PSF Extraction

PYME has a reasonably complete toolchain for extracting PSFs for use in either 3D localization fitting or deconvolution. The process is very similar for both cases.

Single channel extraction

Step 1: Acquire a bead stack

Acquire a stack of bead images using the same objective, tube lens setting (if available) and excitation/emission channel as your data and save as .tif. To get a good PSF, you should consider the following:

Bead selection: The bead size should be smaller (preferably much smaller) than the PSF width, but should also be bright enough to give you good signal to noise. Ideally we’d use ~40 nm diameter beads but these tend to be too dim. In practice we usually use 200 nm beads and correct for the bead size later (see FIXME).

Bead density/spacing: You want to spread beads on the coverslip such that several beads in the field of view (to permit averaging), but also want the beads to be well separated (~ 60 pixels) from each other. This will likely require some trail and error with dilutions and also selection of ROIs when performing the imaging. A good starting point is a dilution of 1:2000 for 200 nm dark red beads, or 1:100 for tetra-speck beads.

Illumination/Integration time: We usually use the arc lamp for illumination as it gives us the largest usable field of view and allows us to average more beads. To get a good PSF, you want high SNR. Choose the integration time and illumination power to optimize this without clipping the signal in focus. We also tend to run somewhat slower so that we minimize any motion artifacts from moving the focus.

Axial step size: For localization, we get good results with 50 nm axial steps. For deconvolution the PSF must be Nyquist sampled in z and it doesn’t hurt to slightly oversample. We typically use 200 nm steps.

Stack z extent: For localization this should span from ~ 2um below the focus to ~2 um above. For deconvolution, ~5 times the axial FWHM on either side of the focus is recommended. In either case, ~ 100 slices should be taken (by default the extracted PSF size is 64 slices deep, and some allowance is needed for axial alignment of different beads).

Step 2: Load the bead stack in dh5view

Load the data in dh5view, if necessary using ImageJ to convert to a format that we can read (see PYME Data Formats). If the data was not generated in PYME, set the voxelsize when prompted.

Step 3: Activate the PSF extraction mode

Choose Modules->psfExtraction from the dh5view menu. This should create a PSF Extraction panel to the right of your data as seen in the folowing image:

_images/psf_extraction_gui.png

The PSF Extraction GUI

Step 4: Select beads to average

Select the beads to use by clicking on the beads and pressing the Tag button. A green square should appear around each bead image when you press Tag.

_images/psf_extraction_tagged.png

This square should automatically ‘snap’ to the centre of the selected bead, and is actually a 3D cube, as evidenced by looking at the x-z projection.

_images/tagged_psfs_xz.png

If the boxes intersect the top or bottom of the xz view (generally only a problem if the stack is too small or poorly centred), this will result in a somewhat cryptic error when it comes to the extraction step.

Notes:

  • The size of the green boxes and the ROI to be extracted is controlled by the ROI Size box. This is a half size, and the actual size extracted will be 2*ROI Size + 1 (i.e. we go 30 pixels to the left of our centre and 30 pixels to the right).
  • When tagging, you should try and avoid beads whose green boxes would overlap with those of other beads. These are too close and contributions from the neighbouring beads will corrupt the measurement
  • You should also try and avoid beads which might be a cluster or aggregate (usually significantly brighter than individual beads)
  • You can clear all tags with the Clear button

Note

If you also enable the blobFinding module you can automatically find beads using a simple threshold and the Object Finding panel. Clicking Find will detect objects (marked with a small green square) and clicking Tag Points will tag each of the found objects. In practice we tend to tag points manually however as we can more easily reject beads which are too close or clustered.

Step 5: Perform the extraction

Clicking on Extract will take each tagged ROI, find the centre of the bead, and use sub-pixel Fourier shifting to align and average all the ROIs. In addition to ROI Size there are 3 configurable parameters for the extraction:

Blur: We pass the extracted PSF through a Gaussian filter to reduce and residual noise. This is the filter sigmas in x, y, and z respectively. The defaults are reasonable for most PSFs, but if you have a lot of averages (or another way of getting good SNR) fit accuracy might be improved by reducing these (especially in z, x-y accuracy is relatively robust against oversize model PSF widths).

PSF Type: This should probably be renamed and dictates whether to normalize the intensity across slices in z. If PSF Type is Widefield we assume that the sum of intensity in the defocussed slices should be equal to that of the in focus slice and normalize each slice to enforce this. If it’s Confocal, no normalization is performed. Because this assumption about intensity doesn’t always hold (e.g. if there is clipping or apodization within the optical path) Confocal is often the right choice even when doing widefield imaging. This is especially relevant for biplane and splitter fits which might misbehave if there is is apodization and Widefield normalization is selected.

Align Z: This is only relevant when extracting multi-channel PSFs (see later).

Once you click on Extract you will get a new window which shows the extracted PSF

_images/extracted_psf.png

This has a number of sanity checks (on left hand side) which indicate potential problems with the PSF with traffic light colours. Click on any of the cells to get a description of what the particular metric means. There is also a tab which lets you look at the estimated CRLB for your extracted PSF and how it compares to a theoretical astigmatic case.

_images/extracted_PSF_CRLB.png

At this point you can save your PSF by choosing File->Save As or proceed to the next step

Step 6: Correcting for the bead size [optional]

Using finite size beads as a PSF estimate for either fitting or deconvolution has the disadvantage that the estimated PSF will be larger than the true PSF. In deconvolution this can lead to artifacts whereby the image is ‘overdeconvolved’ and a continuous structure broken up into a series of points. When performing localization an overly large PSF estimate will typically result in reduced axial localization precision, and z-dependant bias / distortion in the fitted values.

One way to partially address this problem is to try and correct for the bead shape. We do this using deconvolution with a twist. Rather than trying to estimate an unknown object from an image + PSF, we try and and estimate an unknown PSF from image + a known object model (in this case a sphere matching the bead diameter). To correct for bead size choose Processing->Deconvolve bead shape from the dh5view menu of your newly extracted PSF. Enter the correct bead diameter, and click OK. You will get a new window with the deconvolved PSF which can be saved by choosing File->Save As.

Multi-Channel PSFs

The PSF extraction tool has support for extracting PSFs from multiple channels simultaeneously, and for various different ratiometric 3D fitting regimes. If a box or button is not described above it is likely related to multi-channel PSFS.

TODO - flesh out the details.

General Tips and Observations

  • PSF sample preparation should match imaging sample preparation as closely as possible (i.e. same RI etc)
  • New PSFs should be taken regularly (ideally at the start of each imaging session)
  • If imaging with a TIRF objective and the sample RI does not closely match that of the immersion oil (or more specifically if the sample RI < the objective NA) there can be a significant super-critical angle fluorescence component in the PSF if beads on the coverslip are used. This has the potential to make the measured PSF a poor estimate of the true single molecule PSF if you are imaging more than ~500nm deep in the sample. We haven’t fully quantified this, but there might be an argument for abandoning the conventional best practice of preparing beads on the coverslip and using beads prepared on the slide and imaged through a thin layer of medium instead.