Basic SMLM processing
Data processing in SMLM starts with localizing the single-molecule blinking events. As shown in the example above, after a typical acquisition you will have an image stack with raw camera frames, each containing a number of emitting molecules. In the simplest case, the localization algorithm first segments molecule candidates using a filter (e.g. local maximum), then performs a 2D Gaussian fitting to determine the center of the PSF. Importantly, the second step then reaches sub-pixel accuracy eventually allowing localization at nanometer precision.
Even though this process seems trivial, the performance of a localizer can vary dramatically and many excellent implementations have been developed. All with their own subtleties and specialties (e.g. 3D or high-density binking). An overview, comparison and benchmarking can be found on the websites of the SMLM software challenge. Since this might be overwhelming at first and many of these programs require expert knowledge or programming experience, in the Software section I give a few options to get you started.
Following the localization step, the dataset needs to be corrected for sample drift. Since SMLM acquisitions can take a while, the sample will drift during the experiment. While axial drift can often be compensated using an automatic focus-lock system, lateral drift is usually evident after a few thousand frames due to the nanometer precision of the localization. There are two simple ways to correct for lateral sample drift, one is software-based, the other one uses fiducial markers (beads) to track and subsequently correct the lateral drift. Both are explained in the section about drift correction.
Filtering then concludes the initial data processing. Simply speaking, low-quality localization need to be removed to achieve a high-quality representation of your SMLM dataset. Standard filters include:
- number of photons - remove localization with too few photons
- Min/Max sigma - remove out of focus localizations
- uncertainty - remove localization with high uncertainty
- frames - remove localization from high-density periods, i.e. if a pre-bleach was required and recorded
Images in SMLM are reconstructions of the localization data into a pixelated format. The simplest way to render an image is to bin the data into a 2D histogram. Since each localization is measured with a certain precision (i.e. the localization precision) this needs to be taken into account. Thus each binned localization is blurred using a gaussian filter with the width according to the localization precision.
Taken together, an SMLM reconstruction represents a probability map of the positions of all the molecules within the sample.
An important parameter is the pixel size. Although in SMLM we achieve nanometer precision, we might not have enough localizations (density) to populate all pixels within an image adequately (undersampling). Conversely, we don't want to lose resolution by choosing a too-large pixel size. Typically, choosing the pixel size around the localization precision (10-20 nm) is a good starting point. Changing the pixel size or other processing steps such as gaussian filtering will affect the representation of the dataset. See some examples below. It is thus important to consider the performance of your imaging setup, the size of the expected structures, and the density of the dataset when reconstructing an image.
