m
Recent Posts
Connect with:
Tuesday / March 5.
HomemipatientArtificial Intelligence in Eye Care: The Future is Now

Artificial Intelligence in Eye Care: The Future is Now

In this month’s column, Dr Margaret Lam picks the brain of Professor Ming He, one of the world’s pre-eminent thought leaders in artificial intelligence (AI) technology in the field of ophthalmology, Professor of Ophthalmic Epidemiology at the University of Melbourne, and the developer of innovative AI systems for retinal screening.

Q1. How can AI in optometry and ophthalmology currently assist in diagnosing eye conditions, and conversely, what are its challenges and limitations?

AI has evolved from computer software in the laboratory to clinical adaptation in many image-driven practices, including ophthalmology. Enabled by a combination of the availability of large datasets and substantially improved computing power, deep learning algorithms (DLAs) have created unprecedented opportunities for substantially improved accuracy in automated diagnosis and classification of medical images. This automated diagnosis, which makes eye care less dependent on human input, is improving accessibility, efficiency, and cost-effectiveness, making ocular disease diagnosis and management quicker, cheaper, and more consistent.

My team was among the first groups to develop and publish DLAs on diabetic retinopathy (DR), glaucoma, and age-related macular degeneration (AMD) based on full colour fundus photographs.1-3 Up until now, a number of research groups have published their results on a broader range of eye diseases, based on more imaging modalities, for example, optical coherence tomography (OCT), visual field and many others.4 Despite this wide range of development, DLA application fundamentally focusses on three major tasks: classification, segmentation, and prediction based on static images collected.

Most DLA studies have reported robust accuracy (normally described as AUC), sensitivity, and specificity. For example, in the classification of referrable DR, DLA achieves an accuracy of AUC > 0.90 with a sensitivity and a specificity of > 0.90. This sensitivity and specificity is comparable, if not better, than trained human graders.2,5

There are many challenges and limitations to adopting DLAs into clinical practice.6,7 Successful AI deployment in clinical practice requires active engagement of all stakeholders; patients, ophthalmologists, imaging technicians, hospital administration, regulatory bodies, and industry. It is difficult to ensure all these stakeholders benefit from AI deployment and collaboratively support and facilitate the development of best practice. Many challenges, such as ethics, patient consent, privacy protection, data ownership and sharing, medical liability, regulatory approval, integration into existing electronic medical records, and user-friendly software interfaces also need to be resolved.

This automated diagnosis… is improving accessibility, efficiency, and cost-effectiveness

Q2. Where do you see AI assisting in optometry and ophthalmology, now and in the next five years?

AI creates unprecedented opportunities for both optometry and ophthalmology. When installed in non-eye clinical settings, it can be used as a tool to create targeted referrals to eye care professionals, basically bringing more patients to us.8 For instance, installing an AI system in endocrinology and primary care settings will help create more effective referrals for DR, glaucoma, and cataract at a time when most of these conditions are clinically ‘treatable’. This alone, will potentially prevent vision loss.

AI can be used as an assistive tool in parallel with, or after clinical diagnosis, to ensure optometrists and ophthalmologists do not miss cases of ocular diseases. It can also be developed as an ‘audit’ tool, enabling images collected over time to be verified for accuracy and consistency of an optometrist’s glaucoma and DR diagnosis.9

In the next five years, more and more DLAs will be approved by regulatory bodies to detect common eye diseases, such as cataract, DR, glaucoma, and AMD, based on an increasing number of imaging modalities including fundus photography, OCT, visual fields, and more.

Automated lesion detection and structural segmentation tools will be built-in for most ophthalmic imaging devices so that as soon as a fundus photo is taken, lesions such as exudate, haemorrhage, IRMAs (intraretinal microvascular abnormalities), or neovascularisation, will be ‘annotated’.

More than simply discriminating a limited number of predefined eye diseases, AI will accurately perform clinical reasoning based on clinical symptoms and signs from multiple imaging modalities, to make valid diagnoses of all eye diseases, even those it has not been trained for.

Q3. Some practitioners feel AI in optometry is telling us what an eye care professional already knows. What is your opinion?

DLA has already been commonly adopted for disease classification, eg., classifying images as with or without glaucomatous neuropathy, or detecting and delineating the boundaries of specific lesions, such as an optic disc and optic cup, to enable further calculation of the vertical cup disc ratio.7

Detecting DR or AMD based on full colour fundus photographs is relatively easy and objective in clinical practice. However, detection of glaucoma is far more challenging and subjective, particularly in the early phase of disease. This is when it is very difficult to differentiate physiological cupping and perimetric glaucoma by looking at the optic disc appearance alone, which can result in substantial variation among clinicians. AI can help standardise this process within a practice and hopefully, reduce the chance of a missed diagnosis.

In addition to simple classification of an image, when progression data are used as ground truth to train an algorithm, AI can predict the prognosis or outcomes of a treatment. DLA is able to achieve reasonable accuracy in predicting the prognosis of DR, glaucoma and AMD progression.

