With companies such as Google, Microsoft, Meta, Amazon, and Apple making significant investments in artificial intelligence (AI) products, it is clear which direction the future is heading.1
It is up to medicine to embrace these advancements, and in a field such as ophthalmology, where imaging and technology are already so prevalent, we will inevitably see integration into daily practice. The president of the American Medical Association put it elegantly; “AI will never replace physicians—but physicians who use AI will replace those who don’t.”2
This article will provide an evidence-based review of foundational information relating not only to AI, but other emerging topics, including precision medicine, traditional machine/deep learning, and geometric deep learning as well as their current/potential applications in ophthalmology and healthcare.
What is precision medicine?
Although definitions of precision medicine—also commonly referred to as personalized, individualized, or targeted medicine—vary, it is broadly understood to be the use of diagnostic tools and treatments targeted to the needs of the individual patient based on genetic, biomarker, or psychosocial characteristics.3
Precision medicine is therefore an incorporation of technology into medicine to better identify and treat an individual patient’s disease.3 The topic of precision medicine has gained substantial popularity over the years, reflected by publications growing exponentially since the early 2010s (Figure 1).
Figure 1: Frequency of PubMed-indexed publications on precision medicine from the years 1979 (the first believed mention of "precision medicine") to 2024.
It promises the optimization of care for patients through a blend of molecular biology (e.g., genomic sequencing) and “big data” analysis which includes large volumes of biomedical data from electronic health records, of which AI can dramatically speed up the process, as will be discussed.4
It is essentially an evolution of evidence-based care, using modern technology and methods to reduce errors and individualized medicine.4 Compared to the symptom-driven approach in medicine, precision medicine considers the critical fact that all patients do not react to the same treatment or medication in the same way.5
The impact of precision medicine in ophthalmology
Ophthalmology has various avenues in which precision medicine has/will play a role. With regard to genomics, utilizing the concepts of precision medicine can lead to diagnosing a specific disease, as in the case of Leber congenital amaurosis, an inherited retinal disorder.6 Furthermore, the discovery of the role of specific genes has led to a new era of treatments in ophthalmology, known as ocular gene therapy.7
Studies are ongoing for the effectiveness of these treatments and there is great optimism in the field for an increase in clinical trials over the upcoming years, culminating in more treatment options for more genes.7 Other clinical trials have studied single-gene inherited retinal diseases, including X-linked retinoschisis, Stargardt disease, Usher syndrome, choroideremia, and retinitis pigmentosa.6
Another area employing precision medicine is ophthalmic oncology. Tumor biopsy of ciliochoroidal melanoma at the time of treatment has been used to confirm the diagnosis and also assess the DNA/RNA characteristics associated with high or low risk of metastasis.6
This information, combined with tumor pathology, patient family history, medical history/behavior, and lifestyle/environmental factors, allows for the best possible care for the individual patient and their circumstances.6 In the future, tumor genomics may lead to targets for molecular therapy, as is the case in some other cancers.
More broadly, genetically defined cohorts of patients with a multifactorial disease, such as age-related macular degeneration (AMD), can be studied with precision medicine to identify risk factors, treatment options, and other patterns that might help with earlier diagnosis and better prognosis.6
A breakdown of intelligence and learning models
Precision medicine involves the analysis of big data sets and the identification of patterns and predictions. This is where AI can enter into the picture and enhance this practice method.
Artificial intelligence (AI)
This term is used to describe the use of computers and technology to simulate intelligent behavior and critical thinking comparable to a human being.8 While the human mind has constraints in learning due to the time and experience required, computer software, through algorithms, can gain far more experience in a significantly shorter amount of time based on vast amounts of stored data.9
AI is an umbrella term for using software-based algorithms to achieve a certain task.
Artificial neural networks
Artificial neural networks are computational models that mimic the way biological neural networks process information in the human brain. Neural networks are capable of processing several types of data and creating patterns for use in a decision-making process.9 It is a system of virtual neurons connected in many layers that take inputs, process them, and provide outputs.9
There are weights and biases introduced that determine how much influence one neuron in the system has on another and there are functions built in to determine when a neuron should be activated or not based on the input. They are capable of recognizing patterns and learning from data.
Machine learning
The ability of a computer to learn from experience and modify its processing based on newly acquired information.9 This allows the computer to learn and make decisions or predictions without being specifically programmed for each task.
Machine learning involves feeding an algorithm a large amount of data and letting it find patterns.9 An example would be providing a dataset with images labeling cases of normal eyes and eyes with conjunctivitis so that the model learns to distinguish them.
This model should then have the rules learned to make a prediction when given unlabeled images and the accuracy can be checked to fine tune the system.
Deep learning
A subset of machine learning using artificial neural networks to analyze and learn from large amounts of data.10 It is called “deep” learning due to the multiple layers of processing that occur within the network.10 In deep learning, the representations are learned from raw input data, rather than manually extracted features of the data (e.g., labels), as discussed in our example for machine learning.11,12
Deep learning has been found to outperform conventional machine learning when dealing with complex, compute-intensive tasks such as image analysis and natural language processing.11
How AI and deep/machine learning apply to clinical practice
With definitions out of the way, the question is how can all of these concepts apply to ophthalmology and clinical practice?
