Radiomics features¶

pydcm.radiomics computes the full IBSI feature set over an ROI — 135 features across 10 classes (first-order, GLCM, GLRLM, GLSZM, GLDM, GLDZM, NGTDM, shape, intensity-volume histogram, local intensity), with the standard radiomics feature names. The seven classes a conventional extractor also computes are validated feature-for-feature against published reference baselines; GLDZM / IVH / local-intensity add the IBSI features beyond that set, so the output drops into an existing radiomics pipeline unchanged and then some.

135 IBSI features, 145 dict keys

The 135 are the IBSI-defined features. The returned dict carries a few more keys (≈145) — first-order and shape convenience extras (e.g. Energy / TotalEnergy / RMS) on top of the IBSI set — so len(feats) is larger than 135.

From a DICOM image + mask¶

from pydcm.radiomics import radiomics

feats = radiomics("ct.dcm", mask="mask.dcm")     # pixels → HU, spacing from geometry
feats["firstorder_Mean"]                          # bare "<class>_<name>" keys
feats["shape_VoxelVolume"]
feats["glcm_Contrast"]

(The RadiomicsFeatureExtractor.execute path below prefixes keys with original_ — e.g. original_firstorder_Mean — matching the conventional extractor's key scheme.)

Pixels are decoded to real-world values (HU) and the spacing is read from the image geometry (PixelSpacing / SliceThickness), so the feature values are in physical units without any manual setup.

From arrays¶

import numpy as np

feats = radiomics(image_array, mask=mask_array, spacing=(z, y, x))

Binning and preprocessing knobs¶

All the preprocessing knobs are exposed as keyword arguments:

feats = radiomics(
    "ct.dcm", mask="mask.dcm",
    bins=32,                       # fixed bin count …
    bin_width=25.0,                # … or a fixed bin width (mutually exclusive)
    value_range=(-1024.0, 3071.0),
    resample=1.0,                  # isotropic resample (mm) before extraction
    distances=[1, 2, 3],           # GLCM neighbour distances
    normalize=False,
    wavelet=False, log_sigma=None, # optional image-filter families
)

Custom features¶

Register a Python feature with the @pydcm.radiomics.feature decorator — it runs over the same preprocessed + discretised grid the native IBSI features used, so a custom histogram feature lines up bin-for-bin with the standard ones, and its value joins the result dict as "<class>_<name>":

import pydcm, numpy as np

@pydcm.radiomics.feature("firstorder")           # → adds "firstorder_p90"
def p90(roi):
    return float(np.percentile(roi.intensities, 90))

# custom features run on the array / extractor path — decode to arrays first,
# not the all-DICOM-files shortcut
feats = pydcm.radiomics(image_array, mask=mask_array, spacing=(z, y, x))
feats["firstorder_p90"]

The roi argument exposes the exact grid the engine ran over — roi.intensities and roi.levels (1-D over the ROI voxels), roi.n, roi.spacing, roi.bin_edges. Naming a custom feature after a standard one (@pydcm.radiomics.feature("firstorder", name="Mean")) overrides that value, so you can change a built-in formula without recompiling. Manage the registry with pydcm.radiomics.registered_features() and pydcm.radiomics.clear_features().

Extractor object API¶

For code that expects the familiar extractor object, that API is available too — feature-class selection is honoured:

from pydcm.radiomics import RadiomicsFeatureExtractor

extractor = RadiomicsFeatureExtractor()
extractor.enableFeatureClassByName("firstorder")
result = extractor.execute("ct.dcm", "mask.dcm")