Meeting News

Artificial intelligence, machine learning making strides in ophthalmology

Andrew Schachat

WAIKOLOA, Hawaii — Ophthalmologists are using artificial intelligence and machine learning to develop programs to automate diagnoses, improve monitoring and help detect a variety of ophthalmic diseases in patients.

“We are being swamped with papers on this. You’re going to be swamped trying to understand this and reading about it and learning from it. There have been some spectacular things that have come out in the last year or two,” Andrew P. Schachat, MD, editor in chief of Ophthalmology Retina, said in a presentation at Hawaiian Eye 2019.

Several studies published this past year show the impact of deep learning, computers and artificial intelligence on improving patient care in the specialty, he said.

Researchers published a study in November 2018 that demonstrated how a deep learning model could accurately predict age-related macular degeneration progression. The model, DeepSeeNet, trained itself by evaluating 58,402 color fundus photographs and tested itself on 900 images from a longitudinal follow-up of 4,549 participants from AREDS. The program accurately detected large drusen and pigment changes, both key risk factors for AMD progression, better than retina specialists.

Deep learning could fully automate the detection of macular fluid on OCT images, according to an April 2018 study. The computer program could automatically detect and quantify either intraretinal or subretinal fluid. The presence of fluid drives the decision whether to give an anti-VEGF injection, he said.

“The computer could, as well as a retina specialist, say if there’s fluid or not,” Schachat said. “You could take a picture and let the computer tell the patient whether they should go to the office and come in for an injection,” Schachat said.

Machine learning accurately predicted visual acuity outcomes in a July 2018 study of neovascular AMD. Researchers used five different machine learning algorithms to predict visual acuity in patients with neovascular AMD at 3 months and 12 months after initial treatment with three anti-VEGF injections.

The program accurately predicted visual acuity within half a line at 3 months and was “pretty good” at 12 months, Schachat said.

These results matter because it may be possible to encourage adherence in patients who would do well with therapy or to help with patient selection for those who would not do well, he said.

In an August 2018 study, researchers developed a deep learning system to automatically classify glaucomatous optic neuropathy by analyzing color fundus photographs. The system achieved a 95.6% sensitivity and 92% specificity at reliably detecting glaucomatous optic atrophy.

“Where did it tend to make mistakes? When pathologic myopia was also present. That confuses the computer and it confuses us in clinic, so that makes sense,” Schachat said.

Outside of studies published in Ophthalmology, a study published in Investigative Ophthalmology and Visual Science, showed deep learning algorithms could predict refractive error with a mean absolute error of 0.5 D from fundus photographs, Schachat said.

Researchers published a study in Nature Biomedical Engineering in 2018 in which a deep learning system accurately predicted cardiovascular risk by evaluating fundus photographs. The computer trained itself on 284,335 images and validated itself on two independent data sets of 12,026 and 999 patients. The computer algorithm could accurately predict patient age by 3.2 years and gender with a 0.97 area under the curve.

“It can announce what your hemoglobin A1c is within 1.4%. It can announce what your systolic blood pressure is within 11 mm Hg and predict major cardiac adverse events pretty well with an area under the curve of 0.7,” Schachat said. – by Robert Linnehan

References:

Li Z, et al. Ophthalmology. 2018;doi:10.1016/j.ophtha.2018.01.023.

Peng Y, et al. Ophthalmology. 2018;doi:10.1016/j.ophtha.2018.11.015.

Poplin R, et al. Nature Biomedical Engineering. 2018;2:158-164.

Rohm M, et al. Ophthalmology. 2018;doi:10.1016/j.ophtha.2017.12.034.

Schachat AP. The year’s most intriguing papers from the journal, Ophthalmology. Presented at: Presented at: Hawaiian Eye; Jan. 19-25, 2019; Waikoloa, Hawaii.

Schlegl T, et al. Ophthalmology. 2018;doi:10.1016/j.ophtha.2017.10.031.

Varadarajan AV, et al. Invest Ophthalmol Vis Sci. 2018;doi:10.1167/iovs.18-23887.

Disclosure: Schachat reports he receives royalties from Elsevier and a stipend from the American Academy of Ophthalmology; his employer is the Cleveland Clinic and State of Ohio; and he may receive possible future payments from Easton Capital.

