Sivanthi Kanagasundaram

Introduction

Since the launch of ChatGPT, the term “Artificial Intelligence” (AI) is now used colloquially.  Although its launch has catalysed its eminence, AI has slowly seeped its way into society over the past few decades. One of its first avatars was in the form of a navigation system that guided individuals from point A to B, saving travellers from the panic that ensued with reading an A-to-Z map. It later appeared in the form of Siri, whose witty comebacks offered amusement. It is now knocking on the doors of Medicine. With careful consideration of ethical dilemmas, its new incarnation in Medicine, specifically in the field of Ophthalmology, could be ground-breaking.

Interpretation of optical coherence tomography (OCT) images and retinal fundoscopy is integral to the practice of Ophthalmology. Despite tireless efforts to keep up with the seemingly endless clinics and intricate surgeries, Ophthalmologists must dedicate a significant portion of their time analysing images. By training and validating AI systems, AI has the potential to aid in the analysis, diagnosis and perhaps even management of a large array of visual pathologies. With increased NHS demand straining the current pool of human resources, the introduction of AI will be necessary in aiding doctors juggle their caseloads efficiently.  

Machine Learning and Deep Learning

AI is the term used to describe the application of computational algorithms to analyse large data sets. Machine learning (“ML”), a subdivision of AI, uses algorithms to draw inferences from patterns in data. Inputs to ML algorithms sourced from patient notes and imaging can be mapped to an output; the probability of developing a disease (1). The ability to forecast susceptible patients to glaucoma from cumulative risk factors, such as elevated intraocular pressure and family history, demonstrates this clinical application.

A subset to ML is Deep Learning (“DL”), which explores more complex non-linear patterns in the data, creating a neuronal network with multiple layers (2). An example of DL is its ability to interpret OCT images and subsequently classify them into retinal pathologies. Note the disparity from ML – there is no strict mapping of input to output. The algorithm is capable of learning complex patterns from a large dataset, which may not be obvious to the human eye (3). Therefore, its application to new data can provide valuable insights and innovative approaches in managing patients.

Current application of Artificial Intelligence in Ophthalmology

AI has made a significant impact in the diagnosis of Diabetic Retinopathy (“DR”) in the US. The pathogenesis of DR unfolds in the microvasculature of the retina, masquerading as a silent threat to vision, until its advanced symptomatic stage. The capitalisation of DL, as a screening modality, permits timely detection of DR in a large population (4).

In 2018, the FDA approved its first AI-assisted DR device; ‘IDx-DR’, which correctly identified the presence of mild Diabetic Retinopathy 87.4% of the time, by analysing images of patients’ retinas (5). Its ability to provide a screening decision without any human intervention has potential to transform current practice. Imagine a world where patients are solely screened with AI algorithms and only referred to clinicians in instances where it detects anomalies that warrant further investigation. Streamlining referrals with AI allows clinicians to prioritise their caseloads responsibly, by allowing them to first focus on providing the best care to the fraction of individuals that require urgent intervention.

DeepMind and Moorfields Eye Hospital have developed an AI system that triages referrals based on DL analysis of OCT images (2). OCT is the current gold standard for imaging the macula and retina, assisting in detection and classification of over 50 different OCT pathologies. This similarly broadens the possibility of Ophthalmologists to manage time-sensitive diseases, such as retinal detachment, rather than scrutinising scans. AI systems have proven they can analyse this imagery in shorter spaces of time, with a similar performance rating to its human counterpart, essentially offering a win-win situation (2).

The use of OCT imaging has expanded beyond diagnosing pathologies of just the eye. “Oculomics” is a growing field in medicine which has allowed for ophthalmic features to act as biomarkers for systemic disease. From the “AlzEye” database, a sample of 154,830 retinal images were analysed using AI. This elicited subtle differences in the thickness of certain retinal layers in patients prevalent for Parkinson’s Disease (6). Photographing the back of the eye is non-invasive, instant and grants a cost-effective substitute for current expensive standards for diagnosis, like neuroimaging. Moreover, early diagnoses of systemic diseases with accurate treatment, would allow patients adequate time for lifestyle changes and can reduce the impact of debilitating conditions on their quality of life.

