SIRT and SIRT-like Filter Reconstructions of a Cryo Tilt Series

(IMOD 4.11)

University of Colorado, Boulder

 

This example data set illustrates how to use the interface in Etomo for reconstruction with SIRT (Simultaneous Iterative Reconstruction Technique).  This method involves three steps at each iteration:


Because SIRT is relatively time-consuming and there is no clear point at which to stop iterating, the strategy is to do a trial reconstruction on a subarea and examine the results after a selected set of iterations.  One can then pick the number of iterations to use for the full reconstruction, a number that may be suitable for other similar data sets.

 

The SIRT-like filter in IMOD provides a fast way to emulate the results from SIRT.  It involves replacing the regular linear ramp for radial weighting (where the filter value is proportional to the frequency, f), with a curve of the form f * (1 (1 a/f)N).  Here a is a constant and N is an iteration number; in IMOD the value of a and an adjustment to get N from a given iteration number were worked out by matching the output of SIRT.  The following shows the SIRT-like filter for different values of N; the linear ramp would be a diagonal line to the upper right corner.

 

 

The example tilt series is of a preparation of bovine papilloma virus (BPV), taken by Mary Morphew on an F20 with a US4000 CCD camera at a defocus of -3.5 microns.

Getting and unpacking the data:

Making the aligned stack:

Making the SIRT reconstruction:

Using Filter Trials to Vary the SIRT-like filter

The reason for the contrast discrepancy between SIRT and the SIRT-like filter is that the contrast of the SIRT reconstruction depends on its thickness.  The scaling of the iteration number to make an exponent for the SIRT-like filter was initially tuned on this same data set to match SIRT with thicknesses of 500 pixels.  If you remake the SIRT subareas with a thickness of 400 or 500, you will see that the contrast matches much better for each iteration number. (You should make the subarea wider if doing a thickness of 500).  Apparently, SIRT converges more slowly on the R-weighted backprojection when the signal in each ray has to be spread over more pixels.  The SIRT-like filter is more predictable because it does not vary with thickness, but this fact does make it more difficult to transition from using SIRT to the SIRT-like filter.  This discrepancy was discovered just as this document was being revised for the IMOD 4.11 release.  The situation may well change in an upcoming release.