Artificial intelligence (AI) has demonstrated some promise in improving diagnostic and grading accuracy of various ophthalmologic conditions, but it also presents myriad challenges for eye care professionals, according to a review published in the Journal of Optometry.
In recent years, advances in retinal imaging, such as optical coherence tomography (OCT), have led to an increase in raw data. And interpreting that data has become more complex for clinicians. Incorporating AI techniques, which include machine learning and deep learning, has the potential to improve and streamline the diagnostic and disease staging processes in eye care.
At least a third of patients with diabetes will develop diabetic retinopathy (DR), the report states. Early screening and timely treatment initiation are standard practice for decreasing the risk for blindness among this patient population. Deep learning can be used to assess data obtained through multimodal imaging procedures, including OCT, OCT-angiography, fundoscopy, and fluorescein angiography and provide DR diagnosis and grading. Studies indicate that deep learning-based algorithms are noninferior, if not superior, to the diagnostic performance of clinical specialists. One study found that an AI model had a 96.8% sensitivity and 87% specificity for detecting referable DR. Similarly, another study found that an AI model had a 94% sensitivity and 98% specificity for grading DR.
In patients with age-related macular degeneration (AMD), combined imaging techniques are also used for detection and classification. As with DR, previous studies have found potential clinical utility of AI-based models for staging atrophy in eyes with AMD. Several studies have also found that AI algorithms demonstrate the ability to detect findings in images of eyes with AMD that were not otherwise visible.
In the last decade, AI-based diagnostic algorithms for glaucoma have increased exponentially. One study found a sensitivity of 95.6% and specificity of 92% for detecting glaucoma using fundus photography as input data.
These technologies also have a potential role in identifying retinopathy of prematurity (ROP). AI-based models have been used to refine classification metrics. One such model demonstrates 95% accuracy in diagnosing ROP — outperforming the diagnostic accuracy of 10 out of 11 individual clinical experts.
Despite these favorable trends, significant challenges remain for implementing AI-based models in eye care. First, the output of an AI model can only come to conclusions based on the input data. Input data sets are often collected from relatively homogenous populations, which may affect the generalizability of algorithm results.
Second, AI models require a large amount of data to ‘learn’ from in order to come to robust conclusions. This may be problematic for identifying rarer pathologies, such as ocular tumors or inherited retinal dystrophies.
Finally, researchers state that it is often unclear what features the algorithms are selecting as important predictors and why, complicating output interpretation.
“Deep learning models have been proven to be the current state-of-art in the field of AI applied to ophthalmology,” according to the researchers. “To date, despite promising results from deep learning models, some challenges persist, and further clinical studies are needed to overcome these limitations as well as to accurately assess the impact of AI and deep learning models applied to the ophthalmological field.”
Desideri LF, Rutigliani C, Corazza P, et al. The upcoming role of artificial intelligence (AI) for retinal and glaucomatous diseases. J Optom. Published online October 8, 2022. doi:10.1016/j.optom.2022.08.001