Potential Applications of Artificial Intelligence in Ophthalmology

AI coupled with robotics has the potential to revolutionise eye surgery. A steady hand in ophthalmic surgery is a prerequisite, however, retinal microsurgery requires almost super-human precision for desired outcomes (7). Early studies involving robotics have shown its potential in filtering out innate tremors and exhibiting perfect dexterity – breaking boundaries in surgery (8). Additionally, the ‘intelligence’ a robot gains during data learning, could pre-calculate the best surgical approach for a procedure. The input of data could be in the form of real-time surgical tracking during the operation and post-surgical data, such as the outcome of surgery and any complications (9). This a long way off being put into widespread practice, given the rich quantities of data needed for its data learning process, as well as extensive testing required to ensure patient safety. 

With such a vast volume of data being gathered, AI could help train the next generation of ophthalmologists by providing a simulated environment to hone patient diagnosis and treatment (10). Being able to do so in a setting where incorrect management has no harm to patients, allows trainees to essentially ‘learn from their mistakes’, without real consequences. This use of AI could offer instant feedback, highlighting improvements or alternative approaches to clinical scenarios. It is to be emphasised, that any simulated training would need to be paired with regular patient interaction to ensure the human element of care and empathy with patients is not lost.

AI coupled with Telemedicine can expedite specialist input via virtual consultations. Existing AI screening systems like ‘IDx-DR’ could signpost high risk patients for an immediate online consultation with a specialist, specifically for patients based in remote or rural locations (11). Integration of automated AI software into the NHS could allow for a quick and efficient patient referral process for sight saving treatment, significantly cutting down on the average 22-week wait to be followed up in retina clinics (12).

Addressing current challenges posed by Artificial Intelligence

Since its early emergence into society, AI has repeatedly been painted as a sleeping beast set to take over the world. Its villainization may be to blame for the slow incorporation of AI systems into aspects governing our health. It is imperative then, to tackle fears of the public and clinicians to ensure trust and competence of AI algorithms.

Healthcare is a field centred around patient safety and although AI systems are intricately designed, they are still prone to errors. If Artificial Intelligence fails to highlight a patient with a particular eye pathology, who will be liable? For this question to be answered effectively, it will be necessary to introduce a dedicated figure to overlook AI in a healthcare setting and standardise the technology (13). Constant audits with feedback from clinicians and monitoring patient outcomes would help calibrate and refine the use of AI in healthcare (14).

To uphold trust between the caregiver and patient, the former should be transparent about using AI in their management. Patients should be communicated the benefits as well as its shortcomings so that an informed decision is made about whether AI is to be integrated into their care (15).

Finally, there must be a balance between doctors using their own clinical judgement and acknowledging the strengths of AI algorithms(13). Clinicians should not overly rely on AI as it could weaken their decision-making abilities and when alternative management plans are proposed, it could strip doctors from their autonomy. A good doctor will learn to incorporate AI, however, a better doctor will also know when to override its decision.

Conclusion

The integration of Artificial Intelligence into the field of Ophthalmology holds tremendous promise for transforming the delivery of eye care. Established AI systems have already proven that they enhance care but will need to be further developed to make it safe to merge into our healthcare. Introduction of AI can reduce the burden on ophthalmologists and healthcare workers alike, and finally create a nurturing workplace that has always been sought after. The initial implementation of AI will be costly, but once AI systems are in place, they could assist early diagnosis of ophthalmic, and possibly even, systemic conditions. Offering treatment before disease progression can circumvent procedures that the patient would have to go through if they were diagnosed at a later stage. Perhaps AI is not the menacing monster it has been made out to be but an omniscient presence paving the way for the future.

