightly between 0.46 and 0.53, reflecting the consistent ratio of the cylindrical vertebral body to the extended posterior articular processes.
BoneForge3D/
├── README.md # Technical documentation and results
├── requirements.txt # Pinned package dependencies
├── notebooks/
│ └── boneforge3d_reconstruction.ipynb # Full reconstruction workflow
├── src/
│ ├── __init__.py # Package interface
│ ├── mesh_extractor.py # Marching Cubes and Laplacian smoothing
│ ├── geometry_analysis.py # Morphometric metrics calculation
│ └── visualizer.py # Orthogonal and 3D rendering utilities
├── results/
│ ├── 3d_slicer_spine_reconstruction.png # 3D Slicer verification render
│ ├── segmentation_views.png # Orthogonal cross-sections
│ ├── l3_reconstruction.png # Isolated L3 3D mesh render
│ └── vertebra_geometry.csv # Quantitative geometric metrics
└── sample_meshes/ # Exported watertight STL models
git clone https://github.com/username/BoneForge3D.git
cd BoneForge3D
pip install -r requirements.txtimport nibabel as nib
from src.mesh_extractor import extract_and_clean_mesh
from src.geometry_analysis import compute_mesh_metrics
# Load segmentation mask
seg_nib = nib.load("sub-verse500_dir-ax_seg-vert_msk.nii.gz")
seg_data = seg_nib.get_fdata()
spacing = seg_nib.header.get_zooms()
# Reconstruct vertebra L3 (Label 22)
mesh = extract_and_clean_mesh(
seg_volume=seg_data,
label_id=22,
voxel_spacing=spacing,
output_path="sample_meshes/vertebra_L3.stl"
)
# Compute geometric properties
metrics = compute_mesh_metrics(mesh, name="L3")
print(f"Volume: {metrics['volume_mm3']} mm³ | Watertight: {metrics['is_watertight']}")jupyter notebook notebooks/boneforge3d_reconstruction.ipynbThe exported STL files can be loaded into 3D Slicer (v5.0+) or any CAD/3D-printing software:
- Open 3D Slicer.
- Click File -> Add Data and select the exported
.stlfiles. - Verify surface manifold integrity, boundary smoothness, and relative intervertebral alignment.
MIT License


