AI increasing accuracy and efficiency in cancer screening

Dr Helen Frazer is leading the research at St Vincent’s Hospital in Melbourne using AI to improve the accuracy of breast cancer screening and shorten the time to return results to patients.

St Vincent’s Hospital in Melbourne has partnered with BreastScreen Victoria, St Vincent’s Institute of Medical Research, the Australian Institute of Machine Learning at the University of Adelaide and the University of Melbourne to support the research program, and its being overseen by Aikenhead Centre for Medical Discovery.

Despite the screening program being largely considered as a successful public health initiative over its 30 year term, Frazer believes that AI can not only save lives in this space, but improve the experience for women.

“We're now working with five of our top performing algorithms and their ensemble and the exciting thing is that we've moved into this real world setting. So we've been testing and validating our models in a real-world retrospective cohort of women of over half a million women,” she said.

“We’re also working in a digital twin environment where we're prospectively testing those models in real time as women come into the screening pathway.”

Frazer said that the research program is focusing on data central and model centric development to train and test the models.

“We have over 4.2 million mammogram images and the associated non-image data for that image. So we know that the words age, whether she has a family history, has she had hormone replacement therapy? We also know whether the mammogram was abnormal, whether the woman came back to assessment, a biopsy performed, a cancer diagnosed, the cancer removed and the surgical histopathology,” said Frazer.

The biggest challenge that they are facing is generalisability she said.

Some of the issues here include how the program performs in other screening jurisdictions, using other machine vendors.

“We've got many vendors. Often mammography imaging, for instance, does the algorithm that's trained on one vendor, does that generalise to another one? But also not only within the Australian population, but can our solutions actually generalise or work with the same accuracy in global export markets?”

Building trust at all layers is another challenge, both for patients to trust the technology but also for clinicians to trust the outputs from the system she said.