3076 Microsc. Microanal. 28 (Suppl 1), 2022
doi:10.1017/S1431927622011461 © Microscopy Society of America 2022
Fast Automatic Point Spread Function Deconvolution Using Edge Detection
Zachary E. Russell
1*
, Mathieu Therezien
1
, Tomas J. McIntee
1
, Shane T. DiDona
1
, Jeffrey J. Haggen
2
,
and Edward L. Principe
2
1.
Ion Innovations, Boone, NC, United States.
2.
Synchrotron Research Institute, Melbourne Beach, FL, United States.
* Corresponding author: zach.russell@ion-innovations.com
Point Spread Function Deconvolution (PSFD) in electron microscopy has proved challenging and lagged
behind similar work in optical microscopy until advances to the field by William E Vanderlinde [1] and
Eric Lifshin [2] and others over the last 20 years pushed the process forward. In this work we propose
and compare an alternative method for acquiring these Point Spread Functions (PSF).
Acquisition of the PSF can be an involved process in electron microscopy, often utilizing specific
calibration samples and involved processes. This is undesirable as the PSF can change while the
instrument is in use and changes can be an indication of poor instrument health such as a filament
reaching end of life. With large image montages becoming more and more common and desirable, a
shift in the PSF during a longer automated montage acquisition would result in a traditional PSFD being
unable to be performed at all. Therefore, it is desirable to have a method that can perform either a blind
or semi-blind extraction of the PSF during acquisitions.
By utilizing the mathematical relationship between the point spread, line spread, and edge spread
functions in conjunction with edge detection and feature extraction from the field of machine learning
and computer vision image processing [3], we can derive an estimated non-symmetric PSF from
arbitrary samples given a sufficient amount of image contrast. First edges are detected that are suitable
for edge spread function extraction (Figure 1: a, b, c) and then correlated and scaled to produce a PSF
kernel. Additionally, by utilizing denoising processes in tandem with PSFD we are able to dramatically
reduce the mottling effects often encountered when performing Richardson-Lucy [4, 5], Landweber [6],
or Tikhonov [7] PSFD (Figure 1: d, e, f).
Due to the speed at which this method can be performed and the simplicity of its operation, it is possible
to recompute this kernel for every image as they are acquired. By evaluating changes in the kernel the
software can provide feedback to automated acquisition systems to inform the user or routine if a long
montage or delayering process is drifting out of specified tolerances [9].
Downloaded from https://academic.oup.com/mam/article/28/S1/3076/6996818 by guest on 25 January 2023