Will AI-analyzed imaging replace glaucoma diagnosis and monitoring by live clinicians?
Click here to read the Cover Story, "AI in glaucoma: Area ripe for research with unlimited potential."
AI will make glaucoma diagnoses
Artificial intelligence will replace glaucoma diagnosis by live clinicians.
This is no debate on artificial vs. human intelligence. In the past, phones were hooked to landlines, and the thought of a “smartphone-camera-GPS-other” in a single wearable device was unimaginable. AI is a reality, and its capabilities and applications are not debatable. In fields such as pathology and radiology, AI has wide applications and serves to increase accuracy and efficiency. Consider this: While AI can work 24/7, physicians simply cannot.
In ophthalmology, image analysis by physicians leads to the diagnosis of vision-threatening diseases such as glaucoma, diabetic retinopathy and age-related macular degeneration. AI may be able to make the association of images with a differential diagnosis as well as or sometimes better than humans. It is important to state here that AI is not error-proof. It may be great at identifying horses, but not as good with zebras. Ophthalmology and AI are a great fit, and there is no doubt that the technology will play a central role in our specialty.
Glaucoma is asymptomatic until advanced. As such, an unacceptable number (by some estimates half or more) of affected patients remain undiagnosed. Early detection has the biggest impact on prognosis and preserving functional vision. To build on the radiology and pathology example noted above, the sheer workload involved in physician-based “manual” review of images and visual field tests is tremendous. Add to it the associated confounding factors (test reliability and ethnic variability, among others) that further challenge diagnostic accuracy. Therefore, AI-based algorithms for improving diagnostic efficiency are an attractive solution.
Many AI strategies provide promising levels of accuracy (specificity, sensitivity, area under the ROC curve). AI often outperforms expert observers. For the detection of glaucoma, AI can rely on OCT, fundus imaging, visual field testing or a combination of tests. Ongoing work in AI combines structural and functional inputs with other data, including age, ethnicity, IOP, corneal biomechanical properties and other data points to further improve diagnostic ability. The potential with AI then extends beyond detection to monitoring for progression.
From a cost-efficiency perspective, the diagnosis of glaucoma is a complex and expensive effort that is dependent on clinical expertise. AI techniques can have a role in reducing cost, expanding access, and identifying and triaging the most urgent cases to the front line for timely treatment, including surgery. Robotics and AI will replace certain tasks physicians currently do. There are, however, regulatory, legal and ethical challenges that have to be resolved as we embrace AI in health care. What remains clear is that AI is in our future, and by embracing it, we can enhance the role of physicians and reshape the future of glaucoma care.
- Devalla SK, et al. Br J Ophthalmol. 2020;doi:10.1136/bjophthalmol-2019-315016.
- Jammal AA, et al. Am J Ophthalmol. 2020;doi:10.1016/j.ajo.2019.11.006.
- Mursch-Edlmayr AS, et al. Trans Vis Sci Technol. 2020;doi:10.1167/tvst.9.2.55.
- Zarbin MA. Trans Vis Science Technol. 2020;doi:10.1167/tvst.9.2.1.
- Zheng C, et al. Curr Opin Ophthalmol. 2019;doi:10.1097/ICU.0000000000000552.
Albert S. Khouri, MD, is professor of ophthalmology, director of the glaucoma division and director of ophthalmology residency at Rutgers New Jersey Medical School.
AI will be a tool, not a replacement
As methods for health care delivery continue to evolve, there is no question that artificial intelligence-analyzed imaging will play a significant role in glaucoma diagnosis in the future.
The diagnosis and management of glaucoma, however, is fraught with a number of challenges that will not be overcome by the integration of AI alone. Glaucomatous optic neuropathy and its primary risk factor, elevation of IOP, often do not cause symptoms before the onset of moderate to advanced visual field loss. As a result, the diagnosis of early disease, before the onset of visual field abnormalities or in the presence of confounding conditions such as neurodegenerative diseases, often presents a clinical challenge.
Unfortunately, there is no single gold standard test for diagnosing glaucoma and training AI algorithms. Instead, the diagnosis is based on several components that may include, but are not limited to, a combination of IOP, optic disc evaluation, visual field testing, OCT, corneal pachymetry and gonioscopy. While AI will certainly have a role in refining the interpretation of current imaging modalities (eg, is this abnormality on OCT of the retinal nerve fiber layer related to glaucoma or another disease entity?), AI remains limited in its ability to integrate data from multiple test and exam components to reach a consensus decision like a skilled clinician.
Furthermore, it is important to remember that monitoring and treatment differ from diagnosis and screening. Glaucoma is, after all, an umbrella term for a group of diseases with varying phenotypes and both structural and functional components. The skill of the well-trained glaucoma clinician lies in his or her ability to not only determine the presence or absence of glaucomatous optic neuropathy, but also the type of glaucoma (open angle vs. angle closure; primary vs. secondary). In doing so, it is the clinician, and not an AI algorithm, that is best able to manage the disease based on both the choice of the most appropriate therapy for an individual patient and also the safe and effective delivery of the chosen treatment.
While the AI-analyzed image may one day be able to identify and diagnose the glaucomatous nerve associated with visual field loss, monitoring for progression and making management decisions is an entirely different challenge.
In the Early Manifest Glaucoma Trial, the vast majority of patients who demonstrated evidence of disease progression did so based on functional visual field outcomes alone (53 of 58 patients in the treatment group and 64 of 78 patients in the control group) as opposed to structural optic disc or combined optic disc and visual field outcomes. In a 2016 editorial published in Ophthalmology, Medeiros and Tatham wrote: “If methods for assessing structural and functional progression were to agree perfectly, there would be no need for using both in monitoring progression. One test would suffice.”
It is not uncommon to see some degree of disagreement between structure and function in glaucoma, and herein lies the challenge in attempting to consolidate functional testing with AI-analyzed imaging for the purpose of diagnosis, management and staging.
Furthermore, structural testing and imaging, such as OCT of the RNFL or ganglion cell complex, are limited by the “floor effect,” which limits the utility of such parameters for monitoring of more advanced disease, irrespective of the integration of AI for the interpretation of such imaging. It is undoubtedly an exciting time to be a glaucoma clinician, but despite the many diagnostic and therapeutic advancements that have occurred recently in our field, there does not appear to be a reason that AI will pose a threat to the current need for live clinicians. Instead, AI-analyzed imaging will be an important tool to assist, but not replace, live clinicians in their ongoing mission to diagnose and treat glaucoma in the future.
- Heijl A, et al. Arch Ophthalmol. 2002;doi:10.1001/archopht.120.10.1268.
- Hood DC et al. Prog Retin Eye Res. 2007;doi:10.1016/j.preteyeres.2007.08.001.
- Lavinsky F, et al. Ophthalmology. 2018;doi:10.1016/j.ophtha.2018.05.020.
- Medeiros FA, et al. Ophthalmology. 2016;doi:10.1016/j.ophtha.2016.03.023.
Brian J. Song, MD, MPH, FACS, is assistant professor of clinical ophthalmology in the department of ophthalmology at USC Keck School of Medicine.