Kevin C. Zhou

Kevin C. Zhou

Incoming Assistant Prof. @ UMich BME

Research

A major theme of my research is developing both software and hardware for high-throughput computational 3D imaging techniques. My interests span the broadest, most inclusive interpretation of “3D”, ranging from dense 3D tomography to sparse 3D topography, from the microscale to the mesoscale and macroscale. Some projects include:

High-speed 4D fluorescence light field tomography of whole freely moving organisms

We developed a new technique, Reflective Fourier Light field Computed Tompography (ReFLeCT), which can capture synchronized video from perspectives spanning 2π steradians across 10s of cubic mm volumes of view, enabling multimodal volumetric video at up to 120 fps. While existing volumetric imaging approaches applied to model organisms such as zebrafish and Drosophila larvae require immobilization, ReFLeCT’s high spatiotemporal throughput enables 4D imaging of whole organisms as they are freely moving!

ReFLeCT

Demonstration of ReFLeCT applied to a freely moving Drosophila larva expressing GFP in its pericardial cells:

ReFLeCT applied to a freely moving zebrafish larva:

The left 6×9 panels of the above two videos are the synchronized raw multi-view video capture, and the right 2×2 panels are four different visualizations of the 4D reconstruction.

Relevant publications

Code

High-throughput computational 3D imaging with camera arrays

We present 3D-RAPID (3D Reconstruction with an Array-based Parallelized Imaging Device), a computational 3D imaging technique based on an array of 12-megapixel cameras capable of imaging at extremely high spatiotemporal throughts (>5 gigapixels/sec). 3D-RAPID features a spatiotemporally scalable, physics/self-supervised algorithm that enables simultaneous 3D reconstruction and stitching of arbitrarily many cameras across arbitrarily many temporal frames.

3D-RAPID

Our high spatiotemporal throughputs (>5 GP/sec) enable high-resolution 3D imaging of freely behaving organisms over wide FOVs (>130 cm2) at high speed (up to 230 fps):

The left panel shows the zoomed-out view of the entire arena of freely swimming zebrafish; the right panels show the zoomed-in individually tracked zebrafish. The video alternates between photometric (RGB) and 3D height information.

Freely moving harvester ants (left: photometric frame, right: 3D height map).

For more videos, see the supplementary materials of the accompanying publication below. See also https://gigazoom.rc.duke.edu/ for interactive views of full video frames.

Relevant publications

Code

Optical coherence refraction tomography (OCRT)

OCRT is a new computational 3D microscopy technique that incorporates extreme angular diversity to incoherently enhance resolution and reduce speckle of conventional OCT.

OCRT k-space synthesis

OCRT registers multi-angle OCT images or volumes by solving an inverse refraction problem, a process that also computationally reconstructs a 3D refractive index map of the sample. The following videos demonstrate the marked improvement of 3D OCRT over conventional OCT volumetric data (zebrafish larva and mouse esophagus, respectively):

Relevant publications

Code

Multiview imaging over extremely wide angular ranges

We introduce a class of techniques that can achieve multi-view imaging over extremely wide angular ranges (theoretically up to 4pi steradians) based on conic section mirrors (e.g., parabolic or ellipsoidal mirrors). Although traditionally such mirrors are known to exhibit off-axis aberrations and are thus more widely used for on-axis focusing or collimation, we show theoretically and practically that they are particularly well-suited for multi-view 3D imaging applications.

multiangle imaging with parabolic mirror

The above figure shows multi-view OCT imaging of a fruit fly head from up to ±75°. The fruit fly images are color-coded by view angle, as denoted by the ray traces on the left. These images were used for 3D OCRT reconstructions.

Relevant publications

Diffraction tomography with a deep image prior

Optical diffraction tomography is a coherent 3D imaging modality that reconstructs a 3D refractive index (RI) distribution based on multi-angle diffraction patterns. A well-known problem particularly when using low-NA objectives is the “missing cone” centered about the kz axis in the Fourier domain, which reduces the axial resolution. At the same time, however, low-NA objectives tend to have higher space-bandwidth products (SBPs), a property exploited in Fourier ptychography for 2D samples.

deep prior diffraction tomography

To combat the missing cone problem, we reparameterized the 3D RI distribution as the output of an untrained CNN, and optimized its parameters instead of the 3D volume directly during the 3D reconstruction process. Interestingly, the inherent biases of the CNN structure, even without pretraining, are enough to fill in the missing cone!

Relevant publications

Code