76
C.-L. Hsieh
such cases, the temporal median image of the video is often a reasonable estimation
of the static background [54].
Single virus particles and intracellular vesicles have been imaged and tracked in
live cells using temporal median background correction and COBRI microscopy [31,
32]. Due to their small sizes, virus particles and cell vesicles move and diffuse constantly because of thermal fluctuation. By contrast, cell structures move and evolve
slowly. When capturing a COBRI video at high speed (>5,000 fps) for a few seconds,
the median background conveniently represents the scattering background of large
cell structures and other static background contributions (i.e., the non-uniform illumination). Many endogenous macromolecules in the cell membrane or inside the cell are
expected to move rapidly. If these macromolecules provide sufficient optical contrast,
they appear in the background-corrected images and make detection of the particle
signal more difficult. Thus, for optimal performance, the imaging sensitivity should
be sufficient for detecting the particle of interest but not unnecessarily high to ensure
that other smaller entities remain indiscernible. The capability of temporal median
background correction in removing cell background scattering has been evaluated
experimentally [31]. A cell peripheral area was imaged using COBRI microscopy
at 5000 fps for 1 s. A median background was calculated from the recorded video
and this was used to normalize the raw video frame by frame. In the normalized
video, no cell feature remained. Moreover, the temporal fluctuation of each pixel in
the background-corrected video corresponded to the photon shot-noise fluctuation,
indicating no additional cell background fluctuation. These findings show that temporal median background correction has favorable performance for removing cell
scattering background in high-speed COBRI microscopy.
Distinguishing the signal from the background according to their different spatiotemporal characteristics is a powerful strategy in label-free imaging. Temporal
median background correction is convenient, but it requires a highly static background and also a dynamic signal over the space. These requirements cannot be met
in some applications, for example, when the motion of the signal is highly localized
(close to the size of the signal) throughout the observation time. In such cases, some
pixels are continuously occupied by the signal, and the median background estimation at those pixels is biased. In principle, superior background estimation can be
made by accounting for the signal or background spatiotemporal characteristics, if
they are available.
A more general method for background estimation was proposed and demonstrated [50]. Instead of examining the data pixel by pixel independently, as in the
case of temporal median filtering, the new method exploits the information encoded in
neighboring pixels. Specifically, it takes advantage of a priori knowledge of the shape
of the signal. For small particles, their optical image is the point spread function (PSF)
of the microscope, typically an airy disk. Specifically, the background estimation is
optimized by minimizing the residual error of fitting the background-corrected image
with a known PSF through an iterative process. This optimization process repeats
itself until the estimation converges. Intuitively, because the signal spot moves over
C.-L. Hsieh
such cases, the temporal median image of the video is often a reasonable estimation
of the static background [54].
Single virus particles and intracellular vesicles have been imaged and tracked in
live cells using temporal median background correction and COBRI microscopy [31,
32]. Due to their small sizes, virus particles and cell vesicles move and diffuse constantly because of thermal fluctuation. By contrast, cell structures move and evolve
slowly. When capturing a COBRI video at high speed (>5,000 fps) for a few seconds,
the median background conveniently represents the scattering background of large
cell structures and other static background contributions (i.e., the non-uniform illumination). Many endogenous macromolecules in the cell membrane or inside the cell are
expected to move rapidly. If these macromolecules provide sufficient optical contrast,
they appear in the background-corrected images and make detection of the particle
signal more difficult. Thus, for optimal performance, the imaging sensitivity should
be sufficient for detecting the particle of interest but not unnecessarily high to ensure
that other smaller entities remain indiscernible. The capability of temporal median
background correction in removing cell background scattering has been evaluated
experimentally [31]. A cell peripheral area was imaged using COBRI microscopy
at 5000 fps for 1 s. A median background was calculated from the recorded video
and this was used to normalize the raw video frame by frame. In the normalized
video, no cell feature remained. Moreover, the temporal fluctuation of each pixel in
the background-corrected video corresponded to the photon shot-noise fluctuation,
indicating no additional cell background fluctuation. These findings show that temporal median background correction has favorable performance for removing cell
scattering background in high-speed COBRI microscopy.
Distinguishing the signal from the background according to their different spatiotemporal characteristics is a powerful strategy in label-free imaging. Temporal
median background correction is convenient, but it requires a highly static background and also a dynamic signal over the space. These requirements cannot be met
in some applications, for example, when the motion of the signal is highly localized
(close to the size of the signal) throughout the observation time. In such cases, some
pixels are continuously occupied by the signal, and the median background estimation at those pixels is biased. In principle, superior background estimation can be
made by accounting for the signal or background spatiotemporal characteristics, if
they are available.
A more general method for background estimation was proposed and demonstrated [50]. Instead of examining the data pixel by pixel independently, as in the
case of temporal median filtering, the new method exploits the information encoded in
neighboring pixels. Specifically, it takes advantage of a priori knowledge of the shape
of the signal. For small particles, their optical image is the point spread function (PSF)
of the microscope, typically an airy disk. Specifically, the background estimation is
optimized by minimizing the residual error of fitting the background-corrected image
with a known PSF through an iterative process. This optimization process repeats
itself until the estimation converges. Intuitively, because the signal spot moves over