AI prep — volumes, NIfTI, raw pixels, resampling, radiomics¶

Read this when the task is feeding imaging into a model or computing quantitative features. Always dcmdeident first if the data left a PACS.

Tools¶

Need	Tool
3-D volume → NIfTI (+ BIDS sidecar)	`dcm2nii`
Raw decoded pixels (native ints, HU) for NumPy/PyTorch	`dcm2raw`
Resample to a target grid	`dcmresample`
Radiomics features → JSON	`dcmradiomics`
SEG mask → labelmap NIfTI	`seg2nii`

Recipes¶

Series → model input¶

dcmclient dcmdeident ./study/*.dcm --out-dir ./deid    # de-id first
dcmclient dcm2nii    ./deid ct.nii.gz                  # 3-D volume, geometry-correct
# dcm2nii also emits a BIDS .json sidecar where applicable (--bids)

Raw pixels (skip NIfTI, go straight to an array)¶

dcmclient dcm2raw --json scan.dcm pixels.raw   # result has dtype/shape/rescale

Pixels come out as native integers; pass the rescale slope/intercept (in the --json result) to get HU/SUV.

Resample to an isotropic grid¶

dcmclient dcmresample series-dir/ out.nii.gz --spacing 1,1,1   # mm
# or --size W,H,D for an explicit voxel grid

Radiomics over an ROI¶

dcmclient dcmradiomics ct.dcm --mask roi.dcm --json

Notes¶

dcm2nii keeps the true (possibly sheared) affine; orthogonalize on demand via dcmresample rather than expecting deskew as a separate step.
Geometry is single-sourced from the volume model matrix — don't hand-compute affines from IPP/IOP.