Why image processing still matters before a deep model sees the scan

Deep networks are often presented as if they replace classical image processing. In medical imaging they do not. Skin lesions, bone-marrow fields, ultrasound, and brain scans arrive with sensor noise, uneven illumination, overlapping intensities, and compression artefacts. If those effects are ignored, a model can latch onto acquisition quirks instead of pathology.

The first questions I ask of an image set are practical. What is the bit depth? Was the colour space standardised? Are there empty borders, burned-in overlays, or scanner-specific palettes? Histogram matching, denoising, and modest contrast normalisation are not cosmetic steps. They reduce the chance that a later convolutional or transformer stage treats a lighting change as a disease cue.

This is especially visible in noisy neuroimaging, where overlapping grey-level intensities make tumour boundaries hard to read, and in dermoscopic photography, where hair, gel, and camera white-balance vary from clinic to clinic. Feature engineering still has a place when the sample size is small: colour statistics, texture, and lesion geometry can support a learned backbone rather than compete with it.

The rule I use in teaching and in system design is simple. Do not ask a network to invent invariance that a short, documented preprocessing chain can provide. Write down every transform, keep a copy of the raw file, and check that the same pipeline is applied at training and at inference.