SNEC, SERI and NUS School of Computing Develop Artificial Intelligence To Screen For Three Major Eye Conditions

SINGAPORE NATIONAL EYE CENTRE (SNEC),
SINGAPORE EYE RESEARCH INSTITUTE (SERI) AND NATIONAL UNVERSITY OF SINGAPORE
(NUS) SCHOOL OF COMPUTING DEVELOP ARTIFICIAL INTELLIGENCE TO SCREEN FOR THREE
MAJOR EYE CONDITIONS – DIABETIC RETINOPATHY, GLAUCOMA SUSPECT AND AGE-RELATED
MACULAR DEGENERATION

Paper newly published in the Journal of
American Medical Association (JAMA) 12 Dec 2017.

Newly developed AI system can screen for 3 eye
conditions (first in the world):  Diabetic Retinopathy, Glaucoma Suspect
(GS) and Age-related Macular Degeneration (AMD)

Introduction

Singapore National Eye Centre (SNEC) and
Singapore Eye Research Institute (SERI) have partnered National University of
Singapore (NUS) School of Computing to build an Artificial Intelligence (AI)
system to screen for diabetic eye diseases, in collaboration with several
leading eye centres globally (Australia, China, USA, Mexico and Hong Kong).
This AI technology uses a Deep Learning System (DLS), the most novel machine
learning technology, that thinks and makes decision like human intelligence in
differentiating those with and without these conditions.

This study involved 30 co-investigators who,
together, reported high diagnostic performance of an AI-based DLS in a
screening program for diabetic patients in detecting Diabetic Retinopathy (DR),
Glaucoma Suspect (GS) and Age-related Macular Degeneration (AMD) in
multi-ethnic population with diabetes. This paper was newly published in the
Journal of American Medical Association (JAMA) 12 Dec 2017 (attached).

The underlying technology is a DLS that has
the ability to learn to identify and detect retinal images that show signs of
DR and related eye diseases across multi-ethnic populations. 

Diabetic Retinopathy (DR) – Leading Cause of
Blindness in Working Adults in Singapore

DR is the leading cause of preventable
blindness among working adults in Singapore. One in three persons with diabetes
have DR. A study showed that 5 of 6 people who have DR are unaware that they
have the condition (Huang OS, Tay WT, Ong PG, et al. Br J Ophthalmol 2015; 99:
1614–1621).

With Singapore having an estimated 600,000
diabetics aged 18 to 69 (www.moh.gov.sg, 2011), about 180,000 have
DR. The Ministry of Health in Singapore has declared ‘war on diabetes’ to
rally the nation in an effort to reduce the burden of diabetes in our
population and keep Singaporeans healthy as we age.

Currently, the most effective way to prevent
DR-related vision loss is annual screening for DR, a universally accepted
practice and recommended by American Diabetes Association and the International
Council of Ophthalmology (ICO) to prevent vision loss. To address this problem,
the Singapore Integrated Diabetic Retinopathy Program (SiDRP), was set up, and
in 2017 screens 100,000 persons with diabetes across 18 primary care clinics in
Singapore. However, SiDRP relies mostly on “human grading” of the retinal
photographs by a large team of trained professional graders or optometrists.
Given rising prevalence of diabetes, SiDRP and other DR screening programmes
are challenged by availability, training and retention of professional graders
and optometrists, long-term financial sustainability and access. DR screening
remains patchy globally as a result of these challenges.

Using Artificial Intelligence (AI) and Deep
Learning System (DLS) to Screen for DR

To address the challenges, the clinical team
from SNEC and SERI (Professor Wong Tien Yin, Assistant Professor Daniel Ting)
partnered with the technical team from NUS School of Computing (Professor Lee
Mong Li Janice, Professor Wynne Hsu, Dr Gilbert Lim) to jointly develop an
AI-based system that can screen retinal images. They are the co-inventors of
this AI-based DLS, a new machine learning technology that uses
representation-learning methods to process large data and recognise intricate
structures and meaningful patterns that may not be visible to the human eye.

In this study, the researchers developed and
trained the DLS to recognise and classify retinal images to detect DR, glaucoma
suspects and age-related macular degeneration and compared it with the
performance of human evaluators of the images in the SiDRP and other studies.
This DLS was developed and tested using about 500,000 retinal images from
multi-ethnic populations across different countries, including the SiDRP
patients. The system had high rates of correctly identifying retinal images
with and without diabetic retinopathy and related eye diseases.

“The DLS will be useful for aiding DR
screening programmes in Singapore and elsewhere. In countries where there are
existing programmes such as UK and Singapore, it will increase the efficiency
and reduce cost of screening DR by replacing a large proportion of what is now
requiring “human assessment”, said Professor Wong Tien Yin, senior author and
Medical Director, SNEC, and Chairman, SERI.

“It will be easier to set up DR screening
programmes in communities in the future which could largely be done
automatically by DLS. It will also save cost and improve efficiency of
healthcare system by allowing ophthalmologists and optometrists to concentrate
on treating only DR cases that require treatment,” added Prof Wong who is also
Chair of Ophthalmology & Vice-Dean, Duke-NUS Medical School, National
University of Singapore

Collaborative Effort by SNEC-SERI with NUS
School of Computing

“This system is a collaborative effort between
the SNEC-SERI Clinical Team and the NUS School of Computing, which started many
years ago. The paper reports on the use of AI and DLS for detecting three
different retinal conditions – referable DR, GS and AMD.  The system
has sensitivity greater than 90 per cent and specificity greater than 85 per
cent to detect these conditions,” said Professor Lee Mong Li Janice, from
the Department of Computer Science at the NUS School of Computing, who is also
one of the study’s senior authors.

