Deep-Learning Machine Uses MRI Scans to Determine Your Brain Age
Determining brain age from an MRI scan has always been a time-consuming business. Now an AI machine gives the answer in seconds
Human cognitive abilities decline with age. And neuroscientists have long known that this decline correlates with anatomical changes in the brain as well. So it’s no surprise to learn that it is possible to spot the signs of aging in MRI images of the brain and even to determine a “brain age.” The difference between brain age and chronological age can reveal the onset of conditions such as dementia.
But the analysis is lengthy because the MRI data has to be heavily processed before it is suitable for automated aging. This pre-processing includes the removal from the image of non-brain tissue such as the skull, the classification of white matter, gray matter, and other tissue, and the removal of image artefacts along with various data-smoothing techniques.
All this data crunching can take more than 24 hours, and that is a serious obstacle for doctors hoping to take into account a patient’s brain age when making a clinical diagnosis.
Today, all that changes thanks to the work of Giovanni Montana at King’s College London and a few pals who have trained a deep-learning machine to measure brain age using raw data from an MRI scanner. The deep-learning technique takes seconds and could give clinicians an accurate idea of brain age while the patient is still in the scanner.
The method is a standard deep-learning technique. Montana and co use MRI brain scans of over 2,000 healthy people between 18 and 90 years old. None had any kind of neurological condition that might influence their brain age. So their brain age should match their chronological age.
Each scan is a standard T1-weighted MRI scan of the type produced by most modern MRI machines. Each scan is labeled with the chronological age of the patient.
The team used 80 percent of these images to train a convolutional neural network to determine a person’s age, given their brain scan. They used a further 200 images to validate this process. Finally, they tested the neural network on 200 images it hadn’t seen to determine how well it could measure brain age.
At the same time, the team compared the deep learning approach to the conventional method of determining brain age. This requires extensive image processing to identify, among other things, white matter and gray matter in the brain followed by a statistical analysis called Gaussian process regression.
The results make for interesting reading. Both deep learning and Gaussian process regression accurately determine the chronological age of patients when given preprocessed data to analyze. Both methods do this with an error of less than five years either way.
However, deep learning shows its clear superiority when analyzing raw MRI data, where it performs just as well, giving the correct age with a mean error of 4.66 years. By contrast, the standard method of Gaussian process regression performs poorly in this test, giving a rough age with a mean error of almost 12 years.
What’s more, the deep-learning analysis takes just a few seconds compared with the 24 hours of pre-processing required for the standard method. The only data processing required for the deep-learning machine is to ensure the consistency of the image orientation and the voxel dimensions between images.
That has significant implications for doctors. “Given the right software implementation, brain-predicted age data could be made available to a clinician while the patient is still in the scanner,” say Montana and co.
The team also compared images taken using different scanners just to show that the technique can be applied to scans taken on different machines in different parts of the world. They also compare brain ages of twins to show how brain age is linked to genetic factors. Interestingly, the correlation drops with age, suggesting that environmental factors become more significant as time goes on, and suggesting a promising line of future research.
That’s an impressive result that has the potential to significantly influence the way clinicians come to a diagnosis. There is considerable evidence that conditions such as diabetes, schizophrenia, and traumatic brain injury are correlated with faster brain aging. So a way to measure brain aging quickly and accurately could have an important impact on the way clinicians deal with these conditions in the future. “Brain-predicted age represents an accurate, highly reliable, and genetically valid phenotype that has potential to be used as a biomarker of brain aging,” say Montan and co.
Ref: arxiv.org/abs/1612.02572: Predicting Brain Age with Deep Learning from Raw Imaging Data Results in a Reliable and Heritable Biomarker