Artificial Intelligence in Neuro-Ophthalmology
Artificial intelligence (AI) is defined as human critical thinking and intellectual pursuits augmented by utilizing synthetic intelligence technology. With the advent of AI in the 1950s via the help of Alan Turing, AI has skyrocketed in fields such as travel and transportation, finance, shopping, and healthcare technology.  In order to understand potential applications of AI in the field of medicine, and specifically ophthalmology, it is necessary to acknowledge the mechanisms behind AI. Through each succession of data processing, AI systems assess their prior performances to hone their predictive capabilities and improve their accuracy. The ability to perform thousands, if not millions, of calculations and tasks without needed breaks is a benefit of AI over human counterparts.  However, it is important to note that AI is not simply a computer program; it is an ever-evolving process consisting of many moving parts.
Within the field of AI, machine learning and deep learning are two subsets. Machine learning requires hand-crafted features provided by the programmer, which the model utilizes to optimize its learning process and automate outcome prediction with limited data sets. In contrast, deep learning can automate feature extraction and learning but requires considerable data to reach near-perfect accuracy. A deep learning system (DLS) improves upon previous machine learning-based approaches by incorporating multi-stage learning mechanisms to extract manifold features for a better outcome. This module consists of multiple small and large receptive fields connected and stacked on top of each other, similar to a human brain creates connections between neurons to improve processing outcomes.  Understanding and harnessing the power of AI and the neural network is appealing to many fields given the fact that it has the potential to reveal great advancements in an abbreviated amount of time. AI has already permeated many subspecialities within ophthalmology, which we will now discuss.
Current Uses of AI in Ophthalmology
The subspeciality of glaucoma is a rapidly evolving field, with new technologies being consistently reported in the literature.  An initial instance of AI implementation in glaucoma involves measuring intraocular pressure (IOP), with increased levels of IOP leading to glaucomatous changes like an enlarged cup/disc ratio. A continuous monitoring contact lens has been developed to better define IOP parameters for monitoring of glaucoma progression.  Furthermore, many deep learning algorithms have been developed for fundus photography, measuring parameters such as disc size, cup size, cup/disc ratio, and neuroretinal rim thickness. In fact, one DLS outperformed 5 out of 6 ophthalmologists in identifying glaucoma by using these criteria. AI has also permeated imaging such as optical coherence tomography (OCT) and visual field testing, with both achieving 93% sensitivity in detecting glaucoma.
Three major topics in AI in the retina subspecialty include diabetic retinopathy (DR), retinopathy of prematurity (ROP), and age-related macular degeneration (AMD). For diabetic retinopathy, AI has been trained to to detect clinically significant macular edema, allowing for earlier diagnosis and treatment of DR. Retinopathy of prematurity, or abnormal blood vessel growth of the retina due to premature birth, has also been influenced by AI advancements. Computerized algorithms and machine learning systems have been developed to develop scores for quantifying ROP emergence and progression. Finally, OCT and fundus photography are commonly used technologies to detect AMD, a major cause of vision loss in the field of retina. Both detection technologies are prime targets for artificial intelligence. In fact, studies have demonstrated that diagnosis of AMD by AI-guided OCT images and fundus photographs are comparable to human graders, with some studies demonstrating outperformance on the part of AI. Quite recently, a deep learning model has been shown to distinguish between Age-related Macular Edema, Diabetic Macular Edema, Drusen and Choroidal Neovascularization with 99.8% and 100% accuracy.
Further applications of AI in ophthalmology include detecting ocular malignancies. Machine learning algorithms have been developed to detect both basal cell and squamous cell carcinomas and to help with preoperative margin marking and reconstructive planning. Additionally, studies show that AI has demonstrated success in cataract detection, grading, and intraocular lens (IOL) power calculation. An emerging technique in AI called Generative adversarial networks (GAN) have shown to translate between two different ocular imaging modalities with high precision. For example, recent works have shown that these architectures can be utilized for synthesizing fluorescein angiography images from fundus photographs. Moreover, these GAN architectures can also be utilized for accurate retinal microvessel segmentation, which is an important tool for surgical application. It is readily apparent that AI has permeated many subspecialities within ophthalmology, with the potential for incredible advancement in the near future.
