AI Shows Promise for Meibography Grading

A multimodal large language model (LLM) showed promising results in analysing and interpreting meibography images and detecting morphological abnormalities and subtle changes; further studies are expected to prove it as an asset for the diagnosis process, according to Pelin Kiyat MD.

“The medical community has witnessed a significant progression of the artificial intelligence role in the field,” she said. “Large language models have the potential to gain importance due to their capability to process and interpret both visual and textured medical data.”

To demonstrate, Dr Kiyat presented on a study involving a new LLM called Claude Sonnet 3.5 (Anthropic), which contains newly designed features for analysing multimodal data. The purpose of the study was to evaluate the performance of Claude Sonnet 3.5 in interpreting quantitative and qualitative alterations in meibomian glands through meibography images, of which the study included 228. An experienced ophthalmologist evaluated and graded the morphological changes and the drop-out ratio of meibomian glands using upper and lower eyelid meibography captured with an infrared microscope. The images were graded from zero to three according to a previously defined system. A subset of 160 images was selected to ensure 40 images each of the four grades.¹ The study asked Claude Sonnet 3.5 to analyse the meibography images, provide the graded gland drop-out, and describe the morphological abnormalities.

Standard sentence prompts were used to obtain consistent LLM responses, then compared to the ophthalmologist’s manual evaluations and classified as correct or incorrect. Researchers assessed performance by calculating the percentage of correct answers and the LLM’s ability to identify morphological changes. For quality assurance, a new account was created for the study, the conversation history was cleared between the images, and the same ophthalmologist performed all image inputs.

The study results showed human evaluation identified morphological abnormalities in 48 images (30%), with heterogeneous lumen diameters (33 eyes) the most prevalent—including both dilated and narrowed lumens, glandular tortuosity (17 eyes), lumen length shortening (16 eyes), and hyperreflective gland residues (11 eyes). Claude Sonnet 3.5 showed high accuracy in grading meibomian gland drop-out, with only 12 images incorrectly analysed. It achieved 100% sensitivity in detecting morphological abnormalities by correctly identifying all 48 images. In addition, Claude Sonnet 3.5 found that 12 images initially defined as normal in the manual evaluation had subtle morphological changes, such as mild tortuosity, inconsistent spacing, and minor gland fusion.

Furthermore, these results demonstrate Claude Sonnet 3.5 as an effective tool to support early diagnosis, presenting an opportunity to speed up patients’ flow through the clinic system. However, further studies are required to address the incorrect grading or the tendency to underestimate severe cases.

“This technology could particularly benefit high-volume ophthalmology clinics by potentially reducing diagnostic time while maintaining accuracy,” Dr Kiyat concluded.

Dr Kiyat spoke at the 2025 ESCRS Winter Meeting in Athens.

Pelin Kiyat MD is an ophthalmologist at the Department of Ophthalmology, Izmir Democracy University, Buca Seyfi Demirsoy Training and Research Hospital, Izmir, Turkey. pelinkiyat@hotmail.com

1. Arita R, Suehiro J, Haraguchi T, Shirakawa R, Tokoro H, Amano S. “Objective image analysis of the meibomian gland area,” Br J Ophthalmol, 2014; 98(6): 746–755. doi:https://doi.org/10.1136/bjophthalmol-2012-303014