An eye exam paired with artificial intelligence (AI) machine learning technology may provide a novel approach to diagnose Parkinson disease (PD) in early stages, according to research being presented at the annual meeting of the Radiological Society of North America (RSNA).
Characterized by the decay of dopaminergic cells in the substantia nigra, researchers highlight that the current standard for diagnosing PD, which is typically based on symptoms like tremors, muscle stiffness, and impaired balance, occurs once 80% of dopaminergic cells have already decayed.
“The issue with that method is that patients usually develop symptoms only after prolonged progression with significant injury to dopamine brain neurons,” said lead study author Maximillian Diaz, a biomedical engineering PhD student at the University of Florida (UF), in a statement. “This means that we are diagnosing patients late in the disease process.”
While the degradation of these cells may be often associated with the exacerbation of motor issues as well as non-motor issues such as pain and depression, this occurrence has also been shown to create thinning of the retina walls, the layer of tissue that lines the back of the eyeball. In fact, a prior study indicated that patients with PD were more likely to experience vision and eye issues, such as blurry vision, dry eyes, and trouble with depth perception.
Researchers saw an opportunity to leverage the capabilities of AI to examine images of the eyes for signs of PD. Utilizing a type of AI called support vector machine (SVM), researchers derived pictures of the back of the eye from both patients with PD and control participants from 2 age and gender matched data sets.
The first data set included 476 fundus eye images (PD = 238; control = 238) from the UK Biobank (UKB), with the second data set consisting of 100 images (PD = 72; control = 28) provided by UF and 44 control images from UKB. Researchers trained the SVM to detect signs on the images suggestive of PD.
A second set of datasets, UKB-Green and UF-UKBGreen, were created using the green color channels to improve vessel segmentation, which was performed using U-Net segmentation network. “The vessel maps served as inputs to SVM classifying networks,” explain the study authors. “Saliency maps were created to assess areas of interest for the networks.”
In their findings, the top performing SVM network for the UKB and UKB-Green data sets were the sigmoid SVM networks, achieving accuracies of .698 and .719 respectively, whereas the top performing networks for the UF-UKB and UF-UKB-Green data sets were the linear SVM networks, achieving accuracies of .821 and .857.
Expanding on the ability of machine learning networks to classify PD based on retina vasculature, study authors found that smaller blood vessels provided the most successful outcomes.
As traditional imaging approaches with MRI, CT, and nuclear medicine techniques may prove costly, Diaz notes that images can even be captured by a smartphone with a special lens.
“The single most important finding of this study was that a brain disease was diagnosed with a basic picture of the eye,” said Diaz. “If we can make this a yearly screening, then the hope is that we can catch more cases sooner, which can help us better understand the disease and find a cure or a way to slow the progression.”
Diaz M, Tian J, Ramirez-Zamora A, Fang R. Machine learning for Parkinson Disease diagnosis using fundus eye images. Presented at RSNA 2020; November 29–December 5, 2020. Abstract: IN-1A-07.