Alzheimer’s disease affects more than 5 million Americans, a number that’s expected to grow to 16 million by 2050. Early intervention can help to lessen the most debilitating symptoms, namely memory loss and problems with reading and organizing thoughts, but unfortunately, it remains one of the most challenging neurological disorders to detect. No specific test for Alzheimer’s exists.
There might be hope on the horizon, though. Researchers at the University of California at Berkeley’s Radiology & Biomedical Imaging Department and the Big Data in Radiology group (BDRAD, a multidisciplinary team of physicians and engineers studying radiology data science) describe in a newly published study an AI system that can predict Alzheimer’s disease from brain scans.
Their paper (A Deep Learning Model to Predict a Diagnosis of Alzheimer Disease Using 18F-FDG PET of the Brain”) was published in the journal Radiology this week.
“Differences in the pattern of glucose uptake in the brain are very subtle and diffuse,” Dr. Jae Ho Sohn, a coauthor, said. “People are good at finding specific biomarkers of disease, but metabolic changes represent a more global and subtle process.”
They aren’t the first to apply AI to Alzheimer’s disease research — scientists at Unlearn.AI, a startup that designs software tools for clinical research, earlier this year developed a system that predicts its progression. But the UC Berkeley team chose to focus on a chemical marker that hadn’t previously been used to train an AI model.
They trained a deep learning algorithm (layered mathematical functions that mimic the behavior of human neurons) on 18-F-fluorodeoxyglucose positron emission tomography (FDG-PET), a specialized imaging technique in which patients are injected with FDG, a radioactive glucose compound, which allows radiologists — and in this case, an AI system — to measure uptake in brain cells through PET scans. It’s an indicator of metabolic activity.
The researchers built a corpus from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), a project database containing more than 2,100 FDG-PET brain images from 1,002 patients. (About 10 percent of the samples were set aside for validation.) After training on the dataset, the AI system could track minute changes in glucose uptake in certain regions of the brain that would be, under normal circumstances, difficult to detect.
In tests on a separate set of 40 imaging exams from 40 patients, the researchers’ AI system achieved 100 percent sensitivity at detecting Alzheimer’s an average of more than six years prior to the final diagnosis.
“We were very pleased with the algorithm’s performance,” Dr. Sohn said. “It was able to predict every single case that advanced to Alzheimer’s disease.”
The team cautioned that it’s early days — the test sample size was relatively small. But they believe the system could be used to complement the work of radiologists, and that it could serve as the foundation for AI that can identify patterns of beta-amyloid and tau protein buildup, abnormal protein clumps, and other biological markers associated with Alzheimer’s disease.