The detection of breast cancer at an early stage is of vital importance if the high mortality rate from this disease is to be reduced. The national breast-screening programme aims to use regular X-ray examinations amongst the age group most at risk to increase the early detection rate. A result of this is that over 1 million pairs of X-ray images per year will need to be reviewed by radiologists. Over 90% of these images will show no abnormality. Potentially, the successful application of image processing techniques to this task could greatly ease the burden of the very large numbers of examinations performed, and improve detection accuracy by directing the human expert's attention only to those results suspected of being abnormal. The problem is, however, for both technical and ethical reasons, formidable.