Artificial intelligence will change the way eye care providers practice
The concept of artificial intelligence is not new, but its exponential growth over the past 10 years promises to make it an immediate and long-term challenge and opportunity in health care, including optometry and ophthalmology.
The artificial intelligence (AI) concept is complex, yet understandable. Using computer software algorithms, programs that simulate the progressive layers of neuronal functions in the human cortex (neural networks) can learn (deep learning) to recognize, identify and interpret digital patterns, such as images, sound, speech, text, data – almost anything. Then, computer applications using Bayesian probability logic, which is a statistical analysis of increasing available information to confirm the probability of a hypothesis, can be applied to this pattern recognition to interpret reasonable expectations.
Acceptance in eye care
Is the concept feasible in eye and vision care? Researchers tested its performance for diabetic retinopathy and found that deep learning AI matched or exceeded the performance of experts in identifying and grading the severity of the conditions (Gulshan et al.). Even more profound, the software was not explicitly programmed to recognize features from images that might indicate the disease. It simply looked at thousands of healthy and diseased eyes and figured out for itself how to spot the condition.
One of the biggest challenges for clinicians will be convincing them that the computers’ diagnostic conclusions are reliable. Certainly, IBM’s Watson (www.ibm.com/watson) has already made a strong case for AI and, to a lesser, but no less impressive degree, so have chatbots, Amazon’s Alexa, Apple’s Siri and Google’s Home. But AI machine-learning systems capable of explaining how they reach their conclusions will also be needed to reassure practitioners. Google is already working on that (Knight).
The description above of diabetic retinopathy recognition through deep learning is an example of supervised deep learning. The computer program was provided with labeled data, which the machine then analyzed and interpreted, layer by layer to arrive at conclusions.
In unsupervised deep learning, information is extracted from unlabeled data (Bayesian deductions, experiences, observations, memory) and then employed in an active process to improve performance with labeled data.
The highest level of computational deep learning is reinforcement learning. Here, the software tries to maximize the greatest benefit(s) it can receive when interacting with a complex, uncertain environment, according to Sutton and colleagues. It uses a process called memory networking, which is a neural network that mimics the plasticity of the human brain’s hippocampus by using random access memory (RAM) to differentiate and adjust the connections between unsupervised and supervised learned information, according to Fan.
Human vision is an active process that sequentially samples the optic impulses in an intelligent, task-specific way using the high-resolution fovea with large, low-resolution surroundings. This physiological, unsupervised learning process in human vision will prove invaluable in future AI diagnostic and corrective approaches in vision care. Vision systems combining deep learning and reinforcement learning are in their early stages, but they already outperform passive, traditional vision systems at classification tasks, Mnih and colleagues reported.
Well suited for repetitive tasks
Finally, beyond cerebral activities including professional judgment, decision-making and personal (patient) interactions associated with day-to-day eye care practice, a significant, albeit critical, portion of the practice includes repetitive activities such as refraction, tonometry, biomicroscopy, ophthalmoscopy and other data-gathering functions. Some of these functions are amenable to deep learning, and some can be effectively accommodated through the merging of AI and robotics.
Robotics is the engineering science that provides mechanical, electrical, computer-driven AI and neural network machines to substitute for and perform repetitive human tasks. Medical robots have the potential to augment the capabilities of doctors and enable new medical diagnostic and treatment procedures (eg, da Vinci Surgical System) with fewer negative side effects. Combining robotics with unsupervised, supervised and eventual reinforcement AI deep learning for tasks such as refraction, biomicroscopy, ophthalmoscopy, aberrometry and other sophisticated, routine ophthalmic procedures is just around the corner.
Fortunately, most forward-thinking health care professionals, including eye care providers, are proactively beginning to adjust their practices and their thinking to the imminent and forthcoming changes AI is bringing. There seem to be three common denominators these enlightened practitioners have in common.
First, they are reinventing the way they think about and do things in their practice. Whether it currently does or does not relate to computer-dependent activities, they are considering how to adjust and approach given tasks, both mentally and physically, if and when those tasks can be converted to a computerized or robotic process. This takes some thinking outside the box in many areas heretofore of a traditional, routine nature.
The second common denominator becoming the hallmark of all innovative and advancing health care practices is the incorporation of diagnostic and data-gathering automation technologies. AI, neural networks and robotics are sciences that interact with technologies more than with humans. In fact, their goal is to substitute or replace the human element. Thus, those specific tasks that have been relegated to automating technologies are, and will continue to be, the areas most amenable to AI and robotic conversion.
Finally, the most successful health care practitioners are intensively learning and embracing the science and concepts AI represents. They are acknowledging its potential and its current and future role in their professions and careers. They are enthusiastically accepting the idea that the merger of technology and AI is the answer to those things, mental and physical, that represent redundant repetition, time-inefficiency, need for consistency, thoroughness (for a quality outcome) and cost effectiveness in their practice.
The goals of deep learning in AI and neural networks are to create computer platforms that can imitate and duplicate human activities, reasoning and logic and, then, through experience, unsupervised and reinforcement learning, to electronically build maximally efficient artificial intelligence programs. Is this concept uncomfortable? Are you seeing shades of “Hal” from Stanley Kubrick’s 2001?
The debate and questions regarding what AI and neural networks might do in the future will continue to grow. But these questions are themselves “proof of concept.” The biggest question for eye care and its practitioners should be: “What can it do and what can I do right now?” This is only the beginning.
- Alterovitz R, et al. AI Magazine. 2016;37(2):76-84;idm-lab.org/bib/abstracts/papers/aimag15.pdf.
- Ba J, et al. Multiple object recognition with visual attention. Presented at: International Conference on Learning Representations. 2014;arxiv.org/abs/1412.7755.
- Fan S. Google’s new AI gets smarter thanks to a working memory. Singularity Hub.com. November 1, 2016.
- Gulshan V, et al. JAMA. 2016;316(22):2402-2410;doi:10.1001/jama.2016.17216.
- Knight W. AI’s Language Problem. MIT Technology Review. www.technologyreview.com/s/602094/ais-language-problem. August 9, 2016.
- LeCun Y, et al. Nature. 2015;521:436-444;doi:10.1038/nature14539.
- Lewis-Kraus G. The Great AI Awakening. The New York Times Magazine. www.nytimes.com/2016/12/14/magazine/the-great-ai-awakening.html. December 14, 2016
- Liang P. Semi-supervised learning for natural language. DSpace@MIT. 2005;dspace.mit.edu/handle/1721.1/33296.
- Mnih V, et al. Nature. 2015;518:529-533;doi:10.1038/nature14236.
- Nilsson N. The Quest for Artificial Intelligence. Cambridge University Press. 2010. ai.stanford.edu/~nilsson/QAI/qai.pdf.
- Sutton RS, et al. Reinforcement Learning: An Introduction. The MIT Press. 2012. people.inf.elte.hu/lorincz/Files/RL_2006/SuttonBook.pdf.
- For more information:
- Louis J. Catania, OD, FAAO, and Ernst Nicolitz, MD, are in private practice at Nicolitz Eye Consultants in Jacksonville, Fla. Catania is also a Primary Care Optometry News Editorial Board member. Catania can be reached at email@example.com. Nicolitz can be reached at firstname.lastname@example.org.
Disclosures: Catania and Nicolitz reported no relevant financial disclosures.