The “black box” of artificial intelligence in ophthalmology

Evolution in AI is fast, promising and could revolutionise ophthalmological decision-making. Image credit: ©alastis – stock.adobe.com

Don’t be mistaken: artificial intelligence (AI) has been around, and used in medicine, for quite a long time. If you make a query in Pubmed, you will find AI mentioned in papers dating back to 1953.

AI refers to software programmes which mimic human cognitive functions like reasoning, conclusion and decision-making. In recent years, AI has become more and more mature. Superior computer chips paved the way to superior machine learning and deep learning. It has now the true potential to have a great impact on our society altogether, and medicine in particular.

In looking closely at AI we discover a gradation in “intelligence” for these programmes. We can have the simplest programmes like automated detectors, where patterns in data yield an outcome ( diagnosis). These are useful in automated blood screening programmes. The AI software does that by processing large amounts of data and recognising patterns in this vast amount of data. Patterns may not even be clear or explainable to the investigators.

One of the most-heard concerns is that AI behaves as a “black box”. This has been addressed by Juan Durán and Karin Jongsma in the Journal of Medical Ethics, "Who is afraid of black box algorithms?" in which they discussed issues such as potential bias, accountability, responsibility, patient autonomy and compromised trust with so-called black box algorithms.

More complex and advanced is machine learning AI where the machine is trained to navigate a dataset and detect by itself the important features that guide towards a diagnosis. Numerous examples have shown clinicians that good predictability with the training set does not automatically mean good predictability in real-life situations, so these types of AI still need thorough supervision by a human expert. Finally, there are deep learning neural networks that try to closely mimic the very complex decision-making processes of which our human brain is capable.

We can say that current AI programmes are just arriving in the second stage. But evolution in AI is very fast and promising, and could even now revolutionise ophthalmological decision-making. The eye is a particularly good organ in which to implement AI-enabled care. Most of our decision-making is based on images and mathematical formulas, like the Fourier or Zernike transformations in topographic cornea imaging. Images and mathematics are particularly suitable for AI software.We have currently nearly all anatomic structures of the eyewhere AI could be or is at the verge of being implemented, both in therapeutics and diagnostic purposes.

AI has applications in retinal diseases like diabetic retinopathy, retinopathy of prematurity, age-related retinal disease and hereditary retinal diseases. In glaucoma, AI can predict risk of angle closure and progression of disease. AI can also serve in cataract care, ocular oncology and tele-ophthalmology.

Refractive surgery and corneal laser procedures have not lagged behind. The introduction of the InnovEyes algorithm in the WaveLight Excimer Laser System (Alcon) is an interesting tool in cornea refractive procedures. Since the introduction of this AI tool, we can further enhance our clinical results with LASIK corneal refractive procedures. We are now confident that the test group of corneas used by Alcon and OCULUS to develop the AI tool closely match the corneas of the patients in our practise. It also helps that we can see in the planning station, prior to the surgery, what the AI tool is planning to correct.

AI can be a tool of great value in diagnosing and treating complex eye problems. Let’s embrace AI and work closely with it, but remain critical about the AI-solutions brought by industry partners. Only we can measure if the solutions brought by AI tools can be worthwhile for our patients. There is no more reason to be afraid of the black box.