AI-enabled radar: Pushing boundaries of glaucoma diagnosing, monitoring
Technology allows ophthalmologists to determine severitiy of disease.
This article was reviewed by Siamak Yousefi, PhD
Artificial intelligence (AI)-enabled radar may be an improvement over the traditional method of visual field analysis. Investigators note that by using AI-enabled radar they can better assess the functionality of patients with glaucoma and determine which patients have slowly and rapidly progressing visual field damage.
This technology is designed to address the shortcomings of traditional visual field analysis, specifically, i.e., that they rely on traditional paradigms such as linear regression and do not generate detailed results beyond progression or no progression, do not provide objective identification of progression, and lack advanced visualization and interpretation, according to Siamak Yousefi, PhD.
Related: Artificial intelligence in medicine: The good, the bad, and the scary
“We have proposed a glaucoma radar, a dashboard, that is a pipeline of linear and non-linear data transformations and unsupervised machine learning that provides advanced visualization with three layers of glaucoma knowledge-including the global visual functional severity, extent of visual functional loss in the hemifields, and local patterns of visual field loss,” explained Dr. Yousefi, assistant professor, Departments of Ophthalmology and Genetics, Genomics, and Informatics, University of Tennessee Health Science Center, Memphis.
“The system also provides personalized monitoring, data from thousands of previous glaucoma patients, and can identify rapid or slow progression,” he added.
Though the basis for this advanced-visualization technology is highly complex, the investigators sought to develop a screen for monitoring glaucoma that is user-friendly and able to be understood by non-medical individuals. It includes more than 13,000 visual fields cross-sectionally.
“We applied principle component analysis to linearly reduce the number of dimensions and extracted the global characteristics of the visual fields,” Dr. Yousefi said. “We then applied manifold learning to grab the local patterns of visual field loss and eventually generated this to a map.”
Related: Virtual reality and visual fields
Glaucoma data
To make sense of the data from the collected visual fields, the investigators first identified very dense areas, and then applied unsupervised clustering to identify 32 non-overlapping clusters that represented different levels of visual field severity.
According to Dr. Yousefi, when the mean deviation of each cluster was computed, the observation was that the severities of the visual fields increased moving from the top right to the bottom left of the screen.
“The AI pipeline was able to identify the global functional severity of the eyes in the various clusters, i.e., mostly normal eyes on the right and those with severe visual field loss on the left and bottom left of the screen,” he said.
The researchers also computed the glaucomatous severity in the inferior and superior hemifields.
“We noticed that the AI wisely put eyes with similar hemifield characteristics into different clusters, although statistically their global severity may have been similar,” noted Dr. Yousefi, indicating that the AI was able to differentiate similar damage in different locations and cluster those eyes together.
Related: The future: Artificial intellignece gains acceptance in ophthalmology
Testing technology
Investigators conducted a study in which they tested a method to identify and decompose visual fields to 17 different archetypes or patterns. They tried to decompose the visual
fields at each cluster, Dr. Yousefi explained, to the different archetypes and then identify the dominant pattern attributable to each cluster.
“Collectively, each region of the dashboard can demonstrate global visual field loss, the extent of visual field loss in the hemifields, and local patterns of visual field loss,” he said.
Because of the capabilities of the radar, the technology can be used for personalized glaucoma monitoring, retention of the history of previously monitored patients, and identification of rapidly and slowly progressing eyes.
Related: Predicting glaucoma progression: More than meets the eye
The availability of this information allows clinicians to easily adjust treatment regimens based on the glaucoma trajectory in a given patient’s eyes; clinicians will be able to adjust the treatment based on the cluster to which the patient is progressing and possibly positively impact the patient’s quality of life.
As a proof of concept, Dr. Yousefi and colleagues tracked eyes from benchmark data that indicated that the glaucoma was not progressing using 10 visual fields for each eye.
“As expected, most of the eyes did not have substantial changes in the visual fields,” he explained. “Some eyes that had visual field changes that were not in the direction of progression also were considered stable.”
Based on these data, Dr. Yousefi pointed out that the specificity of the radar was about 94%,” he said.
Related: Technology traces differences in structure, glaucoma progression
Investigators also performed the same analysis in eyes based on data from a separate algorithm that indicated that they eyes would progress. Most of these eyes showed glaucomatous changes in the direction of glaucoma worsening.
Dr. Yousefi concluded that with this method, ophthalmologists can perform better functional assessment of patients with glaucoma compared to the currently available tools.
“Advanced computational tools can provide more informative objective outcomes,” he said. “While deep learning typically generates black box and yes/no outcomes, our proposal may be helpful for developing models that can be interpreted more. Next-generation glaucoma assessment tools can provide multiple layers of glaucoma knowledge using advanced visualization.”
Read more by Lynda Charters
Siamak Yousefi, PhD
E: siamak.yousefi@uthsc.edu
Dr. Yousefi has received funding from the National Eye Institute.
