Diagnosing OSD with AI

Artificial intelligence deep learning algorithm makes accurate determinations by assessing OCT images

Cheryl Guttman Krader

Posted: Tuesday, September 1, 2020

A multi-disease predicting artificial intelligence (AI) deep learning algorithm shows excellent accuracy, sensitivity, and specificity for autonomous detection of both dry eye disease (DED) and Fuchs’ endothelial corneal dystrophy (FECD), reported researchers at the 2020 ASCRS Virtual Meeting
The AI software was developed by researchers at Bascom Palmer Eye Institute, University of Miami, Florida, USA. It processes images captured using an ultra-high resolution anterior segment optical coherence tomography (AS-OCT) platform (Envisu R2210, Bioptigen).
The findings from prospective evaluations of its performance for diagnosing the two ocular surface conditions were presented by Collin Chase BSc, and Taher K Eleiwa MD, MSc.
Detecting DED
The availability of an AI platform for diagnosing DED would be a valuable addition for clinicians, said Mr Chase.
“DED affects up to 25% of patients seen in ophthalmology clinics. Although there are currently a number of traditional tests and newer modalities used for its diagnosis, as noted by the members of the Dry Eye Workshop II, there is no gold standard,” he said.
“Artificial intelligence and deep learning networks can help us to discover patterns that exist within medical imaging but that are not obvious to the naked eye. Given the difficulty of diagnosing DED without conducting multiple time-consuming tests, we believe that our autonomous algorithm that is coupled with AS-OCT images could be helpful.”
Mr Chase said that the group was interested in using AS-OCT images to develop the AI system for diagnosing DED, considering previous research that found OCT-identified features in the corneal epithelial layer were prevalent in eyes with DED. In addition, OCT has been reported to detect DED through quantification of ocular surface adhesiveness and tear meniscus measures.
“In this black box process, we are not teaching the system what features to look at but rather asking it to discover features on its own. As we had hoped, we found that the AI was looking at the epithelial layer and the ocular surface,” he told the virtual session.
The AI algorithm for DED was trained and tested using a total of 27,180 images from 151 eyes of 91 patients. Its performance was then tested in a prospective study that included 32 eyes with DED and 28 healthy eyes.
All of the patients in the prospective study underwent testing for DED with a tear breakup time, Schirmer’s test, corneal fluorescein staining, conjunctival lissamine green staining and the Ocular Surface Disease Index. Eyes were categorised as healthy if none of the DED tests was abnormal and they had no corneal ICD-10 diagnoses. Eyes were assigned a clinical diagnosis of DED if they had an ICD-10 diagnosis of DED and at least two abnormal DED tests.
The deep learning model was very successful for diagnosing DED and differentiating DED eyes from the healthy controls. It correctly diagnosed 29 of the 32 eyes with DED and 25 of the 28 healthy eyes. In a receiver operating characteristic (ROC) curve analysis for DED, the test had an area under the curve of 0.993, 96.9% sensitivity and 95.8% specificity, Mr Chase reported.
A prediction score was also calculated for each group of eyes based on the percentage of images diagnosed as DED. The mean DED prediction score was 0.83 for the DED group and 0.11 for the healthy eyes (p<.01). Diagnosing FECD
In a recently published paper, Dr Eleiwa and colleagues reported that the ultra-high resolution OCT platform detected in vivo characteristics of FECD and could be used for both diagnosis and severity grading.
“We found that in healthy eyes, the endothelium/Descemet’s membrane is visualised as a band formed by two smooth regular hyperreflective lines with a hyporeflective space in between. In eyes with FECD, however, the posterior line showed a wavy irregular appearance with areas of focal thickening, which is considered tomographic visualisation of guttae,” he explained.
The researchers conducted a prospective case-control study that used 7,380 AS-OCT images from 15 eyes with FECD and 28 controls to evaluate the performance of the AI algorithm for diagnosing FECD. The ROC curve analysis showed that the AI algorithm was able to autonomously diagnose FECD with an AUC of 0.999, 100% sensitivity and 97.8% specificity. It had a mean FECD prediction score of 0.92 for the FECD eyes and just 0.01 in the healthy controls (p<.01). Dr Eleiwa said that prospective studies are now required to evaluate the utility of the AI algorithm to predict FECD, especially after cataract surgery. He noted that current tools for diagnosing FECD have limitations. In particular, they fall short of detecting the natural course of FECD and for predicting its progression, especially after cataract surgery. “Slit-lamp examination can miss subtle oedema, specular microscopy can have sampling errors that render its measurements inaccurate and pachymetry gives an isolated measurement of central corneal thickness that is not always representative of cornea oedema or FECD severity,” Dr Eleiwa noted.

Latest Articles

escrs members advert