Inherent conflicts of interest should be acknowledged. Manufacturers who develop and market

SaMD have a strong financial interest in presenting their products positively. Thus, conflicts of interest exist if they fund, conduct, and publish results of

studies, including those that might report deficiencies in their products. Many of the published papers in the field of AI-based diabetic retinopathy screening, particularly those using CE-marked and FDA-approved algorithms, were conducted by manufacturers or patent owners.

It should be also remembered that AI algorithms can be designed to perform in unethical ways. For example, Uber’s software Greyball allowed the company to identify and circumvent local regulations, and Volkswagen’s algorithm allowed vehicles to pass emission tests by reducing their

emissions of nitrogen oxide during testing. AI algorithms could be tuned to generate increased profits for their owners by recommending particular drugs, tests, or the like without clinical users’ awareness. AI systems are vulnerable to cybersecurity attacks that could cause their algorithms to misclassify medical information [³¹].

Seven essential

factors to design AI for social good were proposed by Floridi et al. (Table 1.5) [³⁸]. The authors propose falsifiability as an essential factor to improve the trustworthiness of technological application, i.e., for an

SaMD to be trustworthy, its safety should be falsifiable. Critical requirements for a device to be fully functional must be specified and must be testable. If falsifiability is not possible, then the critical

requirements cannot be checked, and the system should not be deemed trustworthy [³⁸].

Essential

Factors to design AI for social good

ability for people to semanticise (that is, to give meaning to, and make sense of) something | Autonomy |

Source: Floridi L, Cowls J, King TC, Taddeo M. How to Design AI for Social Good: Seven Essential Factors. Sci Eng Ethics. 2020 Jun;26(3):1771–1796. Springer

One of the arguments for AI-based medical

devices is that they can reduce medical costs and eliminate unnecessary procedures. A study from Singapore found that a semiautomated model that combined a DL system with human assessment achieved the best economic returns, leading to savings of 19.5% in screening for diabetic retinopathy. An earlier study from the UK reported cost-savings of 12.8–21.0%; however, a simple comparison between them is not possible due to the different models of DR screening in the two countries (two-stage screening in Singapore and three-stage screening in the UK), and their different DR classification systems. The authors of both studies argued that a semiautomated system produces more savings than a fully automated system due to the lower rate of false positives and unnecessary specialist visits [³⁹, ⁴⁰].

This book aims to provide ophthalmologists and other visual professionals and researchers with an

overview of current research into the use of AI in ophthalmology. Together with a team of international experts from Europe, North America, and Asia, we present an overview of the most important documentary research in ophthalmology on ML and AI technologies and their benefits. We discussed the use of AI in the diagnosis of some retinal and corneal disorders, the diagnosis of congenital cataract, neuro-ophthalmology, glaucoma, intraocular lens calculation methods, ocular oncology, ophthalmology triaging, cataract-surgery training, refractive surgery, and the assessment and prediction of systemic diseases through the use of the eye. Chapters on digital-image analysis, AI basics, and technical aspects of AI provide the reader with knowledge not commonly possessed by ophthalmologists, but required to understand the topic in both its

field-specific and broader contexts. The very important chapter on AI safety and efficacy outlines the challenges ophthalmology will face with the introduction and widespread dissemination of this technology. Although we have covered all of the major areas of AI/ML technology in ophthalmology, research in this field is progressing so quickly that some new concepts that emerged at the end of 2020 and in early 2021 do not appear on these pages. However, evidence-based medicine often demands that we await for more evidence to verify early reports and assess the real value of new medical technologies or applications. I would like to thank all the contributors for sharing their knowledge in this new and fascinating discipline, which has great

potential to change ophthalmology.

Acknowledgements

I would like to thank Aleksandra Lemanik, Foundation for Ophthalmology Development, Poznan, Poland and Tomasz Krzywicki, Faculty of Mathematics and Computer Science, University of Warmia and Mazury, Olsztyn, Poland for their help in preparing illustrations, and Szymon Wilk, Faculty of Computing and Telecommunications, Poznan University of Technology, Poznan, Poland, for his valuable discussion on this chapter.

References