Artificial intelligence (AI) has permeated the medical field to enhance the experiences
of clinicians and patients. Jiang et al. summarized the major publications and authors
working on AI development in the ophthalmology field so far (1). To date, most studies
have focused on designing models that help doctors make fast screenings or arrive
at accurate medical decisions by learning from existing clinical data based on machine
learning and convolutional neural networks (CNN) (2). Therefore, institutions that
have extensive data on existing research medical have quickly developed AI models
and have emerged as leaders in academia in this field. Most AI studies have targeted
diabetic retinopathy, age-related macular degeneration, glaucoma, and cataracts based
on epidemiological research data combined with large-scale fundus photographs and
optical coherence tomography (OCT) (3,4). Based on the advances of deep learning technology,
AI-based systems such as IDx-DR (Digital Diagnostics, Coralville, IA, USA) and EyeArt
(Eyenuk, Los Angeles, CA, USA) have been approved for diabetic retinopathy detection
by the US Food and Drug Administration (5). Several clinical trials have demonstrated
that these AI systems can serve as low-cost point-of-care diabetic retinopathy detection
tools (6). AI has been used in various fields, such as refractive correction surgery
(7), ocular disease treatment (8), and oculomics (9). The AI models developed so far
have focused on decision support systems for doctors.
Future perspectives on generative AI
In addition to existing data analyses based on the aforementioned technologies, AI
technology that generates synthetic data has recently been introduced (
Figure 1
) (10). In the future, generative AI systems that promote patient convenience rather
than doctors will be actively used in the new era of AI. Since the advent of ChatGPT,
the public has been able to easily access and interact with AI on chat-like interfaces,
which have rapidly accelerated its progress in all areas, including medicine (11).
ChatGPT was developed based on GPT-3, a large language model trained to generate realistic
texts. Based on its high level of natural language processing ability, it shows the
ability to understand unstructured data, new content generation ability, and high
versatility. Eventually, it is expected to quickly penetrate medical AI fields related
to ophthalmology and expand its scope from simple questions and answers (Q&As) to
diagnostic care assistance.
Figure 1
History of development of generative AI. AI, artificial intelligence; VAE, variational
autoencoder; GAN, generative adversarial network; DCGAN, deep convolutional generative
adversarial network; VQ, vector quantized; ViT, vision transformer; DDPM, denoising
diffusion probabilistic model; DDIM, denoising diffusion implicit model; GLIDE, guided
language-to-image diffusion for generation and editing; GRU, gated recurrent unit;
GPT, generative pre-trained transformer; BERT, bidirectional encoder representations
from transformers; LaMDA, Language Model for Dialogue Applications; LLaMA; Large Language
Model Meta AI.
Wang et al. conducted several studies based on a basic transformer architecture to
analyze electronic medical records (12,13). Medical image generation using generative
adversarial networks (GAN) has been conducted for ophthalmology imaging (14). Until
now, these language processing and generative technologies have been used by ophthalmologists
to solve detailed tasks; however, it is expected that patient-centered AI will be
developed in the future, starting with ChatGPT. First, there is an era of searching
for and inquiring about patients’ ophthalmic knowledge based on chatbots (15). The
future advanced chatbot model does not simply refer to medical text data but also
to the analysis of medical records, including ophthalmology images, and notification
of analysis results. Secondly, patients can predict the course of future diseases;
therefore, they can use various ophthalmological imaging domains based on generative
AI (16). This can convey the characteristics and prognosis of the disease to patients
more effectively than simple language delivery. For example, patients can understand
how the shape of their macula changes after retinal surgery and decide on the surgery
(17). The development of diffusion models beyond GAN will make this possible in the
future (18,19). Third, smoother communication in telemedicine is possible through
new interactive tools based on generated AI, such as virtual humans (digital humans)
(20). Virtual doctors can be shown to patients naturally through generative AI, without
being constrained by language, using text-to-speech technology. In areas where doctors
are scarce, patients can easily understand various types of medical information through
virtual humans.
Limitations of generative AI
However, generative AI currently has certain disadvantages. Currently, conventional
CNN models can operate on personal computers with graphics cards; however, most large
generative models, including ChatGPT, require a large amount of computation and can
operate only in large computational environments. This limits the dissemination of
new technologies. In addition, shortcomings that must be addressed include difficulties
in providing the latest information, frequent errors, and insufficient performance.
Owing to the difficulty of learning large models, the latest medical knowledge, especially
recent research results, is likely to have difficulty providing information. The provision
of inaccurate information in the medical field can cause fatal problems. Although
diffusion models continue to appear in the field of image generation following GAN,
they still exhibit insufficient performance for generating realistic images. Though
so far, the importance of training data has been reported, but the algorithms have
not overcome the limitations of small data. Therefore, it is necessary to continue
development through continuous data collection and advanced algorithm development.
Conclusions
The article by Jiang et al. summarized the flow of international AI research in ophthalmology
over the last decade (1). Just in time, AI in academia and industries are also facing
a new phase. AI development in the field of ophthalmology is also expected to enter
a new era that will not simply help doctors improve screening accuracy. Patients are
increasingly able to use AI and enhance their lives. In the future, research groups
that use large generative models effectively and are competent at fine-tuning to adapt
large foundation models to their specific tasks will emerge as new leaders in the
field of ophthalmology and AI.
Supplementary
The article’s supplementary files as
10.21037/atm-23-1598