PPCNet Dataset

Biplanar DRRs + Projection Matrices + Ground-Truth Point Clouds

A curated lumbar spine dataset for 3D point cloud reconstruction from biplanar radiographs.

---

Overview

This dataset provides paired biplanar DRRs, calibrated 3×4 projection matrices, and dense ground-truth point clouds for 1,037 patients with complete L1–L5 lumbar vertebrae, derived from the publicly available VerSe'19, VerSe'20, and CTSpine1K collections.

---

📁 Dataset Structure

Each of the 1,037 patient folders contains:

File	Description
ct.nii.gz	CT volume (LPS orientation)
seg.nii.gz	Segmentation labels (L1=20, L2=21, L3=22, L4=23, L5=24)
gt_ppc.vtk	Ground-truth point cloud (5,120 points, world-mm)
AP_0/drr_AP_0.png	Anteroposterior DRR (512×512)
AP_0/P_AP_0.txt	3×4 AP projection matrix
LP_90/drr_LP_90.png	Lateral DRR (512×512)
LP_90/P_LP_90.txt	3×4 lateral projection matrix

Label mapping: L1=20, L2=21, L3=22, L4=23, L5=24

---

Data Split

Fixed split in dataset_split.json (seed=42):

Split	Patients	Usage
Train	829	Model training
Validation	103	Hyperparameter tuning
Test	105	Final evaluation
Total	1,037

---

DRR Generation

DRRs are generated using Plastimatch's ray-casting algorithm with the following parameters:

Parameter	Value
Source-to-Axis Distance (SAD)	1,000 mm
Source-to-Imager Distance (SID)	1,500 mm
Detector Size	500 × 500 mm
Image Resolution	512 × 512 px
Views	AP (0°) + Lateral (90°)
Bone Enhancement	2.5× for HU > 300
Post-processing	CLAHE normalisation

Each DRR comes with a calibrated 3×4 perspective projection matrix encoding the exact source–detector geometry.

---

Ground-Truth Generation

The ground-truth point clouds are constructed through a five-step pipeline:

1. Surface extraction — marching cubes on per-vertebra segmentation masks
2. Coordinate transform — vertices mapped to physical (mm) space via NIfTI affine
3. Concatenation — all five vertebrae (L1–L5) merged into a single cloud
4. Alignment — rigid ICP with four flip initialisations to match DRR projection space
5. Sampling — farthest-point sampling to a uniform 5,120-point representation

---

⬇️ Download

from huggingface_hub import snapshot_download

# Download full dataset (~69.2 GB)
snapshot_download(
    repo_id="ppcnet-dataset/PPCNet",
    repo_type="dataset",
    local_dir="./PPCNet_Dataset"
)

Or use the Hugging Face CLI:

pip install huggingface_hub
huggingface-cli download ppcnet-dataset/PPCNet --repo-type dataset --local-dir ./PPCNet_Dataset

---

Source Datasets

This dataset builds upon:

• 📦 VerSe — A Vertebrae Labelling and Segmentation Benchmark for Multi-Detector CT Images
• 📦 CTSpine1K — A Large-Scale Dataset for Spinal Vertebrae Segmentation in CT

---

License

This dataset is released under the MIT License.

---