August 2020


Artificial Intelligence
How AI applies to cornea

by Liz Hillman Editorial Co-Director

“We all know machines have extraordinary memory and processing power, but they don’t have intelligence. They have the ability to execute instructions programmed by humans.”
—David Wallace, MD

Refractive surgery screening, ectasia risk detection and diagnosis, recommendations for corneal crosslinking, improving refractive surgery outcomes—these are just a few of the areas where clinician-scientists are hoping to employ artificial intelligence to enhance doctors’ decision making. And, in some respects, artificial intelligence is ready to deploy in the realm of the cornea.
As Lopes et al. put it in a 2019 paper in the open-access journal Current Ophthalmology Reports, the cornea subspecialty was a “pioneer in aggregating technology to clinical practice,” and the “tremendous amount of information from complementary multimodal imaging devices” is “perfectly suitable for AI.”1

Recognizing ectasia

Renato Ambrósio Jr., MD, PhD, and colleagues have been conducting research that focuses on developing AI indices for describing the susceptibility of the cornea to ectasia and determining the impact of laser vision correction.
“Considering the vulnerability or susceptibility of the cornea for biomechanical decompensation and ectasia progression, multimodal imaging is a factual revolution in evolution,” Dr. Ambrósio said.
Dr. Ambrósio started working with artificial intelligence in 2008, helping design the Belin/Ambrósio Enhanced Ectasia Display for Pentacam (Oculus). Other indices Dr. Ambrósio mentioned on this front include the Pentacam Random Forest Index, and the tomography and biomechanical index (TBI) that combines the data from Pentacam and Corvis ST (Oculus). According to a review article that included a meta-analysis of the published studies involving the TBI, it “had the highest accuracy for the detection of subclinical keratoconus compared to all other parameters tested.”2 However, there are published reports of a lower sensitivity of TBI for detecting abnormalities on very asymmetric ectasia-normal topography (VAE-NT) cases. These cases demonstrate the opportunity and need for optimizing artificial intelligence for the combination of corneal tomography and biomechanical data. At the 2020 ASCRS Virtual Annual Meeting, Dr. Ambrósio presented a paper that showed a significant improvement in accuracy with an optimized machine learning algorithm, including the sensitivity of 85.6% (compared to TBI at 75.7%) in a series with more than 500 VAE-NT cases.
Furthermore, assessing the impact of surgery on the corneal structure has also been developed with AI, Dr. Ambrósio said, mentioning the Ectasia Susceptibility Score,3 which is available on the website of the Brazilian Study Group of Artificial Intelligence and Corneal Analysis (BrAIN). The next step is the development of the Enhanced Ectasia Susceptibility Score, Dr. Ambrósio continued, as one of the main projects of the BrAIN. He also described a new machine learning tool (Relational Tissue Altered) that represents the impact of LASIK when assessing ectasia risk, which was described in another paper from the ASCRS Virtual Annual Meeting.
David Wallace, MD, thinks that AI could help identify the early risk factors for keratoconus. Dr. Wallace is helping develop a smartphone-based, Placido corneal topography system (Delphi, Intelligent Diagnostics) that is more affordable, portable, and accessible than existing topography systems. Delphi will aggregate data in cloud servers to enable assisted analytics, intending to incorporate AI shortly after product launch.
“This may identify a subset of risk factors, including forceful eye rubbing, that lead to early topographic change, which can then, if the behavior is not altered, develop into more advanced ectasia,” he said.
Put another way, Dr. Wallace said, “the hope is that a machine like Delphi could offer the best real-time clinical interpretation to guide diagnostics and possibly therapeutic decisions.”
In general, Dr. Wallace said he thinks AI will exceed human ability looking at single studies one at a time.
“In terms of understanding the contributing factors to ectasia, AI may help us point to or identify risk factors that now perhaps are underappreciated. We also intend our system to be capable of real-time difference mapping in the cloud, which should greatly augment early detection.”
From an imaging standpoint, Dr. Wallace pointed out that Placido reflectance is more sensitive than Scheimpflug elevation mapping for the front surface of the eye and should be able to identify smaller, more subtle irregularities than possible with Scheimpflug-based systems.
“Dr. Ambrosio makes some excellent and valid points, which I wholeheartedly agree with, on relevance of posterior surgical imaging,” Dr. Wallace said. “There has been spirited debate about the relative merits of Placido vs. Scheimpflug for anterior surface imaging and the verdict is already in.”
Corneal topography is an area where neural networks have already been used for assisted diagnostics. Neural networks, Dr. Wallace explained, are fixed-size datasets that help derive comparatives for normal and abnormal diagnoses. The challenge with an open system, as with AI, is that you can’t expect that it’s going to work in the same way as a closed neural net, Dr. Wallace said.
“Any dataset needs to be somewhat curated so that bad data cannot pollute a good pool. That means finding and eliminating testing and sampling artifacts along with other sources of bad data,” he said.
“There are a lot of factors that go into doing this right that are an important part of the process. There is no guarantee that every system that uses big data is going to generate inspired and thoughtful clinical-assisted diagnostics or analytics. It’s the combination of the big data capability, the thoughtful approach, and the selective use of good data that allows advances in this space,” Dr. Wallace said.

