Machine learning helps to assess brain atrophy

Scientists from CSIRO and the Queensland University of Technology have used artificial intelligence to develop a ‘world-first’ benchmark for measuring brain atrophy in neurodegenerative diseases, including Alzheimer’s disease.

Alzheimer’s is the most common form of dementia, accounting for 60–80% of cases. One of the ways of measuring its progress is via magnetic resonance imaging (MRI) that shows cortical thinning, which can start up to 10 years before clinical symptoms of Alzheimer’s appear. Such measurement is challenging, however, as changes in the thickness of the brain’s cortex are extremely small, often in the submillimetre range.

Advanced machine learning techniques are routinely used in brain research to assess changes in cortical thickness, but until now, a lack of a clinically accurate ‘ground truth’ dataset meant scientists could not evaluate their sensitivity to the detection of small atrophy levels. The only way to get a ground truth measure of cortical thickness was by studying the brain post-mortem, but as brains begin to shrink immediately after death, even this resulted in inaccurate readings.

“Extremely accurate methods are needed to observe these signs in brain images when they begin to appear so they can be addressed earlier rather than later,” said CSIRO research scientist Dr Filip Rusak.

“Using the power of machine learning, we were able to produce a set of artificial MRI images of brains with predefined signs of neurodegeneration in the cortex region, the outer layer of the brain most affected by Alzheimer’s.

“Before these findings, there was no way to conclusively determine the sensitivity of the various methods used to measure cortical thickness in Alzheimer’s patients.”

The new technique allows researchers to set the amount and location of brain degeneration they want to compare against so they can get a clear picture of what method of cortical thickness quantification performs the best. It can test the sensitivity of methods to a minuscule level, determining whether a method can detect changes in thickness of just 0.01 mm.

The research has been published in the journal Medical Image Analysis, and has already had international impact. Michael Rebsamen from the University of Bern said his team had strong evidence that DL+DiReCT — a deep learning-based method for measuring cortical thickness — is robust and sensitive to subtle changes in atrophy, but until now they could not quantify what level of atrophy could truly be measured.

“The innovative benchmark from CSIRO closes this gap and marks an important milestone for evaluating cortical thickness methods,” Rebsamen said.

The technique can be applied to research in any brain disease that involves neurodegeneration, representing a significant step forward to better understanding dementia and other debilitating brain diseases. It can also potentially be used to predict the level of cortical degeneration a person can expect over time — and all off the back of commonly used and relatively inexpensive MRI images.

“The findings will help researchers pick the right tools for the job. The right tool increases the chances of accurately assessing disease progression,” Rusak said.

“So there’s no need for new medical infrastructure.”

The synthetic dataset images have been made publicly available so clinicians and scientists can use the synthetic images to conduct their own assessments of cortical thickness quantification methods. They are available at https://data.csiro.au/collection/csiro:53241v1.

Image credit: iStock.com/mr.suphachai praserdumrongchai