Robust electron counting for direct electron detectors with the Back-Propagation Counting method
Joshua Renner, Matthew A. Wright, Kristofer Bouchard, Bruce E. Cohen, Peter Ercius *, Azriel Goldschmidt *, Cassio C.S. Pedroso, Ambarneil Saha, Peter Denes · 2025
Read Robust electron counting for direct electron detectors with the Back-Propagation Counting method by Joshua Renner, Matthew A. Wright, Kristofer Bouchard, Bruce E. Cohen, Peter Ercius *, Azriel Goldschmidt *, Cassio C.S. Pedroso, Ambarneil Saha, Peter Denes (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-03933.txt.
Introduction
Joshua Renner and eight colleagues present Back-Propagation Counting, a new way to count electrons for direct electron detectors in electron microscopy. Single-electron counting improves signal quality, but at high fluence several electrons can strike one pixel and standard counting breaks down.
The method borrows machine learning tools built for large tensor operations yet needs no large training dataset. On synthetic data it counted multiple strikes per pixel and stayed robust as occupancy rose. On real data it reconstructed diffraction peaks from single nanoparticles with higher relative intensity and produced images with better contrast than a standard method. High-flux movies and diffraction work stand to benefit most.
Catalog record
| CREATOR | Joshua Renner, Matthew A. Wright, Kristofer Bouchard, Bruce E. Cohen, Peter Ercius *, Azriel Goldschmidt *, Cassio C.S. Pedroso, Ambarneil Saha, Peter Denes |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 39 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-03933
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Robust electron counting for direct electron detectors with the Back-Propagation Counting method Joshua Renner, Matthew A. Wright, Kristofer Bouchard, Bruce E. Cohen, Peter Ercius *, Azriel Goldschmidt *, Cassio C.S. Pedroso, Ambarneil Saha, Peter Denes Abstract Electron microscopy (EM) is a foundational tool for directly assessing the structure of materials. Recent advances in direct electron detectors have improved signal-to noise ratios via single-electron counting. However, accurately counting electrons at high fluence remains challenging. We developed a new method of electron counting for direct electron detectors, Back-Propagation Counting (BPC). BPC uses machine learning techniques designed for mathematical operations on large tensors but does not require large training datasets. In synthetic data, we show BPC is able to count multiple electron strikes per pixel and is robust to increasing occupancy. In experimental data, frames counted with BPC are shown to reconstruct diffraction peaks corresponding to individual nanoparticles with relatively higher intensity and produce images with improved contrast when compared to a standard counting method. Together, these results show that BPC excels in experiments where pixels see a high flux of electron irradiation such as in situ TEM movies and diffraction. 1 Introduction Improvements in electron microscopy (EM) techniques have been essential to advancing materials science [ 1 , 2 ] and molecular biology [ 3 , 4 , 5 ] via imaging of objects at the nanoscale. Recent developments in direct electron detectors [ 6 ] have increased the capabilities of electron microscopy, enabling improvements in imaging [ 7 , 8 ] , diffraction [ 9 , 10 ] , and spectroscopy [ 11 ] . One type of detector, called an active pixel sensor (APS) [ 12 ] , utilizes a complementary metal-oxide-semiconductor (CMOS) sensor with fast readout in order to reduce the number of electrons per frame to below about 1% of the pixel number. Thus, each electron strike produces a localized deposition of energy that can be differentiated from other electron strikes and essentially “counted” as one electron regardless of the energy deposited [ 13 ] . The process of converting an extended energy deposition to a single electron strike is known as electron counting. This process improves the signal-to-noise ratio relative to the raw sensor response and is central to EM experiments that involve direct electron detection based on the APS design [ 14 , 15 ] . The “effective” detector quantum efficiency can be increased with improved electron counting algorithms, producing high signal-to-noise frames at a wider range of fluences. Counting electron strikes in APS detectors is challenging for several reasons. First, the amount of energy deposited in a thin volume of material by an energetic particle passing through it follows a Landau distribution [ 16 ] . This captures both the most probable value of energy deposition, as well as a long tail extending to the full energy of the