Current Imaging Modalities in Neuro-Ophthalmology
Before discussing applications of AI in neuro-ophthalmology, it is important to review current imaging modalities already available in the field. The major basis of imaging in neuro-ophthalmology is formed by computed tomography (CT) and magnetic resonance imaging (MRI), each having their distinct advantages and disadvantages depending on what an ophthalmologist is attempting to detect. Additionally, CT angiography (CTA) and MR angiography (MRA) are vital for highlighting vascular abnormalities potentially contributing to ophthalmic pathology. Furthermore, an important diagnostic tool is visual field testing via Humphrey Visual Field (HVF) perimetry. Pinpointing sites of visual field defects allows neuro-ophthalmologists to identify where along the optic pathway lesions may be. Finally, as in other subspecialities, fundus photography is also commonly used in neuro-ophthalmology to detect optic nerve changes.
Applications of AI in Neuro-Ophthalmology
Although there is marked overlap between glaucoma, retina, and neuro-ophthalmology, there are certain conditions that are of particular importance to neuro-ophthalmology - these being papilledema, anterior ischemic optic neuropathy (AION), and non-arteritic anterior ischemic optic neuropathy (NAION), as well as distinguishing AION and NAION from glaucomatous optic neuropathy (GON).
In neuro-ophthalmology, the laterality of a pathological process is important in narrowing down differential diagnoses. For example, bilateral optic disc edema and unilateral optic disc edema are typically caused by different disease states. Therefore, the ability of AI to distinguish laterality between the right and left eye is important for beginning the task of making a diagnosis. Liu et al developed a DLS that achieved 98.78% accuracy in this respect, rivaling the much larger data set of Jang et al. This makes the interesting conclusion that DLS can be accurate even with small data sets.
In a landmark retrospective study in 2020 utilizing 14,341 ocular fundus photographs, Milea et al found that deep-learning systems using retinal cameras were able to differentiate normal optic discs from those with papilledema or other non-papilledema abnormalities. After being cross-referenced with four expert neuro-ophthalmologists, it was determined that the DLS had sensitivity of 96.4% and a specificity of 84.7% for detecting papilledema. A variety of retinal cameras were used in this study, including Topcon ©, Zeiss ©, and Cannon ©. This DLS was further tested against two expert neuro-ophthalmologists after evaluating 800 new fundus photographs. Classification was split into normal optic discs, papilledema, or other optic disc abnormalities. In this study, the DLS classified 678 of 800 (84.7%) photographs correctly, while expert 1 correctly classified 675 of 800 (84.4%) and expert 2 classified 641 of 800 (80.1%). These studies highlight the possibility of faster and more accurate recognition of papilledema so treatment can be initiated right away.
AION and NAION are two vision-threatening conditions where quick diagnosis and intervention is vital. Levin et al demonstrated in their retrospective study that their neural network could detect AION 94.7% of the time compared to clinicians, with the capability of distinguishing AION from optic neuritis in the presence of overlapping features. However, a careful literature review shows minimal additional research regarding AI and AION detection, lending itself to potential future direction. Similarly, studies covering AI and NAION recognition are also limited. A 2006 study from Feldon et al described the ability of a computerized classification system to characterize NAION severity based on Humphrey Visual Field testing. However, the study lacked clinical applicability and was meant primarily for research purposes.
Additional research into deep learning systems in relation to HVFs has been conducted, albeit regarding glaucomatous changes. Utilizing 32,443 HVFs, Wen et al demonstrated their algorithm was able to create visual field predictions of glaucoma progression using a single initial HVF. The accuracy of the DLS ranged from 0.5 to 5 years, giving clinicians the predictive tools necessary to create more accurate treatment plans.
Glaucoma as a disease is often in the cross-section between glaucoma, the specialty, and neuro-ophthalmology. Because of this, it is important to differentiate glaucomatous optic neuropathy from non-glaucomatous optic neuropathy (NGON), such as AION or NAION, as it may determine which specialist sees the patient. Yang et al applied the convolution neural network of ResNet-50, a MATLAB® DLS, to 3815 color fundus images, displaying 93.4% sensitivity and 81.8% specificity in distinguishing NGON from GON. This has potential to provide clear demarcation between the two pathologies to guide referral, allowing for better use of resources and time.
