Proposition 3III.1.P0352 of 76 in the corpus
Normalisation is a promise about the numbers the model will meet.
Subtracting a mean and dividing by a deviation fixes what counts as a typical intensity. The statistics must come from training data and be applied unchanged afterwards, or the promise is broken silently.
Demonstration
Raw pixel values arrive in whatever range the format uses — 0 to 255 for an 8-bit photograph, roughly −1000 to 3000 Hounsfield units for a CT. Feeding those directly gives large activations, large gradients, and a first layer that spends its early training just rescaling.
The standard fix is per-channel standardisation:
x ← (x − μ) / σThe important part is not the arithmetic. It is where μ and σ come from.
They are computed on the training set, stored, and applied unchanged to everything afterwards. Recomputing them per batch at test time is a leak: the prediction for one image then depends on which other images happened to be in the batch with it. The number goes up, the model is not better, and nothing raises an error.
The modality decides what is even meaningful.
CT is a measurement. Hounsfield units are absolute and comparable across patients and scanners — water is 0, air is −1000, by definition. So you window by value: a lung window and a soft-tissue window are two fixed intervals, and min-max normalising a CT per image destroys precisely the property that makes it a measurement.
MRI is not. Intensities are arbitrary and depend on the scanner, the sequence and the session, so they are comparable only within a scan. Some normalisation per volume is not optional, and which one is a research question in its own right.
Corollary
The phrase standard ImageNet normalisation names three numbers computed from photographs of objects on the internet. Applying them to a chest radiograph is not wrong exactly, but it is an assumption, and it is one worth making deliberately rather than by copying a snippet.