Considering crosslinking

There are two main possibilities for AI in planning crosslinking, Dr. Ambrósio said. The finite element modeling from William Dupps, MD, PhD, has been used to create customized crosslinking algorithms, he said, and the use of AI for prognostic factors is also promising.
“In my routine, I consider the current parameters as the stiffness parameter described by Cynthia Roberts, PhD. But longitudinal studies are underway to develop such AI prognostic factors,” Dr. Ambrósio said.
As new parameters are added to datasets, such as biomechanics, Dr. Ambrósio said he thinks the ability to detect and characterize disease will be improved.
Dr. Wallace weighed in on the utility of artificial intelligence in crosslinking planning, first noting its potential to identify topographic asymmetry and early keratoconus for referrals for crosslinking. Second, will topography be able to guide custom treatment including combined excimer and CXL? he asked.
“In theory certainly, but the devil is in the details. The details require much more information than just corneal topography from anterior surface Placido reflectance as an input dataset. You would want to know thickness mapping, you might want to know something about corneal biomechanics; that is not assumed just from thickness mapping,” Dr. Wallace said.

Predicting refractive surgery outcomes

Dr. Ambrósio said beyond screening for ectasia risk, AI is promising for augmenting efficiency and predictability of refractive surgery.
Dr. Wallace was guarded in his opinion for AI’s utility in refractive surgery.
“First, laser treatment technology with either excimer (PRK, LASIK) or femtosecond laser (SMILE) has certain limitations. Second, identification of certain corneal anatomy by machine, such as the visual axis, angle kappa, and other features, is to some extent variable from machine to machine, so there may not be enough consistency to agree on the input data to guide output for treatment planning. Third, I honestly think that we are going to have several years of adaptation to AI in a strictly diagnostic mode, and that is going to need to precede any thinking about artificial intelligence in a treatment capacity,” he said.
Ultimately, Dr. Wallace cautioned against machines being “as good or better than a skilled, experienced refractive surgeon.”
“… most of us who are experienced laser refractive surgeons have spent years or decades getting here, and I don’t think it’s going to be possible to distill all of our collective information into a few thousand lines of code and make a machine that can do it better,” he said.

Closing thoughts

Dr. Wallace made some points about the use of the term “intelligence.” He said intelligence is “information that is thoughtfully used or innovatively assessed to create new ways of thinking or new ways of analyzing that weren’t possible or weren’t done before.” As such, he pointed out, a machine isn’t, in and of itself, “intelligent.”
“We all know machines have extraordinary memory and processing power, but they don’t have intelligence. They have the ability to execute instructions programmed by humans,” he said, noting that’s why he prefers the terms “augmented intelligence” or “machine-enhanced analysis.”
Dr. Wallace said that AI may not make doctors better at what they’ve spent their careers doing.
“AI is not necessarily the answer. It’s the thoughtful use of information by thoughtful people gathering good data, the right way, and putting it together that helps the massive processing power of current computer technology to draw smart conclusions from much bigger datasets than humans could ever grasp and process,” Dr. Wallace said.

At a glance

• Artificial intelligence (AI) helps to predict ectasia risk factors prior to refractive laser vision correction procedures (PRK, LASIK, SMILE).
• Datasets for AI need to be carefully curated in order to deliver an accurate algorithm.
• AI is likely to be used, in addition to the diagnosis, for the prognosis and treatment, including crosslinking.

A few AI papers at the 2020 ASCRS Virtual Annual Meeting

•“Diagnostic Performance of an Artificial Intelligence Algorithm in Fuchs Endothelial Cell Dystrophy” found AI to be a viable option for Fuchs detection with “excellent accuracy, sensitivity, and specificity,” though prospective studies are still required to evaluate AI’s utility in assessing Fuchs progression.
•“A Prospective Study for Autonomous Diagnosis of Dry Eye Syndrome Using an Artificial Intelligence Algorithm” found that the algorithm used with anterior segment OCT could be helpful in diagnosing dry eye disease.
•“An Artificial Intelligence (AI) Algorithm for the Autonomous Diagnosis of Corneal Graft Rejection” determined that AI can accurately diagnose graft rejections.

About the doctors

Renato Ambrósio Jr., MD, PhD
Director of cornea and
refractive surgery
Instituto de Olhos
Renato Ambrósio
Rio de Janeiro, Brazil

David Wallace, MD
Medical director and CEO
LA Sight
Founder and managing partner
Intelligent Diagnostics
Los Angeles, California


1. Lopes BT, et al. Artificial intelligence in corneal diagnosis: Where are we? Curr Ophthalmol Rep. 2019;7:204–211.
2. Esporcatte LPG, et al. Biomechanical diagnostics of the cornea. Eye Vis (Lond). 2020;7:9.
3. Ambrosio JR, et al. Assessing ectasia susceptibility prior to LASIK: the role of age and residual stromal bed (RSB) in conjunction to Belin-Ambrósio deviation index (BAD-D). Rev Bras Oftalmol. 2014;73:75–80.

Relevant disclosures

: Oculus, Alcon, Carl Zeiss Meditec, Allergan, Mediphacos
Wallace: Intelligent Diagnostics


Ambrósio: dr.renatoAmbró

How AI applies to cornea How AI applies to cornea
Ophthalmology News - EyeWorld Magazine
283 110
220 138
True, 8