One of the most limiting factors to implementing AI is cost. There are few studies in the literature that focus or even consider the effect of cost on AI practicality. There are even fewer that delve into specific conditions, with the most literature being on implementation with diabetic retinopathy screening. With this being said, Ruamviboonsuk et al found five studies published that reviewed the cost-effectiveness of AI. They concluded that artificial intelligence was generally more cost-effective than manually screening for diabetic retinopathy. However, these studies lacked generalizability. With this scarcity of literature, it is difficult to determine the overall impact AI is capable of having.
This lends itself to another problem facing AI: obtaining devices for study creation. With the time and finances it takes to create AI technology, it is particularly cumbersome to get FDA approval in order to start testing research hypotheses. FDA approval is especially important when it comes to insurance coverage - without coverage, the costs are simply too great to justify the use of AI. (29)
Another limitation to AI is finding large, complete datasets to build and fine-tune algorithms. The “garbage in, garbage out” phenomenon states that when incomplete and inadequate data is given to the AI device, incomplete and inadequate predictions will result. In order to find these datasets, it takes either a large private practice or hospital setting with willing patients as participants. This is likely a less detrimental limitation than cost proves to be.
Hopefully, as AI expands within the medical field and beyond, more literature will become readily available to demonstrate the cost-effectiveness of AI which will lead to easier access to this technology, reducing costs and allowing a great number of people to be reached.
Most research to date has focused on fundus photography for detecting neuro-ophthalmological conditions. As discussed above, however, there exists a wide range of imaging modalities in neuro-ophthalmology. In order to have a greater impact in the field, AI needs to permeate these imaging methods. There have been such studies in other fields, including cardiology, pulmonology, and neurology. Utilizing these frameworks, an explosion of neuro-ophthalmology research is possible.
Additionally, literature searches reveal that most studies pertaining to artificial intelligence and neuro-ophthalmology are retrospective studies. The application of AI in real-time allows for more clinical applicability. Optomed Oy (Ltd), a Finnish medical device manufacturer, is attempting to do this with their Optomed Aurora hand-held fundus camera. Preliminary studies have demonstrated high sensitivity and specificity in detecting optic disc abnormalities in real-time. Furthermore, the portable nature of a hand-held camera is particularly appealing in venues such as the emergency department for quick and accurate screening of retinal pathologies. Further research into the application of real-time imaging and mobile devices will undoubtedly advance the field of neuro-ophthalmology in the near future.
- ↑ Amisha, Malik P, Pathania M, Rathaur VK. Overview of artificial intelligence in medicine. J Family Med Prim Care. 2019;8(7):[]. doi:10.4103/jfmpc.jfmpc_440_19.
- ↑ Global CSU. How does artificial intelligence work?: CSU Global. The Official Blog of CSU Global. https://csuglobal.edu/blog/how-does-artificial-intelligence-actually-work. Published August 16, 2021. Accessed February 24, 2022.
- ↑ Deep Learning vs. machine learning: Beginner's guide. Coursera. https://www.coursera.org/articles/ai-vs-deep-learning-vs-machine-learning-beginners-guide. Published October 20, 2021. Accessed February 23, 2022.
- ↑ Thompson AC, Jammal AA, Medeiros FA. A Review of Deep Learning for Screening, Diagnosis, and Detection of Glaucoma Progression. Transl Vis Sci Technol. 2020;9(2):42. Published 2020 Jul 22. doi:10.1167/tvst.9.2.42.
- ↑ 5.0 5.1 Akkara JD, Kuriakose A. Artificial intelligence in glaucoma. EyeWiki. https://eyewiki.org/Artificial_Intelligence_in_Glaucoma. Published February 2, 2022. Accessed February 24, 2022.
