Useful Answers from AI Start with Asking the Right Questions

AI in ophthalmology is much like AI in any other field: It provokes excitement in some and dread in others. However, its applications are clear, according to Béatrice Cochener-Lamard MD, PhD, and its use should be well established.

“Do you know that we have been talking about AI since 1940? But even in the early 2000s, nobody was even thinking that AI was already in our life. Machine learning was there, but nobody knew that it was AI,” she said.

“Then there is the deep learning we are all using today—neural networks with many layers. Deep learning is no more than an algorithm that will answer questions. The key is asking the proper question. That is your responsibility.”

Being able to ask the proper question requires clinicians to provide deep learning models with a significant quantity of data to create an algorithm comprising millions or billions of parameters to be calculated. Data hygiene is a must, according to Professor Cochener-Lamard, as the information fed to deep learning models must be relevant, accurate, and consistent.

“AI can offer binary, multi-label, and multi-class classifications as responses. You can decide, for example, to ask your AI to say an image is normal or not, or if there are any kind of lesions, but you need to segment your inquiries and make them specific.”

Train your technology

Prof Cochener-Lamard defined three ways to train deep learning models: supervised, unsupervised, and self-supervised. Supervised is the most common. It is frequently used in medical imaging diagnostics and depends on the accuracy of manually labelled image sets to carry out automated disease classification and lesion segmentation.

The unsupervised and self-supervised approaches instead process vast amounts of raw, unlabelled data and learn autonomously by predicting gaps in the data and extrapolating answers. This is generative AI, the model used by programs like ChatGPT, and its autonomy can cause problems.

“You need to understand what type of AI is engaged, the quantity and quality of data, the population on which the data was trained, etc.,” Prof Cochener-Lamard said.

Despite these challenges, Prof Cochener-Lamard said that AI can be trusted when properly trained. Continuing her example of lesion analysis, she said deep learning models can highlight specific pixels on an anterior segment image or video frame to show exactly where a lesion or structural anomaly was flagged.

“When AI starts to work on a big volume of data, you can go from ‘do I have diabetic retinopathy or not?’ to ‘yes, and I can tell you that I have identified some other lesions,’” she said. “Explainability is key—you have to give AI the right data to find the right answers.”

Taking over triage and testing

When mastered, AI’s clinical applications are significant, particularly in the anterior segment. Prof Cochener-Lamard attributed this significance to high patient volumes, standardised diagnostic imaging, and a clinical dependency on ‘structural micrometre accuracy,’ making ophthalmology an ideal environment for deep learning.

“AI in the anterior segment is transitioning gently from diagnostic support to predictive analytics to intraoperative guidance. It is also moving to semi-autonomous microsurgery and will one day move to robotics too,” Prof Cochener-Lamard said.

Not only can AI outperform traditional screening for many anterior segment conditions, but Prof Cochener-Lamard said it can also significantly improve corneal treatment. It can create digital twin corneas, allowing clinicians to map biomechanical responses, customise laser ablation profiles, and map preoperative wound-healing dynamics.

“Keratoconus and ectasia prediction is the most mature AI application in the cornea. AI is transforming keratoconus screening from a static threshold to multidimensional risk prediction based on tomography, epithelial mapping, and biomechanical analysis,” Prof Cochener-Lamard said.

“Infectious keratitis is also a clinically impactful example of AI because in certain parts of the world, where clinicians have low resources, AI can be used to triage patients. You can even use a smartphone to help you to differentiate infection agents.”

AI is here, ready or not

Yet AI is still dependent on national and international legislation, Prof Cochener-Lamard said. The EU’s European Directorate for the Quality of Medicines & HealthCare (EDQM) provides significant data privacy protection for patients when AI is used, but that means its development could also be stymied. Other regions, such as China, are reportedly far ahead.

Whatever the regulatory requirements, however, Prof Cochener-Lamard emphasised that AI is here to stay; it is useful and effective, and those doctors who do not use it will get left behind.

“AI is a tool that has real added value for doctors,” Prof Cochener-Lamard concluded. “Remember, it is not AI that will replace you, but rather the person who knows how to use it.”

Prof Cochener-Lamard presented at the EuCornea annual congress in Porto, Portugal.

Béatrice Cochener-Lamard MD, PhD, FEBO is Head of the Ophthalmology Department at Brest University Hospital, France, a recipient of the Legion of Honour, a former president of the ESCRS, and past president of EuCornea. beatrice.cochener@ophtalmologie-chu29.fr