AI Analysis and the Cornea

AI-driven widefield specular microscopy promises an unprecedented improvement in imaging for Fuchs’ corneal endothelial dystrophy (FCED), recent studies from Singapore suggest.

Specular microscopy, the current standard practice for imaging FCED and the cornea in general, is too inaccurate and can be influenced by too many factors, according to Marcus Ang MBBS, PhD.

“Specular microscopy is the current standard, but this involves a central, less than 0.05 mm² endothelial scan, which covers less than 1% of the cornea. This can also be affected by corneal clarity, optical quality, artefacts, and patient cooperation,” he said.

“Widefield specular microscopy imaging could therefore provide real benefit, as it captures multiple images of the central, paracentral, and peripheral corneal epithelium. It can provide a map of the cornea within seconds, and the operator can then select the best quality images multiple times—from the central to the peripheral regions.”

Professor Ang discussed several studies based at the Singapore National Eye Centre, the first of which confirmed the quality of images taken by the widefield approach. One hundred FCED patients had 15 images taken of each eye, divided accordingly: 1 central, 8 paracentral, and 6 peripheral. Image quality was graded as follows: (Grade 1) less than 25%, (Grade 2) 25–50%, (Grade 3) 50–75%, and (Grade 4) greater than 75%.

These images were then analysed and graded by a deep learning AI algorithm called DenseNet-121. They were also assessed by an independent clinician using a Bland–Altman plot and Kappa coefficient to act as a control.

Deep learning, wide imaging

Deep learning is a form of AI that creates artificial neural networks with multiple layers that are designed to mimic the structure and function of the human brain. It can analyse large volumes of high-resolution digital images, generated from various ophthalmic modalities, to create large data sets on a particular target.

Prof Ang reported DenseNet-121 was able to identify and grade each image correctly, and the accuracy level of both methods was identical. In total, 72.9% of the images were classified Grade 3 or above, and the image quality of the paracentral and peripheral regions was better in eyes with higher endothelial cell density, which he said was “a marker of FCED severity.”

Additional findings demonstrated that endothelial cell density correlated well with corneal oedema, meaning the images can also be used to ascertain cell loss in the paracentral and peripheral regions. As such, he recommended using widefield imaging combined with AI-driven analysis as an alternative to specular microscopy.

In a second study, Prof Ang hypothesised that AI-driven widefield imaging could also help guide treatments such as Descemet stripping only (DSO). As the technique is primarily used in eyes with good peripheral cell counts, he believed he could use AI-driven widefield imaging to identify which eyes would make good candidates for DSO.

Prof Ang trained the same DenseNet-121 model to differentiate specular images of FCED and those with other types of endothelial damage, such as uveitis, from normal eyes. An independent data set of 752 paracentral and 557 peripheral images of patients with mild to moderate FCED was examined.

“Our accuracies were pretty good, ranging from 0.77 to 0.99, and the AI could also identify eyes with cell densities of more than 1,000,” Prof Ang said. “This could be developed into software or an app designed for the treatment of DSO.”

Lastly, Prof Ang reported on a prospective study examining the role AI can play for FCED patients undergoing femtosecond laser-assisted cataract surgery (FLACS). He said he hopes to use AI to mitigate the impact of cataract surgery on the cornea’s endothelium, which can have differential effects on the central, paracentral, and peripheral areas.

Using deep learning again, Prof Ang plans to investigate how an AI detection tool could help improve cell shape detection and create regional maps of cell density in the cornea. Such maps could also include data on hexagonality and coefficient variation viewed both pre- and postoperatively to ascertain the impact of FLACS on corneas.

Prof Ang spoke at the EuCornea 2025 congress in Prague.

Marcus Ang MBBS, MMed(Ophth), MCI, FRCS(Ed), FAMS, PhD is a Senior Consultant at the Singapore Eye Centre and an associate professor at the Duke National University of Singapore Medical School. gmsahnm@nus.edu.sg