Applications of generative artificial intelligence (AI) in ophthalmology continue their rapid and inexorable progress.
It was a few months ago that the study "Deep Learning to Discriminate Arteritic From Nonarteritic Ischemic Optic Neuropathy on Colour Images", published in Jama Ophthalmology, which examined the accuracy with which deep learning-based algorithms are able to discriminate non-arteritic ischaemic optic neuropathy (NAION) from the arteritic form (AAION) from colour images of the fundus in the acute phase. These are the results of an international study that enrolled 961 eyes of 802 patients. The training phase was carried out using images from 21 neuro-ophthalmology centres in 16 countries, while the test phase involved a cohort of five specialised centres in Europe and the United States, over the period August 2018 to January 2023. The performance of the Deep Learning System (DLS) according to the criteria of sensitivity, specificity and accuracy yielded excellent results with specific activation maps with higher levels of accuracy than 90% in distinguishing between AAION and NAION.
This is one of many successful examples of the countless applications of AI in Ophthalmology, to which an extensive review published in the Asian Pacific Journal of Ophthalmology has been devoted, Latest developments of generative artificial intelligence and applications in ophthalmology - ScienceDirect
The last two decades have, in fact, witnessed a great development that has led to ever larger datasets containing an ever wider spectrum of information and to the creation of ever more powerful models with greater capacity to learn from data. From machine learning to deep learning, from discriminative AI to generative AI and the emergence of Large Language Models (LLMs).
From this perspective, what are the most important directions and problems in ophthalmology?
The first issue is the overall verification of the standard survey model. Current studies have mostly evaluated the performance of generative AI on the basis of existing databases in ophthalmology, as the only form of validation. The limitation of this approach is probable shortcomings in the diversity and complexity of the available case histories. The closer the performance of generative AI approaches human performance, the more pressing the need for the validation context of the algorithms to reflect real-world clinical conditions becomes. Moreover, AI outputs are generally random, leading to problems of replicability of resultswhich is an increasingly important requirement.
It is then necessary to increase supporting evidence as only a limited number of studies on AI applications have been reported and verified through peer review. At present, a multitude of studies are made available as preprints on open access platforms without supporting verification. Furthermore, there is not enough evidence to date for a "consensus" on which ophthalmologists and AI experts agree regarding the results of generative AI in ophthalmology and its future development.
Another problem is that the still limited capacity of generative AI to integrate multimodal information and perform multimodal tasks in the clinical practice of ophthalmology. Multimodal generative AI will be a crucial technology to achieve personalised medicine, in which generative AI models need to integrate different types of data, including biosensors, genetic, epigenetic, proteonomic, microbiological, metabolic, imaging, text-based information, clinical data, and social and environmental factors. The 'intelligent agent' is an emerging multimodal generative AI solution to handle multiple tasks, ranging from simple, rule-based tasks to complex tasks in response to a dynamic context. In ophthalmology, as in other areas of medicine, the intelligent agent must aim at specific goals, perceive its context and collect the necessary data for decision -making. For example, the intelligent agent has already been used in emergency room triage to optimise the use of available resources according to patient urgency, in clinical decision-making in the case of adverse drug rations, and in the remote monitoring and care of dementia patients. Similar applications could be developed for care in ophthalmology.
A further step is the development of a integration workflow of generative AI algorithms in routine ophthalmological clinical practice.
On the subject of AI in Ophthalmology, we point out:
- AI & Oculomics - Oculista Italiano
- AI and ophthalmology: listen to the podcast - Oculista Italiano
- Here come the robots! - Italian Ophthalmologist
- Feng X, Xu K, Luo MJ, et al. Latest developments of generative artificial intelligence and applications in ophthalmology. Asia Pac J Ophthalmol (Phila). 2024 Jul-Aug;13(4):100090. doi: 10.1016/j.apjo.2024.100090. Epub 2024 Aug 14. PMID: 39128549.
- Gungor A, Najjar RP, Hamann S, et al. Deep Learning to Discriminate Arteritic From Nonarteritic Ischemic Optic Neuropathy on Colour Images. JAMA Ophthalmol.2024;142(11):1073–1079. doi:10.1001/jamaophthalmol.2024.4269