Machine learning analysis of eye scans may aid diagnosis of MS
Study finds algorithm detects slight changes that may be early MS signs
Using machine learning to analyze eye scans can help detect slight changes that may be early signs of multiple sclerosis (MS), potentially aiding in early diagnosis of the disease, a study found.
The study, “SLO-Net: Enhancing Multiple Sclerosis Diagnosis Beyond Optical Coherence Tomography Using Infrared Reflectance Scanning Laser Ophthalmoscopy Images,” was published in Artificial Intelligence.
MS is caused by inflammation in the brain and spinal cord. In most patients, the disease causes some amount of damage to cells in the eyes, and a building body of research is exploring whether eye-related changes may help diagnose MS.
Infrared scanning laser ophthalmoscopy (IR-SLO) is an imaging technique that uses a laser to create a two-dimensional image of the retina, a region at the back of the eyeball that houses light-sensing cells.
IR-SLO is often performed in conjunction with another type of eye imaging called optical coherence tomography (OCT), which generates a two-dimensional picture of the retinal layers. But while OCT has previously been explored as a potential tool to help diagnose MS, IR-SLO hasn’t been thoroughly studied for this purpose.
Different models for diagnosis of MS
A team of scientists in Iran explored whether applying machine learning to IR-SLO scans could help to identify MS.
In machine learning, a computer is given a dataset alongside a set of mathematical rules, or algorithms, that the computer uses to identify patterns in the data. The computer can then use these learned patterns to make sense of new data, such as distinguishing between people with a given condition and those without.
The scientists started with 265 IR-SLO scans obtained from 32 people with MS and 70 people without the disease in Iran. A subset of these scans was used to train a variety of machine learning algorithms, and the remaining scans were used to test the algorithms.
The best performing algorithm demonstrated more than 82% accuracy: It could accurately identify 83.1% of MS patients and 85.1% of those without the disease.
Similar tests using only OCT images yielded slightly better accuracy. However, using both IR-SLO and OCT scans in combination showed higher accuracy than either type of imaging on its own, reaching up to 96.9% accuracy.
“This study is the first work in which machine learning models were trained with two different imaging modalities for the diagnosis of MS,” the researchers noted.
The scientists further tested the machine learning tool on an external dataset of IR-SLO and OCT scans from people in the U.S., and again the tool showed good accuracy, correctly identifying 97.3% of MS patients and 84.6% of those without MS.
The study represents “a significant step toward automated and precise detection of MS using a non-invasive, low-cost, and easily accessible technology,” the scientists wrote, though they noted a need for additional studies with more patients to validate and further refine the tools.
About the Author
