BioPhotonics spoke with Kevin Zhou, a postdoctoral fellow in the Biomedical Engineering Department at Duke University, who received the Barry Goldwater Scholarship, the National Science Foundation Graduate Research Fellowship, and the Schmidt Science Fellowship. Zhou and his colleagues recently published a paper (www.doi.org/10.1364/optica.454860) about a new computational 3D microscopy method — called 3D optical coherence refraction tomography (OCRT) — that enhances the resolution and contrast provided by OCT over a 3D field of view. The researchers envision that this technique will be useful for in vivo imaging of the eye or skin, and they are working to miniaturize the imaging system and interpret the data using machine learning algorithms.

Is the goal of your method to reduce the speckle found in traditional OCT images and to expand the field of view?

There is an inverse relationship between the axial field of view and lateral resolution, due to the physics of light propagation. Thus, in most biomedical imaging applications, people give up lateral resolution to obtain large axial imaging ranges. Furthermore, OCT exhibits speckle noise due to interference between reflections from unresolved microscopic and nanoscopic structures within the sample, which can obscure useful information. In our paper, we presented a new computational extension of OCT, called 3D OCRT, featuring a parabolic mirror that increases the angular diversity, and a new reconstruction algorithm, which together enhance the lateral resolution and nearly eliminate the speckle noise. As a result, we were better able to observe elusive features in several different biological samples, such as retinal structures in the zebrafish eye and muscle fibers in mouse esophagus.

Can you describe the previous developments that led to this technique?

Previously we had published a proof-of-concept paper in Nature Photonics, in which we had cut up and inserted the sample into a glass tube, which we then mechanically rotated to acquire OCT images from multiple angles. We got some nice results, but there were several hurdles we had to overcome to make OCRT more widely applicable. Most people don’t want to cut up their samples and stuff them into tubes. Further, tube rotation is slow, and you can only rotate about one axis, so we thought of alternate ways to access wide angular ranges at high speed without moving the sample. We initially thought about using high-numerical-aperture objectives, but we ran into vignetting issues at higher angles. We also thought about mechanically rotating the imaging system, which has been done before, but we wanted the potential for it to be fast. Finally, we thought about using parabolic mirrors, which at the time sounded like a crazy idea, since they are known to have a lot of aberrations and are almost never used for imaging. After doing some theoretical modeling, we found that parabolic mirrors are uniquely well suited for OCRT because the aberrations are no worse than the defocus effects in conventional OCT that limit lateral resolution.

What applications will this new technique most benefit, and what obstacles remain?

We’re still thinking about applications, but we believe the specialties that will benefit the most include dermatology, ophthalmology, and gastroenterology. We could, for example, image the anterior segment of the eye and characterize the refractive properties of the cornea and crystalline lens, which in the past have been difficult to measure in vivo. To make OCRT more useful in gastroenterology, we want to miniaturize the components, including the parabolic mirror, so that they can fit in an endoscopic probe. We also want to improve the speed of the data acquisition by improving the hardware and exploring subsampling strategies, compressive sensing, and motion-correction algorithms that have been used in other 3D reconstruction problems. Finally, we want to explore how we can tailor the image contrast to different applications. In particular, for each sample, we collect a very rich, high-dimensional data set, which we then reduce to a 3D representation of the sample. In our paper, we used the simplest strategy, which was simply to average across all angles, but we can imagine alternative dimensionality-reduction strategies that could highlight specific features that could be diagnostically relevant, depending on the application.