Furthermore, DLA is able to see features in colour fundus photographs, for example, that are not differentiable by humans. Already, this is enabling AI to estimate the risk of developing cardiovascular events.

Q4. What are you working on now to increase the scope and efficiencies that your AI system can offer eye care?

Supported by the Medical Research Future Fund (MRFF), a 10-year investment plan from the federal government, our research group at Centre for Eye Research Australia, in collaboration with Eyetelligence Ltd (industrial partner), and Monash e-Research Centre (technologic partner) and Inner Maven Pty Ltd (consumer and commercialisation partner), has launched a three-year project to prove how AI technology can:

1) Be adopted to create a new model of care to enable and improve efficiency of early diagnosis of eye diseases in primary care settings such as GPs and Aboriginal Medical Services,

2) Provide standardisation and an objective reassurance to eye diseases with minimal error and variation in eye care service, and

3) Incorporate an AI retinal risk score to increase the accuracy of risk prediction on cardiovascular diseases within cardiology services. This project is on-going and will aid in providing the best evidence to understand the best approach to adopt AI in clinical practice.

AI and Glaucoma Diagnosis: A Case Study

A series of typical patient images, collected by a busy optometry practice, were run through a DLA. None of the patients had been identified as glaucoma suspects or received visual field testing at their initial consultation.  The DLA identified several glaucoma suspects among the images.

Figure 1 highlights localised retinal nerve fibre layer (RNFL) defects on both eyes.


Figure 2 demonstrates a normal rim on the left eye but localised notching, consistent with RNFL defect, on the right eye.


Figure 3 demonstrates a combination of high myopia, parapapillary atrophy, and rim loss at the superior quadrant, where glaucoma diagnosis is easily missed in busy practices. AI identified the patients as either high risk (red), moderate risk (amber) or low risk (green).

 

Dr Margaret Lam is the National President of Optometry Australia and a Director of Optometry NSW/ACT. She practices optometry in Sydney and teaches at the School of Optometry at UNSW as an Adjunct Senior Lecturer.

Professor Mingguang He is the Professor of Ophthalmic Epidemiology at the University of Melbourne and a global expert in vision-related clinical and epidemiologic research. He has led important epidemiological studies and clinical trials, including the first population-based study on myopia in China; the first population-based study on glaucoma in China; a clinical trial to prove the efficacy of increased outdoor time on myopia prevention, published in JAMA 2015, and a prophylactic clinical trial on angle closure glaucoma, published in Lancet 2019. His publications have attracted more than 13,000 citations.

Prof He has received more than AU$10 million in major research grants in Australia, including the prestigious NHMRC investigator Grant and Medical Research Future Fund for his research on artificial intelligence.

He is the deputy editor-in-chief of British Journal of Ophthalmology. He founded and served as the first president of the Asia Pacific Tele-Ophthalmology Society, is a founding council member of Asia Pacific Myopia Society, and the Deputy Secretary-General for the Asia Pacific Academy of Ophthalmology.

References

  1. Li Z., He Y., Keel S., et al., Efficacy of a deep learning system for detecting glaucomatous optic neuropathy based on color fundus photographs. Ophthalmology 2018; 125:1199-1206.
  2. Li Z., Keel S., Liu C., et al., An automated grading system for detection of vision-threatening referable diabetic retinopathy on the basis of color fundus photographs. Diabetes Care 2018; 41:2509-2516.
  3. Keel S., Wu J., Lee P.Y., et al., Visualizing deep learning models for the detection of referable diabetic retinopathy and glaucoma. JAMA Ophthalmol 2019; 137:288-292.
  4. De Fauw J., Ledsam J.R., Romera-Paredes B., et al., Clinically applicable deep learning for diagnosis and referral in retinal disease. Nat Med 2018; 24:1342-1350.
  5. Ting D., Cheung C.Y., Lim G., et al., Development and validation of a deep learning system for diabetic retinopathy and related eye diseases using retinal images from multi-ethnic populations with diabetes. JAMA 2017; 318:2211-2223.
  6. Tan Z., Scheetz J., He M. Artificial intelligence in ophthalmology: accuracy, challenges, and clinical application. Asia Pac J Ophthalmol (Phila) 2019; 8:197-199.
  7. He M., Li Z., Liu C., et al., Deployment of Artificial Intelligence in Real-World Practice: Opportunity and Challenge. Asia Pac J Ophthalmol (Phila) 2020; 9:299-307.
  8. Keel S., Lee P.Y., Scheetz J., et al., Feasibility and patient acceptability of a novel artificial intelligence based screening model for diabetic retinopathy at endocrinology outpatient services: a pilot study. Sci Rep 2018; 8:4330.
  9. Li Z., Keel S., Liu C., et al., Can artificial intelligence make screening faster, more accurate, and more accessible? Asia Pac J Ophthalmol (Phila) 2018; 7:436-441.