particle resulting from relatively rare interactions that leave behind large amounts of energy. For a large number of individual electron strikes in a given pixel the total energy deposition tends towards the most probable value, but for one to several hits, the Landau tail makes it difficult to quantify exactly the number of electrons that hit a pixel. Second, an electron can undergo multiple scattering in the medium, changing its direction of travel, producing clusters of energy responses with an inconsistent shape and extent [ 17 ] . Finally, there is electronic noise [ 18 ] intrinsic to the CMOS hardware that can mask electron strikes with low energy deposition (false-negatives) or be incorrectly counted as electron strikes (false-positives). A counting algorithm that attempts to normalize the response to a single electron is therefore important in improving image quality. However, developing such an algorithm is challenging due to issues in segregating low-energy depositions from electronic noise and distinguishing multiple electron hits from single hits. These challenges are exacerbated by variations in the magnitude and shape of the energy deposition across the detector. APS sensors are being used to implement a new technique called four-dimensional scanning transmission electron microscopy (4D-STEM) [ 19 ] , a mode of operation capable of mapping crystallographic information with atomic- to nanoscale detail. In this technique, a focused electron probe is scanned over a sample in a two-dimensional (2D) raster pattern and a 2D diffraction pattern is measured at each probe position (see Figure 1 a). This results in a four-dimensional (4D) dataset with the potential to be hundreds of gigabytes to terabytes in size. 4D-STEM has become practical with the availability of fast, sensitive detectors leading to the need for computer hardware and algorithms to process such large datasets. To deal with this, electron counting algorithms are typically optimized for and built into specific camera hardware enabling in-line data processing during the experiment. Thus, a fast implementation of a counting algorithm is used (see [ 20 ] for an open source implementation) which relies on simple thresholds to eliminate noise and energetic x-rays. These simple algorithms can only accommodate a low flux per frame on the camera and do not consider the shape of energy deposition over multiple pixels. Other approaches to electron counting use artificial intelligence, which has been increasingly employed in electron microscopy in recent years [ 21 , 22 , 23 , 24 ] . In particular, convolutional neural networks (CNN) have been studied to improve localization accuracy and reduce coincidence loss [ 25 , 26 ] . However, the large numbers of parameters in these models, which must be optimized through back-propagation and stochastic gradient descent, require large amounts of simulated data and computation for training, and retraining is necessary for different detectors and electron energies. An approach that takes into account the energy deposition shape of individual electrons on the detector but does not require extensive detector/energy specific training could lead to substantial improvements that are broadly applicable, but are nascent [ 13 ] . Figure 1: Illustration of 4D-STEM and Back-Propagation Counting (BPC). (a) A scanning electron nanobeam incident on a sample produces a diffraction pattern captured by a pixelated APS-based detector. Simultaneously, high-angle scatterers are captured by a monolithic high-angle annular dark field (HAADF) detector. (b) An estimated counted frame is computed from the raw frame and convolved with the pre-computed 2D kernel to produce a reconstructed frame. Back-propagation is used to iteratively update the estimate according to a loss function. (c) The final optimized count grid is rounded to integer values. The ground-truth is shown for comparison but is not used in the algorithm. To enable more accurate nanoscale imaging, we developed a novel method called Back-Propagation Counting (BPC) for electron counting. The algorithm makes use of modern machine learning techniques and software [ 27 ] and back-propagation [ 28 , 29 ] to fit the energy deposition of a raw 2D frame to a “counted” frame of the same dimensions based on the expected average energy deposition of single electron strikes. The process does not involve training deep neural networks and thus does not require large amounts of training data. It can be readily applied with minimal understanding of the detector used in the acquisition. BPC processes low-occupancy frames with similar accuracy to standard methods and accommodates higher electron flux experiments with