Dataset Viewer
Auto-converted to Parquet Duplicate
source
stringclasses
1 value
split
stringclasses
1 value
projections
dict
angles_deg
stringclasses
1 value
geometry
dict
gt
stringclasses
1 value
LUNA16 series 1.3.6.1.4.1.14519.5.2.1.6279.6001.154837327827713479309898027966 (LIDC-IDRI), CC BY 3.0
reserved test (never seen in training)
{ "shape": [ 180, 384, 768 ], "dtype": "float16", "note": "DRR line integrals divided by the global max (scale below); the model normalises by the global max anyway", "scale": 1067728.25 }
gantry angle per projection, training convention, 0..358 step 2
{ "H": 384, "W": 768, "psize_mm_at_iso": 0.7273, "SAD_mm": 1000, "volume": [ 256, 256, 128 ], "voxel_mm": 2 }
gt_hu.nii.gz int16 HU, 256x256x128 @ 2 mm

EUReCA demo case

One simulated cone-beam acquisition for trying the EUReCA CBCT reconstruction models without any data preparation.

  • Source CT: LUNA16 / LIDC-IDRI series 1.3.6.1.4.1.14519.5.2.1.6279.6001.154837327827713479309898027966 (thorax), CC BY 3.0. This case is in EUReCA's reserved test split and was never used for training or model selection.
  • Geometry: Varian Halcyon on-board imager as simulated for training: SAD 1000 mm, detector 384 × 768 pixels at 0.7273 mm (isocenter plane), volume 256 × 256 × 128 voxels at 2 mm.

Files (luna16_27966/)

file content
projections.npy (180, 384, 768) float16, DRR line integrals at gantry angles 0, 2, …, 358°, divided by their global maximum (scale in meta.json)
angles.npy (180,) float32 gantry angles in degrees, training convention
gt_hu.nii.gz ground-truth CT resampled to the 2 mm reconstruction grid, int16 HU
aug.ini resampling transform used (identity)
meta.json provenance, geometry and scaling

Usage

git clone https://github.com/Jiening666/EUReCA.git && cd EUReCA
pip install -r requirements.txt
python download_weights.py --model eureca_main --demo
bash demo/run_demo.sh eureca_main 20        # -> PSNR / SSIM against gt_hu.nii.gz

The models expect views drawn from the arc they were trained on (a 211° arc for eureca_main); demo/run_demo.sh selects them from this 360° library. Any subset of views and angles can be passed to test_npy.py --view_indices.

License and attribution

Derived from LUNA16 (https://luna16.grand-challenge.org/), which is based on LIDC-IDRI, distributed under CC BY 3.0. Please cite LUNA16 and LIDC-IDRI when using this case: Setio et al., Medical Image Analysis 42 (2017); Armato et al., Medical Physics 38 (2011).

Downloads last month
56