Analyzing FLIO data from the AI-READI project with Aaron Y. Lee, MD, MSCI

Fluorescence lifetime imaging ophthalmoscopy offers a novel approach to retinal imaging, enabling AI-driven insights into systemic health through the AI-READI project. (Image credit: AdobeStock/Sam Efendi)

As researchers continue to explore the retina as a window into systemic health, fluorescence lifetime imaging ophthalmoscopy (FLIO) offers a novel and underexplored dimension, according to Aaron Y. Lee, MD, MSCI.

Lee is the C. Dan and Irene Hunter Endowed Professor at the University of Washington Department of Ophthalmology in Seattle. He was recently appointed head of the Department of Ophthalmology and Visual Sciences at Washington University School of Medicine in St Louis, Missouri, where he will also hold the title of Arthur W. Stickle Distinguished Professor in Ophthalmology and Visual Sciences. He starts his new role on September 1, 2025.¹

Lee’s presentation at the 2025 International SPECTRALIS Symposium – And Beyond (ISS), held in June in Heidelberg, Germany, highlighted how FLIO data is being integrated into the Artificial Intelligence Ready and Equitable Atlas for Diabetes Insights (AI-READI) project. Sheryl Stevenson, executive editor with the Eye Care Network, caught up with Lee to learn more.

Please explain the significance of FLIO in retinal imaging and how its data contribute to the goals of the AI-READI project.

FLIO is a relatively new experimental modality where we hope the fluorescence lifetime imaging will provide insights that may help with understanding how diabetes affects the macula. We have many well-established retinal imaging modalities in AI-READI: color fundus photography, infrared imaging, fundus autofluorescence, optical coherence tomography (OCT), and OCT angiography. These modalities have provided well-proven associations with systemic health markers. We hoped by systematically collecting a large cross-sectional dataset of FLIO data, we could provide a research resource to the larger community as well as unlock new associations from the retina to the rest of the body.²

What preprocessing steps are necessary before FLIO data can be effectively used in machine learning or AI models?

For machine learning, the current form of the data in AI-READI is easily analyzed. The data is stored as a 3D array in DICOM, where any DICOM reader can easily read the data. We encoded the raw photon counts per picosecond in the z-axis so that when you scroll through the volume, you are scrolling through the time series. The challenge of our dataset is that it does not work well with existing tools that have traditionally been used for curve fitting or phaser plots.

What challenges have you encountered when aligning FLIO data with other multimodal datasets in AI-READI, and how did you address them?

Aligning the FLIO data with the other modalities is fairly simple. We have provided direct linkage files for each participant and have shown how they can be used to pull consistent data from each participant. We have a series of metadata files for each modality that allow researchers to easily find the images that they may be interested in without crawling through all the files individually.

How do you ensure the interpretability and clinical relevance of the insights derived from AI models trained on FLIO data?

The current state of the art in FLIO analysis is relatively immature for understanding what is clinically relevant or interesting. One of the many challenges with FLIO is that the traditional approaches require manual tuning of parameters to achieve good fit, and it is hard to replicate and reproduce the results at scale.

Have you observed any novel patterns or biomarkers in the FLIO data that were revealed only through AI-driven analysis?

We presented some of our preliminary work at the ISS 2025 meeting. We showed that there was a dose-dependent relationship between some of the blood test markers and the FLIO signal from the central subfield of the macula in the AI-READI dataset. This is very preliminary but also very exciting to see because it could indicate that there are latent associations in the fluorescence lifetimes that are directly associated with systemic health markers.

References

1. Lee named new head of Ophthalmology and Visual Sciences. News release. Washington University School of Medicine. April 1, 2025. Accessed June 17, 2025. https://ophthalmology.wustl.edu/lee-named-new-head-of-ophthalmology-visual-sciences/

2. Owsley C, Matthies DS, McGwin G, et al; AI-READI Consortium. Cross-sectional design and protocol for Artificial Intelligence Ready and Equitable Atlas for Diabetes Insights (AI-READI). BMJ Open. 2025;15(2):e097449. doi:10.1136/bmjopen-2024-097449