Proposition 1III.7.P0171 of 76 in the corpus
A medical image is a measurement, and what it measures decides the pipeline.
CT reports absolute attenuation in calibrated units; MRI reports relaxation in arbitrary ones. The right preprocessing is a fact about the physics, not a default in a library.
Demonstration
X-ray measures attenuation integrated along a ray. It is a projection, so everything between the source and the detector is superimposed — a nodule can hide behind a rib, and no amount of contrast adjustment separates them. Pixel values have no absolute meaning.
CT is the same physics reconstructed into a volume. Its values are calibrated: Hounsfield units, water at 0 and air at −1000 by definition, so intensities are comparable across patients and scanners. This is why clinicians window by value — a lung window and a soft-tissue window are two fixed intervals — and why min-max normalising a CT per image destroys the property that makes it a measurement.
MRI measures proton relaxation, and the intensities are arbitrary. They depend on scanner, field strength, coil, sequence and session, and they are comparable only within a scan. Some per-volume normalisation is mandatory, and which one is a research question. The same anatomy in T1, T2 and FLAIR gives three different pictures; those are channels, and dropping one is dropping a measurement.
Fundus photography is reflected visible light. It is closest to an ordinary photograph, and it is the case where colour carries the diagnosis directly — which is what makes the augmentation warning of Chapter V concrete.
Two facts that cut across all of them.
Volumes have anisotropic spacing: 0.7 mm in plane and 5 mm between slices is ordinary, so a 3 × 3 × 3 kernel covers a very different physical volume along each axis. Resampling to a common spacing is why nnU-Net computes it from the dataset rather than accepting a default.
And a DICOM file carries the acquisition parameters. Reading them is the difference between processing images and processing measurements.
Corollary
Ask what physical quantity the numbers represent before choosing any transform. Half the preprocessing decisions answer themselves once that question is answered, and the other half become visible as decisions.