🔬 optics

Event-sparse stack denoising for 4D-STEM applications

This paper introduces event-sparse stack denoising for 4D-STEM, a method that leverages time-resolved electron event data and sparsity-based algorithms to achieve comparable image quality at significantly reduced electron doses while enhancing defect detection sensitivity and enabling critical dose estimation.

Original authors: Gregory Nordahl, Rebekka Klemmt, Espen Drath Bøjesen

Published 2026-07-24

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Gregory Nordahl, Rebekka Klemmt, Espen Drath Bøjesen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to take a photograph of a tiny, fragile snowflake using a camera that is incredibly sensitive to light. If you use a bright flash, the snowflake melts instantly, ruining the picture before you can even see it. If you use a dim light, the picture comes out grainy and full of static, making it impossible to see the delicate details. This is the daily struggle for scientists who use a powerful tool called 4D-STEM (Four-Dimensional Scanning Transmission Electron Microscopy). Instead of taking a normal photo, this microscope shoots a beam of electrons at a material, point by point, to build a map of its atomic structure. The problem is that to get a clear picture, you usually need to blast the sample with so many electrons that you accidentally damage or destroy the very thing you are trying to study. It's a classic catch-22: you need a lot of light to see clearly, but the light itself destroys the view.

To solve this, scientists have tried various tricks, like using special filters to smooth out the grainy noise or grouping pixels together to make them brighter. But these methods often blur the fine details or require complex computer processing that takes forever. The big question has always been: Can we get a crystal-clear image of a delicate material without blasting it with enough electrons to melt it? The answer lies in a clever new way of taking pictures that treats electrons not as a continuous stream of light, but as individual, rare events, and then uses a smart computer program to clean up the mess.

The Paper's Story: Catching Ghosts to Clean the Image

In this paper, a team of researchers from Aarhus University in Denmark introduces a new method they call "event-sparse stack denoising." Think of it as a way to take a series of super-short, super-dim snapshots of a sample and then use a digital magic trick to combine them into one perfect image, all while using only a tiny fraction of the electron "dose" that would normally be required.

Instead of taking one long exposure where the electrons hit the detector all at once (which creates a noisy, blurry mess), the researchers use a special camera to take hundreds of tiny snapshots at each spot, each lasting just 1 microsecond (one-millionth of a second). Because the exposure is so short, most of the time, no electrons hit the detector at all. When they do, it's a rare, isolated "event." This creates a stack of images that are mostly empty black space with just a few scattered dots of light.

The researchers then run these stacks through two different "cleaning" pipelines to separate the real signal (the actual structure of the material) from the random noise (the electronic static).

The first method uses a technique called DBSCAN. Imagine you are at a crowded party where most people are just chatting in small groups, but a few people are shouting randomly in the corners. DBSCAN is like a bouncer who looks for clusters of people standing close together. If a person is standing alone in the corner (a random noise dot), the bouncer ignores them. If they are part of a group (a real electron hitting the detector to form a pattern), they get to stay. The researchers found that by looking for these clusters and then checking which spots appeared repeatedly across the hundreds of snapshots, they could build a clean map of the material.

The second method uses Sparse PCA. Think of this as trying to find the "main melody" in a song that is being played by a hundred different instruments, some of which are out of tune. This algorithm looks at all the snapshots and asks, "What pattern keeps showing up in a consistent way?" It keeps the consistent patterns (the real signal) and throws away the random, chaotic noise.

What They Found

The results are quite impressive, especially for delicate materials. When the researchers tested their method on simulated data (computer-generated models of gold nanoparticles), they discovered something remarkable. Their cleaned-up images, built from just 31 of the 200 tiny snapshots, looked just as good as a standard image built from all 200 snapshots. In plain English, they achieved the same image quality using only 16% of the usual electron dose. This means they could see the same details while exposing the sample to less than one-sixth of the radiation, which is a huge win for preserving fragile samples.

