3.6 Super-Resolution
Image Reconstruction
Unlike in conventional microscopy, the image in SMLM cannot be
acquired in a direct optical way, but should be calculated from the
list of localizations, obtained after fitting. Therefore, there are
several methods to calculate an image from SMLM data:
1. Histogram mode. The gray value of a super-resolution pixel
equals to the number of events detected within the pixel’s area.
The fluorescence intensity (number of photons) is not taken
into account. The pixel size chosen by the user strongly affects
the gray levels of the image: in case of too small pixel sizes, the
image would contain only “zeros” and “ones,” i.e., become
binarized. This mode is fast to calculate, but it can produce
noisy and pixilated images, especially for weak density of
localizations.
2. “Gaussian” mode. Every localization is represented as a Gaussian function with the width equivalent to the localization precision of the fluorophore, taking into account the number of
detected photons. This image is slower to calculate than the
histogram image, but it provides a smooth representation and
allows to decrease pixel size without risk of pixilation. However, this mode can reduce the resolution of the image [27].
3. Local density mode. The pixel’s gray values are calculated to be
proportional to the local density of fluorophores in the neighborhood of the pixel. The local density can be estimated, e.g.,
using Voronoi diagrams [9] or Delaunay triangulations
[27]. Even though this method is slower than the others (but
still reasonable as in the order of minutes), it provides smooth
images with preserved resolution even for weak signals.
3.7 Post-processing
of Localization Data
Post-processing of localization data includes methods such as correction of drift and chromatic aberrations that are essential for most
of experiments and advanced processing methods, such as clustering and colocalization analysis that are necessary for some experiments, depending on biological question and sample type or
acquisition setup (e.g., multi-color imaging or 3D data
acquisition).
For drift correction, there are two most feasible possibilities:
(1) correction using fiducial markers and (2) correction with crosscorrelation. For the first approach, photostable fiducial markers,
such as fluorescent beads, gold nanoparticles, or quantum dots
[28], need to be incorporated and fixed within the sample, usually
on the surface of the coverslip. The second approach does not need
fiducials and uses properties that are intrinsic to localization data
and rely only on computing [29]. The cross-correlation method
works well for contrasted structures with well-defined shape but
can be less efficient for diffuse structures. It can also be recommended for imaging far from the coverslip because the fiducials
usually can be robustly immobilized only on a glass surface.
280
Leonid Andronov et al.
Image Reconstruction
Unlike in conventional microscopy, the image in SMLM cannot be
acquired in a direct optical way, but should be calculated from the
list of localizations, obtained after fitting. Therefore, there are
several methods to calculate an image from SMLM data:
1. Histogram mode. The gray value of a super-resolution pixel
equals to the number of events detected within the pixel’s area.
The fluorescence intensity (number of photons) is not taken
into account. The pixel size chosen by the user strongly affects
the gray levels of the image: in case of too small pixel sizes, the
image would contain only “zeros” and “ones,” i.e., become
binarized. This mode is fast to calculate, but it can produce
noisy and pixilated images, especially for weak density of
localizations.
2. “Gaussian” mode. Every localization is represented as a Gaussian function with the width equivalent to the localization precision of the fluorophore, taking into account the number of
detected photons. This image is slower to calculate than the
histogram image, but it provides a smooth representation and
allows to decrease pixel size without risk of pixilation. However, this mode can reduce the resolution of the image [27].
3. Local density mode. The pixel’s gray values are calculated to be
proportional to the local density of fluorophores in the neighborhood of the pixel. The local density can be estimated, e.g.,
using Voronoi diagrams [9] or Delaunay triangulations
[27]. Even though this method is slower than the others (but
still reasonable as in the order of minutes), it provides smooth
images with preserved resolution even for weak signals.
3.7 Post-processing
of Localization Data
Post-processing of localization data includes methods such as correction of drift and chromatic aberrations that are essential for most
of experiments and advanced processing methods, such as clustering and colocalization analysis that are necessary for some experiments, depending on biological question and sample type or
acquisition setup (e.g., multi-color imaging or 3D data
acquisition).
For drift correction, there are two most feasible possibilities:
(1) correction using fiducial markers and (2) correction with crosscorrelation. For the first approach, photostable fiducial markers,
such as fluorescent beads, gold nanoparticles, or quantum dots
[28], need to be incorporated and fixed within the sample, usually
on the surface of the coverslip. The second approach does not need
fiducials and uses properties that are intrinsic to localization data
and rely only on computing [29]. The cross-correlation method
works well for contrasted structures with well-defined shape but
can be less efficient for diffuse structures. It can also be recommended for imaging far from the coverslip because the fiducials
usually can be robustly immobilized only on a glass surface.
280
Leonid Andronov et al.