Some examples are covered below:
Diabetic retinopathy
Multiple different studies have used AI for detecting referable diabetic retinopathy (DR) in standard color fundus photography, ocular coherence tomography (OCT) images, ultra-widefield (UWF) imaging, and even smartphone-captured retinal images with high sensitivity and specificity (90%+).13
Age-related macular degeneration
Similar to DR, there have been multiple studies incorporating machine learning of databases of hundreds of thousands of retinal images to develop a deep learning algorithm for automated detection of AMD.13
There has also been work done on AI algorithms to accurately quantify the volume of fluid in neovascular AMD, and other important AMD features such as intraretinal fluid, subretinal fluid, pigment epithelial detachment, drusen, fibrosis, ellipsoid zone loss, and subretinal hyperreflective material.13
Another study by De Fauw et al. used a deep learning model to identify referable retinal disease in general (such as neovascular AMD, geographic atrophy, drusen, macular edema/holes, epiretinal membrane, vitreous traction, and central serous retinopathy) using OCT images.14
Conjunctivitis
AI has been used in the form of a neural network trained in grading the severity of conjunctival hyperemia with high degrees of accuracy.15
Dry eye
Studies have been done on deep learning in meibography images and machine learning in tear meniscus thickness using OCT with reproducible and high grading accuracy.13
Keratoconus
Multiple studies using neural networks, machine learning, and deep learning have demonstrated high accuracy in identifying keratoconus, even subclinical keratoconus, using corneal topography and OCT images.13
Glaucoma
An area with a significant amount of AI research is glaucoma. A deep learning algorithm trained with OCT to assess monoscopic optic nerve photographs was able to outperform glaucoma specialists in terms of accuracy in identifying glaucomatous optic nerve damage.16 Several other studies have shown agreement in the ability of AI to detect glaucomatous optic nerve changes.13
In addition to identifying damage alone, progression has been detected on fundus photographs using deep learning algorithms confirmed by OCT.13 Finally, machine learning in perimetry has been able to detect glaucoma ahead of diagnosis.13,17
Pediatrics
In the world of pediatric ophthalmology, AI has been used in the early detection of retinopathy of prematurity, in the prediction/scheduling of appointments/follow-up of congenital cataracts, and in detecting amblyopia.13
How have researchers used ChatGPT for investigations?
Some other areas that have been explored in ophthalmology surround large language models such as ChatGPT developed by OpenAI. It has shown some ability to serve as an assistant offering diagnostic and therapeutic suggestions based on patient data and assisting in triage.18 It can generate practice questions, provide explanations, and create patient education materials.18
In research, it can facilitate literature reviews, data analysis, manuscript development, and peer review.18 All of these applications come with important drawbacks to consider with regard to data security concerns, accuracy of results, ethics, and bias.18 Research is active in these areas within ophthalmology and medicine in general.
What is geometric deep learning?
Geometric deep learning is a subtype of machine learning that deals with geometric data, such as graphs or manifolds and typically preserves the invariance of geometric data under transformations and can be applied to 3D structures.19 In other words, traditional machine learning works well with data in neat, grid-like structures (e.g., a spreadsheet or a photo).
In contrast, geometric deep learning works with data that doesn't fit into these neat structures, such as a social media network of friends, the 3D shape of a molecule, or functional networks in brain imaging.20 Geometric deep learning models can therefore understand and make predictions based on complex shapes and networks.
For example, a model could understand how people are connected within social media and then make predictions for friend recommendations. In a more medically related example, such as drug discovery, geometric deep learning models can be used to predict structures for new medications.21
The term used for the types of data used in traditional machine learning (i.e., spreadsheets and 2D images) is Euclidean data. For the data analyzed by geometric deep learning models (3D shapes, graphs), the term is non-Euclidean data.20 Euclidean data remains useful and well-suited for traditional machine and deep learning applications such as speech, image, and video signals however has limitations when applied to more complex, real-world scenarios.20
Euclidean data fails to capture the complexity of molecular structures due to their 3D arrangements, for example, and forcing such data into an Euclidean structure would result in a loss of important information and relationships.
Two more key concepts in geometric deep learning are graph convolutional networks and manifolds.
A graph convolutional network (GCN)
GNC is similar to a convolutional neural network (CNN) which is used for Euclidean data and is a neural network layer that gathers and processes information from surrounding pixels, known as the “receptive field,” to create a condensed, lower-dimensional representation.22
A GCN is similar, however, instead of gathering and processing information from neighboring pixels, it uses neighboring nodes in the graph.22 The resulting information being aggregated can be far more complex as each node may contain detailed data sets.
For example, if a node was a person in a social network, the person would have all kinds of details about where they were born, went to school, occupation, relationship status, interests, etc. A GCN in that setting would gather all of that information of the individual and do the same for all other individuals, process it, and make it useful for accurate friend recommendations, as one example.
Manifold
A manifold is a mathematical space that, at a small enough local scale, resembles Euclidean space (flat surfaces like a plane), but can have a more complex global structure. Manifolds are helpful for modeling data on curved surfaces or in high-dimensional spaces. For a more in-depth and technical explanation we would refer the reader to a crash course article previously written.23
A simple example would be the surface of the Earth when viewed from a small area, appears flat; however, when all these small areas are combined, the Earth becomes a sphere, not flat. A similar line of thinking would apply to the curvature of the cornea or retina of the human eye, making manifolds an important concept for geometric deep-learning applications in ophthalmology.
New frontiers: Geometric deep learning in ophthalmology
To date, there are very few studies that have been published using geometric deep learning in ophthalmology. Of these, most focus on diagnosing glaucoma using the 3D structure of the optic nerve head.24-27
In a study by Thiéry et al., they were able to demonstrate a geometric deep learning technique that requires significantly less information as input to perform better than a 3D CNN, and with an area under the curve superior to that obtained from retinal nerve fiber layer thickness measurement, which may lead to improved and more simplistic diagnosis and prognosis applications in glaucoma.24
One might think that a similar application of geometric deep learning to the 3D structure of the cornea may lead to a more accurate and simplified diagnosis of diseases such as keratoconus. Or perhaps in assessing the macular structure and detecting early disease/monitoring progression and treatment responses.
Where this technology is still young, the possibilities are unknown and represent an exciting area to target for research in the future.
