Bringing eye images into focus with AI

Researchers at the University of Waterloo have developed a better way to enhance the clarity and detail of eye images used to diagnose disease by teaching artificial intelligence (AI) software the science behind the imaging process.

The new AI model precisely reverses quality loss and reconstructs reliable images, providing a powerful tool for more accurate diagnosis of eye diseases affecting the cornea, the transparent tissue in the front of the eye.

To detect evidence of ocular disease and monitor treatment, doctors rely on scans of microscopic tissues captured using a technique that bounces light off tiny structures within the eye. As the light reflects, it creates a blurring effect and a grainy pattern called "speckle noise" that obscures fine details and makes image analysis difficult.

"The trade-off with cellular-level images is that they can appear out of focus and noisy," said Dr. Kostadinka Bizheva, a professor of physics and astronomy who supervised the study until suddenly passing away recently.

"It's like trying to read something through frosted glass. Restoring the image quality is an essential step to ensuring an accurate diagnosis."

Before (top) and after (bottom) closeup images of a cucumber (left) and a human cornea (right) show how their clarity and detail are improved by new AI software. (University of Waterloo)

The Waterloo-developed solution reverses defocus and suppresses speckle noise using a physics-informed diffusion model (PIDM). Researchers trained the model on the physics of how light moves and interacts with tissue at a cellular level so that it understands how defocus and speckle noise are formed.

The AI model then accounts for those issues when it progressively refines the image, verifying each step against real-world physics to ensure scientific accuracy.

"Typical diffusion AI models can sometimes misinterpret or 'hallucinate' details when the image is reconstructed," said Dr. Alexander Wong, professor of systems design engineering and the Canada Research Chair in Medical Imaging Systems.

"By merging the power of AI with the knowledge of physics, our model can methodically reduce such errors and produce more trustworthy results."

In tests on images of plant tissue and the human cornea taken with optical coherence tomography (OCT) - a non-invasive scan similar to ultrasound but uses light instead of sound waves - the PIDM outperformed current reconstruction methods to reveal crisp cell outlines and details of internal structures.

Wong said the results demonstrate how embedding scientific principles in AI models can create more trustworthy and effective tools to improve human health.

Dr. Lyndon Jones, an optometry and vision science professor who was not involved in the study, said the AI model could help doctors diagnose diseases of the external eye much earlier and catch problems that might have been missed without it.

"This technology comes at a time when OCT imaging of the eye is becoming more common and will be crucial to its widespread adoption by eyecare practitioners worldwide," said Jones, Principal Scientist at the Centre for Ocular Research and Education at Waterloo.

Dr. Bizheva's collaborators now hope to build on the work she began by incorporating additional physics principles in the AI model and extending its application to other eye tissues, such as the retina, to support the diagnosis of more diseases.

The research team also included Nima Abbasi, a PhD candidate in systems design engineering at Waterloo.

A paper on this work, A Physics-Informed Diffusion Model for Super Resolved Reconstruction of Optical Coherence Tomography Data, appeared in IEEE Transactions on Biomedical Engineering.

Feature Image: Nima Abbasi (left) and Dr. Alexander Wong helped develop a new AI model that enhances the clarity of medical images used to diagnose eye diseases. (University of Waterloo)