Deep Learning OCT

Timothy Norris reports

Imaging advancements in OCT have reached extremely high standards in terms of quality and resolution, according to Pierre Zéboulon MD during the “OR Digitalisation: Truth and Myths” clinical research symposium at the ESCRS Congress, adding cutting-edge OCT potentiality is still under-exploited in clinical practice.

“We now have access to beautiful anterior segment OCT images with high levels of detail of the cornea and the lens,” he said. “We need to build new tools to take advantage of this advanced imaging.”

Using deep learning with OCT scans could be a very important step for ophthalmology, helping design new techniques and approaches to different diagnoses and bringing valuable insight to ocular pathology.

“With this, we will create new standards in the future, entering a new era of anterior segment diagnosis,” Dr Zéboulon said.

“Why do we need objective cataract grading? It is obviously better for the patient’s follow-up. It helps in complicated cases of visual activity loss. It will certainly be useful for the near future of telemedicine.”

Dr Zéboulon used Heidelberg’s Anterion for high-quality OCT scans of the anterior segment to show his cases. “We wanted to build something entirely new—that takes full advantage of the high-resolution images, a new approach not based on a calculated subjective clinical classification. We wanted it to be as objective as possible, with a clinically relevant output,” he said. “So, we went for individual continuous metrics for each lens layer with good interpretability.”

The cornerstone, he noted, is a deep learning model detecting pixel-wise cataracts on the OCT images, correctly and directly detecting any cataract type with great accuracy, as demonstrated in a study published by the author in the Journal of Optometry in 2022.¹

“And so, we built our grading solution on top of that model to access layer-wise information,” he said.

Global Cataract Score was created using linear discriminant analysis and the six variables on a development set of 331 eyes and tested on 217 eyes. “The output of this score is very simple,” Dr Zéboulon explained. “It’s a single value: If it’s positive, there’s clinically significant cataract.”

Deep learning-empowered OCT can also be used for the early detection of corneal oedema, a very useful tool given the constant growth of DMEK procedures worldwide. “Our solution is capable of detecting an oedema corresponding to a change of 20 μm with an AUC of 0.96, also providing an easy to read and en-face map,” Dr Zéboulon said.

Some limitations are still present in the model, especially in the en-face oedema map. “With a flat posterior surface, it can be a bit tricky to know if there’s a posterior displacement, and our model doesn’t fit the image. Also, if the posterior surface is prolate, such as in the case of keratoconus, it cancels out every feature of the presentation itself. The third limitation is in case of homogeneous thickness increase,” Dr Zéboulon explained.

Dr Zéboulon presented at the 2023 ESCRS Congress in Vienna.

Pierre Zéboulon MD, MSc is a corneal surgeon in the Anterior and Refractive Surgery Department of the Rothschild Foundation Hospital, Paris.

pierrezeboulon@gmail.com

1.       José M González-Méijome et al., J Optom., 2022; 15: 1–2.