Medical Application: Computer Aided Mammography

Introduction

Breast cancer is the most common cancer amongst women in the Western world; it affects more than eight percent of women at some time in their lives. The success of treatment depends on early detection, before the cancer causes any symptoms. The most effective method for detecting early breast cancer is X-ray mammography, and in many countries, women at risk of developing the disease by virtue of age or family history are screened regularly for signs of the disease.

Computer-Aided Mammography

The UK breast screening program alone generates over 1.5 million mammograms per year, all of which must be carefully searched for signs of disease which are frequently small or subtle. Occasionally, cancers are missed, or normal women unnecessarily recalled for further investigation. For the last ten years, we have been working in close collaboration with radiologists from the Manchester Breast Screening Service to develop computer-based aids for radiologists interpreting mammograms, with the aim of improving the accuracy of the process.

We believe that prompting may be the key to improving radiologists' detection performance. In a prompting system, potential abnormalities are detected automatically, and their locations draw the radiologist's attention to suspicious areas of the film. Such a system appears to be effective only if the prompts provided by the computer are sufficiently accurate.

Research at the WIAU

Our ongoing research has focused on two main areas; the automated detection of potential abnormalities, and the investigation of the efficacy of prompting.

We have developed methods for detecting a number of the most important signs of abnormality including microcalcifications, which appear as clusters of small bright dots, spiculated lesions, which appear as ill-defined central masses with radiating linear structures, and asymmetry between the left and right breasts. Our methods are generally model-based, trained on example images drawn in sequence from those available at the breast screening centre. Expert radiologists annotate the films to provide a gold standard against which we can compare the results of our algorithms, and results are presented as receiver operating characteristic curves which plot the proportion of abnormalities detected (true positives) against the number of normal features incorrectly labelled (false positives).

Our prompting experiments have demonstrated that, for a single abnormality type (microcalcifications), it is possible to gain an improvement in detection performance using prompting. Current experiments are designed to determine the conditions under which this result holds good; for example, we know that there is no improvement when the false positive rate is high, but we do not yet know whether this is due to the increase in the total number of prompts, or due to the altered true positive prompt to false positive prompt ratio. These experiments use synthetic mammograms and abnormalities, generated using fractal and other techniques.

Other Applications

We are also developing a system to aid the training of radiologists. In order to maintain a sufficiently high level of skill, radiologists should ideally read several thousand mammograms per year. In some cases this is impractical, so we are fusing synthetic abnormalities generated by modelling real examples with normal backgrounds to produce multiple data sets of `new' images for training and testing purposes. These will be integrated into a computer-aided learning environment.

We have also been developing the technology to monitor the ways in which an individual's mammograms change over time. By measuring textural changes to the glandular tissue, and other quantitative changes between successive mammograms, it may be possible to quantify alteration in risk status caused by drugs (such as HRT) or due to aging. For this purpose we have developed a calibration system comprising a step-wedge and set of markers to measure breast thickness.

Acknowledgements

This work has received support from the EPSRC and the NHSBSP. We are grateful to Dr Caroline Boggis and her colleagues at the Nightingale Breast Screening Centre for their active and expert participation in this research.

Sue Astley: Sue.Astley@man.ac.uk

Research Courses Clinical Radiology Industrial liaison Contact Us Join Us Search Home

Top Page