References

1. Balyen, Lokman MD; Peto, Tunde MD, PhD. Promising Artificial Intelligence-Machine Learning-Deep Learning Algorithms in Ophthalmology. Asia-Pacific Journal of Ophthalmology 8(3):p 264-272, May-June 2019. | DOI: 10.22608/APO.2018479
2. Ahuja AS, Halperin LS. Understanding the advent of artificial intelligence in ophthalmology. J Curr Ophthalmol. 2019 May 28;31(2):115-117. doi: 10.1016/j.joco.2019.05.001. PMID: 31317087; PMCID: PMC6611924.
3. Choi RY, Coyner AS, Kalpathy-Cramer J, Chiang MF, Campbell JP. Introduction to Machine Learning, Neural Networks, and Deep Learning. Transl Vis Sci Technol. 2020 Feb 27;9(2):14. doi: 10.1167/tvst.9.2.14. PMID: 32704420; PMCID: PMC7347027.
4. Ting DSW, Pasquale LR, Peng L, et al Artificial intelligence and deep learning in ophthalmology British Journal of Ophthalmology 2019;103:167-175.
5. Padhy SK, Takkar B, Chawla R, Kumar A. Artificial intelligence in diabetic retinopathy: A natural step to the future. Indian J Ophthalmol. 2019 Jul;67(7):1004-1009. doi: 10.4103/ijo.IJO_1989_18. PMID: 31238395; PMCID: PMC6611318.
6. Wagner SK, Romero-Bascones D, Cortina-Borja M, Williamson DJ, Struyven RR, Zhou Y, Patel S, Weil RS, Antoniades CA, Topol EJ, Korot E, Foster PJ, Balaskas K, Ayala U, Barrenechea M, Gabilondo I, Schapira AH, Khawaja AP, Patel PJ, Rahi JS, Denniston AK, Petzold A, Keane PA; for UK Biobank Eye & Vision Consortium. Retinal Optical Coherence Tomography Features Associated With Incident and Prevalent Parkinson Disease. Neurology. 2023 Aug 21:10.1212/WNL.0000000000207727. doi: 10.1212/WNL.0000000000207727. Epub ahead of print. PMID: 37604659.
7. Urias, M.G., Patel, N., He, C. et al. Artificial intelligence, robotics and eye surgery: are we overfitted?. Int J Retin Vitr 5, 52 (2019).
8. Pandey SK, Sharma V. Robotics and ophthalmology: Are we there yet? Indian J Ophthalmol. 2019 Jul;67(7):988-994. doi: 10.4103/ijo.IJO_1131_18. PMID: 31238393; PMCID: PMC6611303.
9. Qiu L, Li C, Ren H. Real-time surgical instrument tracking in robot-assisted surgery using multi-domain convolutional neural network. Healthc Technol Lett. 2019 Dec 5;6(6):159-164. doi: 10.1049/htl.2019.0068. PMID: 32038850; PMCID: PMC6945802.
10. Bakshi SK, Lin SR, Ting DSW, Chiang MF, Chodosh J. The era of artificial intelligence and virtual reality: transforming surgical education in ophthalmology. Br J Ophthalmol. 2021 Oct;105(10):1325-1328. doi: 10.1136/bjophthalmol-2020-316845. Epub 2020 Aug 14. PMID: 32816750.
11. Chia MA, Turner AW. Benefits of Integrating Telemedicine and Artificial Intelligence Into Outreach Eye Care: Stepwise Approach and Future Directions. Front Med (Lausanne). 2022 Mar 11;9:835804. doi: 10.3389/fmed.2022.835804. PMID: 35391876; PMCID: PMC8982071.
12. Foot B, MacEwen C. Surveillance of sight loss due to delay in ophthalmic treatment or review: frequency, cause and outcome. Eye (Lond). 2017 May;31(5):771-775. doi: 10.1038/eye.2017.1. Epub 2017 Jan 27. PMID: 28128796; PMCID: PMC5437335.
13. Abdullah YI, Schuman JS, Shabsigh R, Caplan A, Al-Aswad LA. Ethics of Artificial Intelligence in Medicine and Ophthalmology. Asia Pac J Ophthalmol (Phila). 2021 May-Jun 01;10(3):289-298. doi: 10.1097/APO.0000000000000397. PMID: 34383720; PMCID: PMC9167644.
14. Kohli A, Mahajan V, Seals K, Kohli A, Jha S. Concepts in U.S. Food and Drug Administration Regulation of Artificial Intelligence for Medical Imaging. AJR Am J Roentgenol. 2019 Oct;213(4):886-888. doi: 10.2214/AJR.18.20410. Epub 2019 Jun 5. PMID: 31166758.
15. Anderson M, Anderson SL. How Should AI Be Developed, Validated, and Implemented in Patient Care? AMA J Ethics. 2019 Feb 1;21(2):E125-130. doi: 10.1001/amajethics.2019.125. PMID: 30794121.