A team of researchers at UC Davis and UC San Francisco have successfully found a way to program a computer to accurately detect one of the indicators of Alzheimer’s disease in human brain tissue, demonstrating the concept for a machine-learning approach to recognize critical markers of the disease. The machine-learning tool was developed by researchers at the University of California. The device can ‘see’ if a type of amyloid plaque is present in a sample brain tissue quickly using a similar mechanism that Facebook uses to recognize faces using captured images. Amyloid plaques are clumps of protein fragments that are typically found in the brains of people with Alzheimer’s disease that can harm and destroy nerve cell connections. The findings of the team have been published in Nature Communications.
The study suggests that machine learning can amplify the expertise and analytical skills of an expert neurologist. The tool allows them to analyze a thousand times more data than the most highly skilled human eyes. The study was led by Brittany N. Dugger, Ph.D., Assistant Professor, UC Davis Department of Pathology and Laboratory Medicine at UC Davis along with Michael J. Keiser, Ph.D., Assistant Professor, UCSF’s Institute for Neurodegenerative Diseases and Department of Pharmaceutical Chemistry. In their research, they attempted to determine if they could program a computer to automate the arduous process of identifying and analyzing small amyloid plaques of several types in large slices of human brain tissue that had been autopsied. Keiser and his team developed a “convolutional neural network” (CNN), a computer program designed to identify patterns by referring to thousands of human-labeled examples.
The UCSF team used their database of labeled example images to teach their CNN machine-learning algorithm to identify different types of brain changes observed in the disease, including the ability to differentiate between cored and diffuse plaques and identifying abnormalities in blood vessels. The algorithm could successfully process a complete whole-brain slice slide with nearly 98.7 percent accuracy, its speed depending on the number of computer processors used.