Research

Optical Coherence Tomography

Optical coherence tomography (OCT) is a non-invasive label-free imaging technique that utilizes light waves to capture high-resolution cross-sectional and volumetric images of biological tissues. Similar to how B-mode and C-mode ultrasound provide images of the body's internal structures, OCT can rapidly detect optical backscattering in tissues. This allows for the creation of detailed 3D images with microscale resolution, providing valuable insights into the structure and function of living tissues. Invented by Prof. Fujimoto and his colleagues at MIT in 1991, OCT has grown into a standard examination in ophthalmology and an emerging tool for intravascular research in cardiology. 

Imaging human or animal biotissue in vivo is often challenging due to the limited number of backscattered photons resulting from weak scattering in the sample and the need to adhere to maximum permissible exposure for laser safety. OCT overcomes this challenge by optically amplifying the backscattered light through interference with a higher-intensity reference plane wave. The backscattered photons interact with a larger number of plane-wave photons, creating an orders-of-magnitude larger signal. A depth profile along a single axial line, called axial scan or A-scan, can be recorded by measuring this amplified interference signal across a wide wavelength band.

An OCT system can be implemented in various configurations. Early-day OCT systems used a broadband light source and a physically moving reference arm scanning one depth position at a time, thereby receiving the name time-domain OCT (TD-OCT). Current generation of OCT systems are called Fourier-domain OCT (FD-OCT), which involves a wavelength-swept laser (swept-source OCT) or a spectrometer based on a line-scan camera (spectral-domain OCT), to detect the whole interference spectrum almost instantaneously without moving the reference arm. One of the core research topics in our lab is to develop multi-MHz wavelength-swept lasers to enable snapshot-like 3D OCT imaging.

3D Imaging of Retinal Neurons and Capillaries In Vivo

We develop hardware and computational techniques to visualize cells and capillaries of living tissue  by manipulating data in the 3D Fourier domain (k-space).  Imaging of retinal neurons (including photoreceptors) is important for detection of serious diseases such as age-related macular degeneration and diabetic retinopathy that can lead to blindness. However, achieving cellular-scale resolution is challenging due to refractive errors such as astigmatism.

Traditionally, these refractive errors were removed physically using adaptive optics, which can be understood as applying a very precise contact lens to a virtual pupil. However, adaptive optics devices are bulky and difficult to maintain. Therefore, we use our ultrahigh-speed swept lasers and computational techniques to solve this problem.

Our approach using multi-MHz swept lasers and 3D k-space (spatial frequency domain) image processing allows visualization of micron-scale structures such as photoreceptor cells and retinal nerve fibers with simple OCT architecture widespread in the clinic. Cellular-resolution 3D images of 3 mm x 3 mm area in the human retina can be acquired within less than 3 seconds. Our goal is to develop a powerful, easy-to-operate cellular-scale imaging system and image processing software that can be used by clinicians. We are preparing for a collaboration with ophthalmology clinic at Keimyung Univserity Dongsan Medical Center to demonstrate the efficacy of our technique in real patients.

Our imaging hardware and computational techniques can be also applied to small animals for preclinical studies. Rapid defocus-free angiography of various rodent organs such as the eye, the brain, and the kidney are all possible with multi-MHz OCT combined with 3D k-space image processing. Small animal imaging is powerful for disease model studies and functional studies such as neurovascular coupling.

AI-Based Cellular-Resolution Image Processing

We are also deeply interested in "visualizing the invisible" with artificial intelligence (AI) techniques. Many neurons in the retina are not visible on OCT because of their translucent nature. Other important cells are difficult to visualize due to heavy scattering from neighboring structures. These cells include retinal ganglion cells  and retinal pigment epithelium (RPE) cells. Without AI enhancement, such cells can be only visualized by averaging more than 100 adaptive optics OCT volumes. Since the high number of averaged images greatly limit the imaging field of view, there is a need for potential AI solutions which generates plausible guesses of these cells from just one OCT volume.

Pioneers in our field have shown that using adaptive optics and a novel type of generative adversarial network (GAN), individual cells in the RPE can be visualized from just one 3D OCT volume. They trained the network with individual RPE images as the input and 100+ times averaged RPE images as the target. The results were impressive although the network's performance in patients with diseases is unknown.

Currently, we are investigating if we can enhance the performance and credibility of the RPE images using phyics-informed neural network (PINN). The heavy scattering that prevents clear visualization of RPE is caused by photoreceptors, which we already can image. We think that by simulating the scattering at the photoreceptors using PINN, more reliable and trustworthy enhancement of the RPE images could be possible, even in diseased patients.

Wavelength-Swept Lasers

Wavelength-swept lasers stand as the powerhouse driving some of the fastest OCT systems to date. These lasers can typically be assembled in standard laboratory environments, utilizing fiber-optic components based on technology originally developed for telecommunications. Having an advantage in high-speed light source technology significantly enhances the generation of new imaging results. In our lab, we are spearheading the development of a groundbreaking wavelength-swept laser technique known as stretched-pulse active mode locking (SPML). This work is being done in collaboration with the Harvard-MGH Wellman Center for Photomedicine and KAIST.

SPML achieves wavelength sweeps at rates of multiple million repeats per second (multi-MHz), or even tens of MHz, enabling snapshot-like 3D or video-rate 4D imaging across wide fields of view or with high transverse pixel resolution. We envision that these research efforts will lead to new imaging capabilities with significant potential for clinical and biological research applications.

One of the unmatched advantages of SPML is its inherent consistency in sweep-to-sweep time, which results from not needing mechanical tuning filters. This consistency allows for the acquisition of 3D complex-valued data that maintains phase consistency, enabling the use of 3D k-space image processing techniques explained above.