Saliency maps highlight the pixels that most affected a prediction. They are ubiquitous in medical imaging AI papers and vendor demonstrations because they look like an explanation a radiologist can evaluate.

Where This Gets Misread

Published work has shown saliency methods producing similar-looking maps for correct and incorrect predictions, and maps that change substantially with trivial input perturbations. A heat map centered on a real lesion is not proof the model used the lesion. Treat them as a debugging aid, not as evidence.