Fighting the ‘War on Diabetes’ and ‘Smart Nation’ Initiatives

“To date, this is the world’s first and largest dataset (with
close to half a million of retinal images) evaluating the use of a DLS not only
to screen for DR, but also other potentially vision-threatening eye conditions
such as GS and AMD. In alignment with ‘War on Diabetes’ and the Smart Nation
initiatives, we also hope to share our experience with other groups involved in
AI-related research projects,” said Assistant Professor Dr Daniel Ting from the
Singapore National Eye Center, SingHealth Duke-NUS, the lead author for the
paper and the clinical lead of the team.

Dr Ting is also awarded the highly prestigious 2017 US-ASEAN
Visiting Fulbright Scholar Program by the US Government, representing Singapore
to visit Johns Hopkins University to share and exchange his domain expertise in
the field of AI and Medicine. “AI is deemed to be the 4th industrial
revolution in the human history. In healthcare, we need to embrace this
technology earlier to improve work efficiency, while maintaining the high
standard of clinical care. Although this is an exciting result, there is still
a lot more areas in AI that we need to research further,” he added.

Next Steps

The team is now beta testing the AI system in
the Singapore Diabetic Retinopathy Screening Programme (SiDRP) alongside human
graders. They are also increasing datasets from around the world, aiming to
achieve five million images over the next five years.

“We are also developing more complex
algorithms for different DR severity levels, predictive algorithms for DR
incidence and progression, diabetes-related systemic complications for e.g.
stroke, coronary diseases and chronic kidney diseases,” said Professor Wynne
Hsu, from the Department of Computer Science at the NUS School of Computing,
who is also one of the senior co-authors of the paper.

Left picture: A patient with referable Diabetic Retinopathy
(DR)

Right picture (same
image): The heat map attention
generated by the Artificial Intelligence (AI) system, showing the areas
characterized by the presence of diabetes eye-related changes. This image is
diagnosed as referable DR.

The study authors would like to thank the National Medical
Research Council (NMRC) and

The Tanoto Foundation for the grant funding support 

Annex A

Useful Questions and Answers

(by Prof Wong Tien Yin)

1) What is deep learning system, and how does it
detect diabetic retinopathy and related eye diseases using retinal images?

Deep learning system (DLS) is an artificial intelligence
(AI)-based machine learning technology that uses methods to process large
quantities of data in their raw form, recognising intricate structures and patterns that may not be
visible to the human eyes. DLS uses convoluted neural networks (CNN) as their
“brain” to learn and train. Using this DLS, target-specific features are automatically
learnt by CNNs in the feature extraction stage and then fed into a classifier
for classification. The DLS approach does not involve any objective judgement
and the feature extraction process is entirely automatic, so that features that
are neither noticed by humans previously nor examined before will also be
assessed.

In this study, we developed and trained a DLS to classify retinal
image into those with and without referable diabetic retinopathy, referable
glaucoma suspects and age-related macular degeneration (AMD) and once this DLS
was trained, we tested and validated the DLS to detect these eye conditions
using separate datasets, including 10 external datasets from multi-ethnic
populations across different countries.

2) How might this research influence patient
care? What has to happen before it may reach the clinic?

The DLS will be useful for aiding DR screening
programmes. In countries where there are existing programmes (e.g., UK,
Singapore), it will increase the efficiency and reduce cost of screening
diabetic retinopathy by replacing a large proportion of what is now requiring
“human assessment”.

In communities and countries without existing
programmes and without sufficient ophthalmologists (e.g., developing countries,
parts of China, India, South America), it can be used as a first line screening
tool to accurately screen for cases of that require referral to an
ophthalmologist for treatment.

We are now beta testing this in the Singapore
national screening programme alongside human assessors so once the comparison
is adequate, it will be implemented

3) What is the background for this study? What
are the main findings?

Currently, annual screening for  DR is a
universally accepted practice and recommended by American Diabetes Association
and the International Council of Ophthalmology (ICO) to prevent vision loss.
However, implementation of DR screening programs across the world require human
assessors (ophthalmologists, optometrists or professional technicians trained
to read retinal photographs). Such screening programmes are thus challenged by
issues related to a need for significant human resources and long-term
financial sustainability.

To address these challenges, we developed an
AI-based software using a deep learning, a new machine learning technology.
This DLS  utilises representation-learning methods to process large data
and extract meaningful patterns. In our study, we developed and validated this
DLS using about 500,000 retinal images in a “real world screening programme”
and 10 external datasets from global populations.

The results suggest excellent accuracy of the
DLS with sensitivity of 90.5 per cent and specificity of 91.6 per cent, for
detecting referable levels of DR and 100 per cent sensitivity and 91.1 per cent
specificity for vision-threatening levels of DR (which require urgent referral
and should not be missed). In addition, the performance of the DLS was also
high for detecting referable glaucoma suspects and referable age-related
macular degeneration (which also require referral if detected).

The DLS was tested in 10 external datasets
comprising different ethnic groups: Caucasian whites, African-Americans,
Hispanics, Chinese, Indians and Malaysians.

4) What are the takeaways from your research?

First, it will be easier to set up DR
screening programmes in communities in the future which could largely be done automatically
by DLS.

Second, it will save cost and improve
efficiency of healthcare systems by allowing ophthalmologists and optometrists
to concentrate on treating only DR cases that require treatment.

5) What are the recommendations for future
research as a result of this study?

We are now beta testing this in the Singapore
national screening programme alongside human assessors so once the comparison
is adequate, it will be implemented.