- ↑ Al-Aswad LA, Kapoor R, Chu CK, et al. Evaluation of a Deep Learning System For Identifying Glaucomatous Optic Neuropathy Based on Color Fundus Photographs. J Glaucoma. 2019;28(12):1029-1034. doi:10.1097/IJG.0000000000001319.
- ↑ 7.0 7.1 7.2 Cole ED, Al-Khaled T, Valikodath N, et al. Artificial Intelligence in retina. EyeWiki. https://eyewiki.org/Artificial_Intelligence_in_Retina. Published December 18, 2021. Accessed February 24, 2022.
- ↑ Kamran, S. A., Saha, S., Sabbir, A. S., & Tavakkoli, A. (2019, December). Optic-net: A novel convolutional neural network for diagnosis of retinal diseases from optical tomography images. In 2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA) (pp. 964-971). IEEE
- ↑ Kamran, S. A., Tavakkoli, A., & Zuckerbrod, S. L. (2020, October). Improving robustness using joint attention network for detecting retinal degeneration from optical coherence tomography images. In 2020 IEEE International Conference On Image Processing (ICIP) (pp. []). IEEE.
- ↑ 10.0 10.1 Valikodath N, Cole E, Al-Khaled T, et al. Introduction to artificial intelligence in ophthalmology. EyeWiki. https://eyewiki.org/Introduction_to_Artificial_Intelligence_in_Ophthalmology. Published November 27, 2019. Accessed February 25, 2022.
- ↑ Goh JHL, Lim ZW, Fang X, et al. Artificial Intelligence for Cataract Detection and Management. Asia Pac J Ophthalmol (Phila). 2020;9(2):88-95. doi:10.1097/01.APO.0000656988.16221.04
- ↑ Kamran, S. A., Fariha Hossain, K., Tavakkoli, A., Zuckerbrod, S., Baker, S. A., & Sanders, K. M. (2020, October). Fundus2Angio: a conditional GAN architecture for generating fluorescein angiography images from retinal fundus photography. In International Symposium on Visual Computing (pp. 125-138). Springer, Cham.
- ↑ Kamran, S. A., Hossain, K. F., Tavakkoli, A., & Zuckerbrod, S. L. (2021, January). Attention2angiogan: Synthesizing fluorescein angiography from retinal fundus images using generative adversarial networks. In 2020 25th International Conference on Pattern Recognition (ICPR) (pp. []). IEEE.
- ↑ Tavakkoli, A., Kamran, S. A., Hossain, K. F., & Zuckerbrod, S. L. (2020). A novel deep learning conditional generative adversarial network for producing angiography images from retinal fundus photographs. Scientific Reports, 10(1), 1-15.
- ↑ Kamran, S. A., Hossain, K. F., Tavakkoli, A., Zuckerbrod, S. L., & Baker, S. A. (2021). Vtgan: Semi-supervised retinal image synthesis and disease prediction using vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. []).
- ↑ Kamran, S. A., Hossain, K. F., Tavakkoli, A., Zuckerbrod, S. L., Sanders, K. M., & Baker, S. A. (2021, September). RV-GAN: segmenting retinal vascular structure in fundus photographs using a novel multi-scale generative adversarial network. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 34-44). Springer, Cham.
