NEI applies artificial intelligence to OCT for retinal imaging

16 Apr 2024

Combination of AI and adaptive optics speeds up image generation and improves contrast.

Researchers at the National Eye Institute (NEI), part of the US National Institutes of Health, have applied a particular artificial intelligence approach to OCT imaging of the retina.

The research, published in Nature Communications Medicine, builds on work at NEI to enhance OCT via adaptive optics (AO), known for some time to be a route towards improved image quality achievable from retinal imaging.

Examples have included the NEI's 2021 development of adaptive optics scanning light ophthalmology (AOSLO), able to image the rod and cone photoreceptors in the retina through a combination of AO and annular pupil illumination, and which was able to improve imaging resolution by a third.

AO can potentially bring similar improvements to OCT imaging, but the cost is the longer time needed to process the OCT optical data into final pictures.

"Adaptive optics takes OCT-based imaging to the next level," said Johnny Tam from the Clinical and Translational Imaging Section at the NEI. "With AO, we can reveal 3D retinal structures at cellular-scale resolution, enabling us to zoom in on very early signs of disease. Artificial intelligence now helps overcome a key limitation of imaging cells in the retina, which is time."

A particular hurdle for AI to tackle is the speckle present in OCT images of the retinal pigment epithelium (RPE) cells, behind the retina's light-sensing layer. Removing the speckle can require researchers to repeatedly image cells over relatively long periods of time to capture the speckle shifts, and then undertake the time-consuming task of piecing together many separate images to create a clean picture of the RPE.

A paradigm shift for AI

The NEI team applied a novel deep learning algorithm termed a parallel discriminator generative adversarial network, or P-GAN. As implied by the name, generative adversarial networks involve two neural networks contesting with each other, with one trained on prior OCT data and generating outputs that could be mistaken for the real thing, and the other network tasked with identifying true results from false ones.

For its P-GAN implementation, the NEI fed the network 6,000 manually analyzed AO-OCT images of human RPE, each one paired with its corresponding speckled original, and trained the network to identify and recover speckle-obscured cellular features.

When subsequently tested on new images of RPE, the P-GAN algorithm successfully de-speckled images and recovered results comparable to the manual method. For a variety of objective performance metrics assessing cell shape and structure, P-GAN outperformed other AI techniques, according to the NEI team. The algorithm "reduced imaging acquisition and processing time by about 100-fold, and yielded greater contrast, about 3.5 greater than before," noted the project.

The NEI believes that this removes a significant obstacle for routine clinical imaging using AO-OCT, especially for diagnosis of diseases that affect the RPE.

"Thinking about AI as a part of the overall imaging system, as opposed to a tool that is only applied after images have been captured, is a paradigm shift for the field of AI," commented Johnny Tam. "Our P-GAN artificial intelligence will make AO imaging more accessible for routine clinical applications and for studies aimed at understanding the structure, function, and pathophysiology of blinding retinal diseases."

NEI video