Posted: 15th July 2020
Having a mammogram can be painfully unpleasant.
After the uncomfortable experience comes the anxious wait for a result to be returned.
In that time – about 2 weeks in Australia – the mammogram is ‘read’ by two experts; if their interpretations of the image disagree, then a third arbitrates.
And while mammogram screening is credited with saving countless lives, there is room for improvement. Even when a result comes back positive, in most cases this does not lead to an eventual cancer diagnosis. And more concerning is that mammograms miss about one in five breast cancers.
Holyoake Research Fellow, Dr Davis McCarthy, along with Dr Helen Frazer, Clinical Director at St Vincent’s BreastScreen, is hoping to improve the process.
“One in seven Australian women will develop breast cancer before the age of 85 and around 3,000 Australians die from the disease each year. The key to boosting survival rates is early and accurate detection,” says Davis.
Davis, Helen and her colleague Dr Peter Brotchie have been training a computer, using a machine learning, or ‘artificial intelligence’ algorithm, to improve the analysis and interpretation of mammograms.
The team’s current algorithm was trained to distinguish between ‘normal’ and ‘cancer’ samples from small patches of mammograms. The team then tested the model using previously unseen images. Their model has 88% accuracy, which is on par with human performance.
The researchers don’t imagine that they will be able to do away with human interpretation altogether: their vision is to replace one of the initial two reads of each mammogram with a read done by artificial intelligence. With more accurate results delivered more quickly, they hope to reduce the burden on the individual as well as on the health-care system, which wastes considerable resources following up innocent abnormalities.
Google recently made headlines in a paper showing that artificial intelligence could be more effective in spotting breast cancer than humans. The key difference in the work being done here is that the algorithm is being designed in the Australian setting – training with machines and images that are relevant to Australian women.
Davis and the team are about to embark on a ‘real world ‘ study where they run their algorithm alongside the current system on the approximately 240 scans that are done each day at the BreastScreen clinic, over a period of 3 months. Their aim is to obtain a thorough and honest assessment of the model’s performance.
Even more ambitious, they hope to develop the algorithm further so that it can provide an explanation of its prediction – for example, giving an annotated version of a mammogram with the cancerous region highlighted. They will go through iterations of this process until they are confident that the algorithm can be advanced to clinical use.
In this way, with computers aiding in the prediction of breast cancer, the experts hope that they can outperform themselves.