By Rich Haridy
Richard is based in Melbourne, Australia and has a strong interest in film, VR and new media. He has written for an number of online and print publications over the last decade and also acted as film critic for several radio broadcasters and podcasts. Richard was Chair of the Australian Film Critics Association for two years (2013-2015) and when not writing or making videos for New Atlas he can be found in darkened cinemas yelling at the screen.
A powerful new deep learning algorithm has been developed that can study PET scan images and effectively detect the onset of Alzheimer’s disease up to six years earlier than current diagnostic methods. The research is part of a new wave of work using machine learning technology to identify subtle patterns in complex medical imaging data that human clinicians are unable to pick up.
One of the clearer diagnostic tools we currently have to identify the onset of Alzheimer’s disease is a type of brain imaging scan called an 18-F-fluorodeoxyglucose PET scan (FDG-PET). This scan is traditionally used to identify several types of cancers but in recent years has proved itself useful in identifying Alzheimer’s, as well as several other types of dementia.
This new research trained a machine learning algorithm on over 2100 FDG-PET brain images. While human clinicians are adept at evaluating these scans, co-author on the new study Jae Ho Sohn says new deep learning technology has the ability to identify more subtle patterns in dense imaging data.
“Differences in the pattern of glucose uptake in the brain are very subtle and diffuse,” says Jae Ho Sohn. “People are good at finding specific biomarkers of disease, but metabolic changes represent a more global and subtle process.”
When the algorithm was ultimately tested on a small independent set of brain scans it was able to predict every single case that advanced to Alzheimer’s on average around six years earlier than the disease was eventually diagnosed. On this metric, the algorithm significantly outperformed human radiologists.
While other researchers have applauded these promising results, many are urging caution, suggesting much more work needs to be done to validate the results before it moves into clinical applications.
“This is a tiny data set, only looking at 40 people,” says John Hardy from University College London. “It’s also a very selected data set and not representative of the whole population. So we can’t know yet whether this is relevant to most people.”
It is also important to note that these kinds of PET scans are not broadly available to most patients. This kind of machine learning innovation is undeniably impressive in an academic context but it doesn’t offer up a useful tool for doctors hoping to better diagnose patients en masse.
“Currently in the UK, the use of PET scanning is mainly limited to research studies and clinical trials, to ensure that potential new medicines are tested in the right people,” explains Carol Routledge, from Alzheimer’s Research UK. “PET scans are a powerful tool, but they are expensive and require specialist facilities and expertise.”
However, this study does offer yet another indication that deep learning algorithms can greatly assist clinicians in wading through the growing mass of big data that is rapidly being accumulated. Alongside a recent study from McGill University revealing an algorithm that can evaluate an assortment of diagnostic data to predict the earliest stages of cognitive decline, this new research adds to a body of work that is recruiting computers to help us predict neurodegenerative disease before major symptoms take hold.
The new study was published in the journal Radiology.