Open to misinterpretation (Image: Mehmed Zelkovic/Getty)
IT’S a case of diagnosis without doctors. Software could soon be working out what’s wrong with you based only on medical data.
Advertisement
Machines have already transformed healthcare. MRI scanners can peer inside the body, for example, and blood samples are analysed automatically, but human skill has always been an integral part of the process: a scan reveals a shadow – the oncologist recognises its significance. But doctors are often busy and overworked; they can make mistakes or . If computers could understand health on their own terms, perhaps they could speed up diagnosis and even make it more accurate.
Central to the new approach are advances in machine learning – the way software can be trained to recognise important features in an image, for example. It’s a powerful tool, but generally the software requires a lot of virtual hand-holding: images might have to be carefully aligned, and human experts are needed to make sure the software is trained to recognise the right features.
Deep learning is more flexible. Here, the software works at multiple levels simultaneously. Given a simple image, the computer might process the edges and lines while also considering what the image as a whole portrays – “dog” or “cat”, say. The approach means deep learning can make inferences about sets of data containing quite different concepts without human guidance.
This could be a game-changer for medicine, says at the University of Queensland in Australia. “[It can] readily combine images obtained from multiple views and multiple modalities,” he says.
Take breast cancer detection. Diagnosis potentially requires information from three sources: an X-ray, an MRI scan and ultrasound – and cross-referencing is laborious and time-consuming. Not with deep learning. Bradley and colleagues have a prototype system that cross-references automatically. They will present it in October at the International Conference on Medical Image Computing and Computer Assisted Intervention in Munich, Germany.
“Deep learning software automatically cross-references information from several sources”
Researchers at Tel Aviv University in Israel have been using deep learning to . So far, their system can distinguish between enlarged hearts and fluid build-up around the lungs.
Meanwhile, a group at the National Institutes of Health Clinical Center in Bethesda, Maryland, is using similar methods to . Both groups are getting results that equal or better existing state-of-the-art detection algorithms.
But will doctors – or patients – ever accept the word of a machine? That remains a problem, says Bradley. Deep learning’s complex networks are inscrutable, spitting out conclusions without giving reasons.
For instance, if you have ever had Facebook suggest you tag someone you don’t know as one of your friends, not even a Facebook engineer could tell you why that happened. Apply that level of mystery to medicine and, understandably, people may well get uneasy.
“Give them a black box? The clinicians are never going to embrace that,” Bradley says. That’s why he has a second system. Once the deep neural network is trained, Bradley uses its outputs to train another, transparent model – a “white box” whose answers humans can inspect and understand, and which will fail in certain circumstances. “In traditional systems the expert will build in sanity checks,” he says. “Is the thing the right size, colour, place? If not then don’t go further.”
We can expect deep learning to have an impact on medicine, says Brendan Fray at the University of Toronto, Canada, particularly with the rise of personalised healthcare and the focus on genes. His startup, Deep Genomics, is bringing the approach to genetic analysis.
“Deep learning will transform personalised medicine, genetic testing and pharmaceutical development,” he says. “It provides the glue between data and medical outcomes.”
Leader: “Smart machines may discover things we can’t, but we still matter”
This article appeared in print under the headline “Deeply healthy”
