Perspectives Ophthalmology and the emergence of artificial intelligence Rapid advances in AI in ophthalmology are a harbinger of things to come for other fields of medicine he autonomous detection and triage of eye for retinopathy using retinal photography. This disease, or even accurate estimations of gender, vast demand for diabetic eye screening services Tage, and blood pressure from a simple retinal has stimulated the development of AI algorithms to photo, may sound like the realms of science fiction, identify sight-threatening disease. Several algorithms but advances in artificial intelligence (AI) have already have achieved performance that meets or exceeds that made this a reality.1 Ophthalmology is at the vanguard of human experts.4,5 Accordingly, in 2018, the United of the development and clinical application of AI. States Food and Drug Administration approved an AI Advances in the field may provide useful insights into system to detect referable diabetic retinopathy from the application of this technology in health care more retinal photographs, the first autonomous diagnostic broadly. system to be approved in any field of medicine.6 Advances in deep learning have extended to other Artificial intelligence imaging modalities that are commonly used in ophthalmology. Ocular coherence tomography is an Once described as the capacity of intelligent machines imaging technology that produces highly detailed, to imitate human intelligence and behaviour, AI now depth-resolved images of the retina. A recent describes many theories and practices used to achieve collaboration between researchers and clinicians computer intelligence (Box 1).2 Machine learning is at Google DeepMind, Moorfields Eye Hospital an application of AI that uses algorithms or statistical and University College London culminated in the models to make decisions or predictions. Complex development of a deep learning system capable patterns and relationships are learned from data to of detecting and triaging more than 50 different generate an outcome.2 Machine learning traditionally retinal conditions at levels equivalent to a panel of relies on the extraction of features from the data by experienced ophthalmologists.7 AI systems with the human operators which then serve as input variables capacity to detect a wide range of diseases, such as to optimise algorithm performance. The performance this, are likely to be most useful in clinical practice. of these systems is constrained by the features that are recognised as important by humans. A highly anticipated innovation is the development of AI systems capable of accurate disease prediction. Such In contrast, artificial neural networks are an advanced tools could assist in managing patient expectations, method of machine learning able to extract features improve the quality of care and reduce treatment without explicit programming.2 Deep learning is the costs.3 In ophthalmology, prediction models have construction of multiple layers of artificial neural been trained to personalise re-treatment intervals for networks which can identify features in data that are not recognisable by humans. Although deep learning systems may be powerful, they lack human-crafted inputs, meaning that large 1 Relationship between artificial intelligence and its subtypes quantities of data are typically required to train algorithms. Jane Scheetz1 1,2 Mingguang He Artificial intelligence in ophthalmology Peter van Wijngaarden1,3 As a discipline, ophthalmology is at the forefront of AI system development and translation in 1 Centre for Eye Research Australia, Royal clinical practice. Leading uses of the technology MJA Victorian Eye and include detecting, classifying and triaging a Ear Hospital, University of Melbourne, range of diseases, such as diabetic retinopathy, 214 Melbourne, VIC. ( age-related macular degeneration (AMD), 4 ) 2 State Key Laboratory of ▪ glaucoma, retinopathy of prematurity, and 2021 1 March Ophthalmology, Zhongshan 3 Ophthalmic Center, retinal vein occlusion, from clinical images. The Sun Yat-sen University, Guangzhou, increasing global burden of eye diseases, coupled China. with the development of new therapies for 3 University of Melbourne, previously untreatable conditions, has served as a Melbourne, VIC. major driver for AI innovation in ophthalmology. peterv@unimelb. As a case in point, there are presently over 430 edu.au million people living with diabetes, most of 155 doi: 10.5694/mja2.50932 whom require annual or biennial screening Perspectives patients with neovascular AMD,8 predict progression is not certain how data withdrawal requests will from early to late AMD,9 estimate the extent of future be dealt with when an individual’s data have been visual field defects in patients with glaucoma,10 and used in the process of training a deep learning predict diabetic retinopathy progression.11 Although system. Accordingly, developments in AI need to these models presently achieve only moderate levels be accompanied by advanced data protection and of accuracy, their performance has been shown to security measures. be superior to humans in several studies.3,8 Future Another challenge to the acceptance of deep advances in the accuracy of prediction models will learning algorithms in medicine is the difficulty likely come from the use of large longitudinal datasets in determining the basis for clinical decisions drawing on multiple data sources, together with the made by these systems, informally described as development of more advanced AI systems.3 the “black box” problem. Visualisation tools have Despite these significant advances, AI systems are not been developed to assist clinicians by highlighting in widespread clinical use and in some cases real-world the salient image features that contribute to the AI performance has been inferior compared with in silico system classification (Box 2).12 This has the potential validation.2,3 Training and validation of deep learning to create trust in system-generated decisions, algorithms with large, representative data (eg, data particularly if the features correspond with those from people of different ethnicities) acquired using used by experienced clinicians for clinical decision multiple devices (eg, different retinal camera models) making.14 and data collection protocols (eg, retinal photographs Interpretability is particularly important when acquired with and without pupil dilation) are key to considering legal liability in the event of patient achieving clinical applicability.4,5 This approach was harm arising from the use of AI in medicine. In used in the development of deep learning systems for traditional malpractice cases, a physician may be retinal photographic screening for diabetic retinopathy, asked to justify the basis for a particular clinical AMD and glaucoma which are now being used in large decision and this is then considered in light of scale screening programs in Singapore and China.4,5 In conventional medical practice.15 In comparison, these programs, AI is used to identify images without challenges in identifying the basis for a given evidence of disease, so that human graders can focus decision made by AI might pose problems for their efforts on the images of those with disease, clinicians whose actions were based on that decision. enabling improved efficiency and cost savings.12 The extent to which the clinician, as opposed to the technology manufacturer, should be held accountable Challenges to the clinical adoption of artificial for harm arising from AI use is a subject of intense intelligence Several obstacles to the adoption 2 Original retinal photograph of right eye with macular degeneration (A). of AI in health care remain. The Heat map of image A showing visualisation of traditional features associ- training of deep learning systems ated with macular degeneration, such as central scarring (B). Original retinal requires access to large amounts photograph of left eye with referable diabetic retinopathy (C). Heat map of of medical data which has image C showing visualisation of traditional features, such as micro-aneu- significant implications relating rysms and haemorrhages (D) to privacy and data protection. In the context of ophthalmology, this is particularly pertinent, as the retinal vasculature may be considered biometric data, making it impossible to completely anonymise retinal photographs.3 Furthermore, characteristics that are not visible to human examiners, such as age and sex, can now be accurately predicted from a single retinal photograph using deep learning.1 Several recent major breaches of data protection laws relating to AI system development 13 1 March 2021 1 March have already come to light. ▪ ) While individual patient data 4 ( used to train an algorithm do 214 not remain within the system, MJA incorrect handling and sharing of data may lead to patients withdrawing consent to the use 156 of their data under General Data Protection Regulation laws. It Perspectives debate.15 Factors such as the manner in which these development, Australia is aiming to advance its AI systems are used and their classification as either AI competitiveness. These framework documents products or software are likely to have important provide guidance for developers, clinicians and health bearings on how cases are litigated.15 Further care consumers to navigate this rapidly evolving field. challenges for existing regulatory frameworks come Broad dissemination of these documents should form from algorithms that continue to learn and evolve