Skip to content

DATA.md

Latest commit

 

History

History
78 lines (59 loc) · 3 KB

DATA.md

File metadata and controls

78 lines (59 loc) · 3 KB
Raw

Data Preparation

Preprocessing is adapted from X2CT-GAN, with a dataset-specific COLONOG pipeline. Install the Python requirements and Plastimatch, then run the commands below from the repository root:

pip install -r requirements.txt
sudo apt-get install plastimatch

LIDC-IDRI

Use converted CT volumes named <case_id>.mhd and binary body masks named <case_id>_outMultiLabelMask.mhd. CT and mask must share the same voxel grid with identity/LPS orientation. .mha, .nii and .nii.gz are also supported.

python -m preprocessing.prepare_lidc \
  --ct-root /path/to/LIDC-volumes \
  --mask-root /path/to/LIDC-masks \
  --input-value-space hu \
  --output-root data/LIDC-HDF5-256 \
  --run

Use --input-value-space shifted_1024 for CT already stored as HU+1024.

CT intensities are shifted by +1024 and masked, resampled to approximately 1 mm spacing with cubic interpolation, then cropped or padded to 320³. The crop keeps the last 320 slices in z and centers y/x. This volume is used for projection and linearly resized to 256³ for CT storage, clipped to [0,2500].

COLONOG

Use CTSpine1K COLONOG volumes in physical HU, named <case_id>.nii.gz:

python -m preprocessing.prepare_colonog \
  --volume-root /path/to/CTSpine1K/raw_data/volumes/COLONOG \
  --output-root data/CTSpine1K-COLONOG-HDF5-256-2views-clean \
  --run

Volumes are reoriented to LPS and resampled around the body center within a 448³ mm field of view. A −500 HU body threshold, component filtering, hole filling and dilation remove background and table regions. CT and projection volumes are linearly resampled to 256³ and 320³, respectively. Intensities are clipped to [-1000,1500] HU and shifted by +1000.

X-ray Generation

Plastimatch generates orthogonal frontal (0°) and lateral (90°) DRRs with source-to-axis distance 541 mm, source-to-detector distance 949 mm and a 500 × 500 mm detector. Projection intensities use plastimatch adjust --pw-linear "0, -1000". Each DRR is normalized by its maximum, mapped to [0,255] and bilinearly resized to 256². The frontal image is horizontally flipped. Views are saved as xray1 and xray2.

The commands process the supplied train, validation and test lists in splits/. Use --split splits/lidc/train.txt to select a split, or --limit 1 to process one case. Omit --run to check input paths without generating data.

Output Format

Each case is saved to <dataroot>/<case_id>/ct_xray_data.h5:

Key Content
ct 256³ CT array in [z,y,x] order, stored as shifted intensities
xray1 256² frontal X-ray
xray2 256² lateral X-ray
spacing, ori_size Projection volume spacing and size

During training, CT values are clipped to [0,2500], divided by 2500 and resized to 128³ with trilinear interpolation. X-rays are divided by 255.

Source attribution and licenses: preprocessing/NOTICE.md.