- AI can enhance brain imaging techniques to predict Alzheimer’s years before diagnosis.
- Convolutional Neural Network developed in this study was able to identify brain scans (with 100% sensitivity) that lead to Alzheimer’s disease.
More than 5 million people in the United States have Alzheimer’s disease and this number is projected to reach 14 million by 2050. Every 65 seconds someone in the US develops the Alzheimer’s, which makes it the 6th leading cause of death in the country.
Typically, the disease is diagnosed when all symptoms have manifested, and by this time the loss of brain cells becomes so significant that it is too late to intervene. No therapy can stop or reverse the progression of Alzheimer’s, however, we can detect it in earlier stages to slow down its progress and improve symptoms.
Now, researchers at the University of California have described how artificial intelligence can enhance brain imaging techniques to predict Alzheimer’s years before diagnosis. The findings could help millions of patients and caregivers.
Deep Learning Analyzes Brain Metabolism
Previous studies showed that Alzheimer’s disease changes the brain metabolism: reduction in brain glucose metabolism is characteristically observed. However, recognizing these subtle changes can be an extremely challenging task.
In this study, researchers have applied deep learning method to detect changes in brain metabolism predictive of Alzheimer’s. They trained the method on thousands of images obtained from a nuclear medicine functional imaging technique called Positron Emission Tomography (PET).
They had access to data from a major multi-site project ADNI (short for Alzheimer’s Disease Neuroimaging Initiative) that focuses on preventing and treating this disease.
The dataset contained over 2,100 PET brain scans from more than 1,000 patients. They used 90% of this dataset to train their deep learning method and the remaining 10% was used to test it.
Examples of PET scans | Brain of a 76-year-old man (zoomed-in) with Alzheimer’s | Courtesy of researchers
They then tested the method on a new, independent group of 40 images from 40 patients that the algorithm had never examined. It was able to detect Alzheimer’s 6 years before (on average) the final diagnosis, on all 40 PET scans.
Researchers trained their convolutional neural network using NVIDIA TITAN Xp GPUs with CUDA Deep Neural Network library. It predicted every single scan (with 100% sensitivity) that lead to Alzheimer’s disease.
Even though the outcomes are very impressive, researchers cautioned that their validation set wasn’t large enough to make the algorithm fully reliable. They need more data to make this AI tool more mature.
At present, the tool can be used to complement radiologists’ jobs that include a wide range of imaging and biochemical tests. With large-scale external validation on multi-institutional data and model calibration, the tool may be integrated into clinical workflow to aid doctors with early predictions of the disease.
The team plans to further train their neural networks to recognize patterns linked with the abnormal accumulation of protein clumps, beta-amyloid and tau proteins, and other markers specific to Alzheimer’s disease.