Compare original and filtered-image radiomics

Extract the same features from an original CT image and from its 3D Mean-filtered version. Both runs use the same morphological ROI. Each feature image is discretized separately into 32 bins, so the GLCM values reflect both filtering and per-image binning.

from zrad.filtering import Mean
from zrad.image import Image
from zrad.preprocessing import IntensityMaskBuilder, RoiData, TextureDiscretizer
from zrad.radiomics import Radiomics

image = Image.from_nifti("path/to/phantom.nii.gz")
mask = Image.from_nifti_mask("path/to/mask.nii.gz", reference=image)
results = {}

for label, use_filter in (("Original", False), ("Mean filtered", True)):
    roi = RoiData(image=image, morphological_mask=mask)
    if use_filter:
        # The filter sets roi.filtered_image and leaves the original image and ROI in place.
        roi = Mean(padding_type="reflect", support=3, dimensionality="3D").apply(roi)
    roi = IntensityMaskBuilder().apply(roi)
    roi = TextureDiscretizer(number_of_bins=32).apply(roi)
    results[label] = Radiomics(aggr_dim="3D", aggr_method="AVER").extract_features(
        roi_data=roi,
        features=["stat_mean", "cm_contrast_3D_avg"],
        include_metadata=True,
    )

for label, features in results.items():
    print(label)
    print(f"  Mean: {features['stat_mean']:.2f} HU")
    print(f"  GLCM contrast: {features['cm_contrast_3D_avg']:.3f}")
    print(f"  ROI voxels: {features['no_voxels']}")

no_voxels is the same in both runs because the anatomical ROI is the same. Mean smoothing retains HU units, but its feature values describe the smoothed image. See Python image workflows for how RoiData carries the original and filtered images.