Re-segmentation guidelines#
Re-segmentation selects which voxels inside a region of interest (ROI) contribute to intensity and texture analysis. Apply it after resampling and before discretization. It can exclude voxels outside a selected intensity range or remove intensity outliers.
Which mask changes?#
Z-Rad follows the Image Biomarker Standardisation Initiative (IBSI) distinction between two masks:
Mask |
Purpose |
Effect of re-segmentation |
|---|---|---|
Morphological mask |
Defines the ROI shape used for morphology. |
Unchanged. |
Intensity mask |
Selects voxels for intensity and texture analysis. |
Voxels can be removed, leaving holes or disconnected regions. |
Intensity statistics, intensity histogram, intensity-volume histogram (IVH), and most texture features use the intensity mask. Grey level distance zone matrix (GLDZM) features use both masks: the morphological mask defines distances to the ROI boundary. See Feature extraction concepts for the feature families.
Range re-segmentation#
Range re-segmentation keeps voxels whose image intensities fall within the
selected lower and upper bounds, including both endpoints. For example, a
CT protocol might use [-50, 150] Hounsfield units (HU), while a PET protocol
might use a lower standardized uptake value (SUV) threshold such as
[3, infinity). These are examples, not default settings for every study.
Use a range that is meaningful for the modality and analysis:
CT and PET have calibrated units, so choose and report a range appropriate to the tissue and study protocol.
Raw MRI intensities depend on acquisition and scanner settings. Use a common range only when the intensity scale has been standardized and the range can be justified.
In Z-Rad, range selection uses the original image supplied for extraction, even when a filtered image supplies the intensities for feature calculation. This is one reason filtered-image extraction also requires the original image.
Outlier removal#
Outlier removal uses the mean and standard deviation of the valid intensity
values inside the ROI. For example, a setting of 3 keeps values within
mean - 3 * standard deviation and mean + 3 * standard deviation.
The accepted interval therefore depends on each ROI’s intensity distribution.
For filtered-image extraction, these statistics use the filtered intensities.
When both methods are enabled, Z-Rad applies the range first and calculates outlier statistics from the remaining voxels:
Keep voxels whose original image intensities fall within the selected range, for example
[-50, 150]HU.Calculate the mean and standard deviation of the retained intensity-mask values.
Remove values outside the selected standard-deviation interval.
The final intensity mask contains only voxels accepted by both rules. Check that enough voxels remain for extraction; see Feature extraction concepts for mask-size requirements.
Configure re-segmentation#
In the GUI’s Radiomics tab, use Intensity Range and Outlier Removal.
In Python, apply Resegmenter after IntensityMaskBuilder and before
TextureDiscretizer or IVHIntensityDiscretizer. Re-segmentation clears
previously prepared texture and IVH images because the intensity population
has changed. See Python image workflows for a complete pipeline.
A configured intensity range also affects bin origins and intensity ranges used during discretization; see Discretization guidelines.
What to report#
Record these settings with the extracted features:
whether you used range re-segmentation, outlier removal, both, or neither
the intensity bounds and units, including whether the upper bound was finite
the standard-deviation multiplier for outlier removal
that re-segmentation followed resampling and, when both methods were used, range selection preceded outlier statistics
whether feature intensities came from the original or a filtered image
that re-segmentation changed the intensity mask while retaining the morphological mask