- ↑ Kakaria AK. Imaging in neuro-ophthalmology: An overview. Oman J Ophthalmol. 2009;2(2):57-61. doi:10.4103/0974-620X.53033
- ↑ Kimura T, Matsumoto C, Nomoto H. Comparison of head-mounted perimeter (imo®) and Humphrey Field Analyzer. Clin Ophthalmol. 2019;13:501-513. Published 2019 Mar 14. doi:10.2147/OPTH.S190995
- ↑ Neuro-ophthalmology Illustrated-2nd Edition. Biousse V and Newman NJ. 2012. Thieme
- ↑ 20.0 20.1 20.2 Liu TYA, Ting DSW, Yi PH, et al. Deep Learning and Transfer Learning for Optic Disc Laterality Detection: Implications for Machine Learning in Neuro-Ophthalmology. J Neuroophthalmol. 2020;40(2):178-184. doi:10.1097/WNO.0000000000000827
- ↑ 21.0 21.1 Jang Y, Son J, Park KH, Park SJ, Jung KH. Laterality Classification of Fundus Images Using Interpretable Deep Neural Network. J Digit Imaging. 2018;31(6):923-928. doi:10.1007/s10278-018-0099-2
- ↑ 22.0 22.1 Milea D, Najjar RP, Zhubo J, et al. Artificial Intelligence to Detect Papilledema from Ocular Fundus Photographs. N Engl J Med. 2020;382(18):1687-1695. doi:10.1056/NEJMoa1917130
- ↑ Biousse V, Newman NJ, Najjar RP, et al. Optic Disc Classification by Deep Learning versus Expert Neuro-Ophthalmologists. Ann Neurol. 2020;88(4):785-795. doi:10.1002/ana.25839
- ↑ 24.0 24.1 Levin LA, Rizzo JF 3rd, Lessell S. Neural network differentiation of optic neuritis and anterior ischaemic optic neuropathy. Br J Ophthalmol. 1996;80(9):835-839. doi:10.1136/bjo.80.9.835
- ↑ Feldon SE, Levin L, Scherer RW, et al. Development and validation of a computerized expert system for evaluation of automated visual fields from the Ischemic Optic Neuropathy Decompression Trial. BMC Ophthalmol. 2006;6:34. Published 2006 Nov 20. doi:10.1186/1471-2415-6-34
- ↑ 26.0 26.1 Wen JC, Lee CS, Keane PA, et al. Forecasting future Humphrey Visual Fields using deep learning. PLoS One. 2019;14(4):e0214875. Published 2019 Apr 5. doi:10.1371/journal.pone.0214875
- ↑ Yang HK, Kim YJ, Sung JY, Kim DH, Kim KG, Hwang JM. Efficacy for Differentiating Nonglaucomatous Versus Glaucomatous Optic Neuropathy Using Deep Learning Systems. Am J Ophthalmol. 2020;216:140-146. doi:10.1016/j.ajo.2020.03.035
- ↑ Ruamviboonsuk P, Chantra S, Seresirikachorn K, Ruamviboonsuk V, Sangroongruangsri S. Economic Evaluations of Artificial Intelligence in Ophthalmology. Asia Pac J Ophthalmol (Phila). 2021;10(3):307-316. Published 2021 Jul 13. doi:10.1097/APO.0000000000000403
- ↑ Chen MM, Golding LP, Nicola GN. Who Will Pay for AI?. Radiol Artif Intell. 2021;3(3):e210030. Published 2021 Mar 3. doi:10.1148/ryai.2021210030
- ↑ Weng SF, Reps J, Kai J, Garibaldi JM, Qureshi N. Can machine-learning improve cardiovascular risk prediction using routine clinical data?. PLoS One. 2017;12(4):e0174944. Published 2017 Apr 4. doi:10.1371/journal.pone.0174944
- ↑ Schoepf UJ, Schneider AC, Das M, Wood SA, Cheema JI, Costello P. Pulmonary embolism: computer-aided detection at multidetector row spiral computed tomography. J Thorac Imaging. 2007;22(4):319-323. doi:10.1097/RTI.0b013e31815842a9
- ↑ Ambastha AK, Leong TY; Alzheimer's Disease Neuroimaging Initiative. A Deep Learning Approach to Neuroanatomical Characterisation of Alzheimer's Disease. Stud Health Technol Inform. 2017;245:1249.
- ↑ Upadhyaya S, Agarwal A, Rengaraj V, Srinivasan K, Newman Casey PA, Schehlein E. Validation of a portable, non-mydriatic fundus camera compared to gold standard dilated fundus examination using slit lamp biomicroscopy for assessing the optic disc for glaucoma. Eye (Lond). 2022;36(2):441-447. doi:10.1038/s41433-021-01485-2
- ↑ Kubin AM, Wirkkala J, Keskitalo A, Ohtonen P, Hautala N. Handheld fundus camera performance, image quality and outcomes of diabetic retinopathy grading in a pilot screening study. Acta Ophthalmol. 2021;99(8):e1415-e1420. doi:10.1111/aos.14850