Andrew Schachat

WAIKOLOA, Hawaii — Ophthalmologists are using artificial intelligence and machine learning to develop programs to automate diagnoses, improve monitoring and help detect a variety of ophthalmic diseases in patients.

“We are being swamped with papers on this. You’re going to be swamped trying to understand this and reading about it and learning from it. There have been some spectacular things that have come out in the last year or two,” Andrew P. Schachat, MD, editor in chief of Ophthalmology Retina, said in a presentation at Hawaiian Eye 2019.

Several studies published this past year show the impact of deep learning, computers and artificial intelligence on improving patient care in the specialty, he said.

Researchers published a study in November 2018 that demonstrated how a deep learning model could accurately predict age-related macular degeneration progression. The model, DeepSeeNet, trained itself by evaluating 58,402 color fundus photographs and tested itself on 900 images from a longitudinal follow-up of 4,549 participants from AREDS. The program accurately detected large drusen and pigment changes, both key risk factors for AMD progression, better than retina specialists.

Deep learning could fully automate the detection of macular fluid on OCT images, according to an April 2018 study. The computer program could automatically detect and quantify either intraretinal or subretinal fluid. The presence of fluid drives the decision whether to give an anti-VEGF injection, he said.

“The computer could, as well as a retina specialist, say if there’s fluid or not,” Schachat said. “You could take a picture and let the computer tell the patient whether they should go to the office and come in for an injection,” Schachat said.

Machine learning accurately predicted visual acuity outcomes in a July 2018 study of neovascular AMD. Researchers used five different machine learning algorithms to predict visual acuity in patients with neovascular AMD at 3 months and 12 months after initial treatment with three anti-VEGF injections.

The program accurately predicted visual acuity within half a line at 3 months and was “pretty good” at 12 months, Schachat said.

These results matter because it may be possible to encourage adherence in patients who would do well with therapy or to help with patient selection for those who would not do well, he said.

In an August 2018 study, researchers developed a deep learning system to automatically classify glaucomatous optic neuropathy by analyzing color fundus photographs. The system achieved a 95.6% sensitivity and 92% specificity at reliably detecting glaucomatous optic atrophy.

“Where did it tend to make mistakes? When pathologic myopia was also present. That confuses the computer and it confuses us in clinic, so that makes sense,” Schachat said.

Outside of studies published in Ophthalmology, a study published in Investigative Ophthalmology and Visual Science, showed deep learning algorithms could predict refractive error with a mean absolute error of 0.5 D from fundus photographs, Schachat said.

Researchers published a study in Nature Biomedical Engineering in 2018 in which a deep learning system accurately predicted cardiovascular risk by evaluating fundus photographs. The computer trained itself on 284,335 images and validated itself on two independent data sets of 12,026 and 999 patients. The computer algorithm could accurately predict patient age by 3.2 years and gender with a 0.97 area under the curve.

“It can announce what your hemoglobin A1c is within 1.4%. It can announce what your systolic blood pressure is within 11 mm Hg and predict major cardiac adverse events pretty well with an area under the curve of 0.7,” Schachat said. – by Robert Linnehan

References:

Li Z, et al. Ophthalmology. 2018;doi:10.1016/j.ophtha.2018.01.023.

Peng Y, et al. Ophthalmology. 2018;doi:10.1016/j.ophtha.2018.11.015.

Poplin R, et al. Nature Biomedical Engineering. 2018;2:158-164.

Rohm M, et al. Ophthalmology. 2018;doi:10.1016/j.ophtha.2017.12.034.

Schachat AP. The year’s most intriguing papers from the journal, Ophthalmology. Presented at: Presented at: Hawaiian Eye; Jan. 19-25, 2019; Waikoloa, Hawaii.

Schlegl T, et al. Ophthalmology. 2018;doi:10.1016/j.ophtha.2017.10.031.

Varadarajan AV, et al. Invest Ophthalmol Vis Sci. 2018;doi:10.1167/iovs.18-23887.

Disclosure: Schachat reports he receives royalties from Elsevier and a stipend from the American Academy of Ophthalmology; his employer is the Cleveland Clinic and State of Ohio; and he may receive possible future payments from Easton Capital.

    See more from Hawaiian Eye/Retina Meeting