They also tested this on real-world samples. First, they looked at a thin film of gold. Even though gold is a strong scatterer that usually creates a lot of background "fog," their method cleared it up, making the tiny crystals stand out much more clearly.

Next, they looked at guanine nanocrystals (tiny crystals found in biological materials). These crystals often have subtle bends or defects that are hard to see. The researchers found that their denoising method made these defects pop out like never before. Specifically, they could detect shifts in the crystal patterns that were 4.1 times more sensitive than before. It's like upgrading from a pair of regular glasses to a pair of high-powered binoculars that can spot a single crack in a distant window.

Watching the Damage

One of the coolest side effects of this method is that it lets scientists watch the sample get damaged in real-time. Because they are taking hundreds of snapshots in sequence, they can see exactly when the electron beam starts to hurt the material. In their guanine crystals, they noticed that the signal started to drop after about 31 frames. By stopping there and throwing away the later, damaged frames, they could create an "exposure-selective" image that only shows the pristine, undamaged parts of the sample.

They also used this to measure the "critical dose"—the point where the material starts to break down. For guanine, they calculated that the critical dose is approximately 1.66 × 10⁴ e-/Å². This number is a vital piece of information for other scientists, as it tells them exactly how much they can "flash" the sample before it gets ruined.

The Catch and the Future

The paper does admit there are some limitations. The cleaning process isn't perfect; sometimes, if the signal is very weak, the method might accidentally remove a little bit of the real detail along with the noise. Also, the current method is limited by how fast the camera can write data to the computer's hard drive. They can currently capture about 100 snapshots per spot, but to get even better results, they would need to capture more, which requires faster data transfer.

However, the authors suggest that as new types of detectors (called event-based detectors) become available, these speed limits will disappear. These new cameras don't store empty space, making them much faster and more efficient.

In summary, this paper suggests that by taking many, many tiny, sparse snapshots and using smart math to find the patterns, we can see the invisible world of atoms much more clearly without destroying what we are looking at. It's a promising step toward a future where we can study the most fragile materials in the universe without melting them.

Technical Summary: Event-Sparse Stack Denoising for 4D-STEM

Problem Statement

Four-dimensional scanning transmission electron microscopy (4D-STEM) is a powerful tool for quantitative nanoscale characterization, yet it remains susceptible to noise, particularly in low-dose regimes required for beam-sensitive materials. While direct electron detectors (e.g., MAPS, HPAD) offer improved signal-to-noise ratios (SNR) and frame rates compared to scintillator-coupled detectors, they still introduce noise from readout electronics and energy deposition variations (e.g., Landau distribution noise).

Traditional denoising methods face significant trade-offs:

Filtering (Gaussian, Wiener, Non-local means): Often removes high-frequency components, leading to the loss of fine details.
Pixel Binning: Increases SNR but sacrifices spatial and reciprocal space resolution.
Dimensionality Reduction (PCA, NMF, Tensor SVD): These non-local methods analyze spatially separated points to extract variance. While effective, they can struggle with aperiodic data (e.g., polycrystalline or disordered materials) and are computationally intensive.
Deep Learning: Unsupervised convolutional neural networks show promise but are computationally demanding, slow, and inherently non-local.

There is a need for a denoising approach that is local, independent of scan position, computationally efficient, and capable of preserving structural information in low-dose, event-sparse data.

Methodology

The authors introduce Event-Sparse Stack Denoising, a method that transforms standard 4D-STEM data collection into Locally Time-Resolved STEM (LTR-STEM). Instead of integrating a single diffraction pattern per scan position, the detector records a stack of multiple, ultra-short exposure (approx. 1 µs) diffraction patterns at each position. This creates a five-dimensional (5D) dataset: $(x, y, \text{stack}, \text{pixel}_x, \text{pixel}_y)$ .

Two distinct sparsity-based denoising pipelines are applied to the pattern stack at each scan position before summation:

Clustering-Based Denoising (DBSCAN):
- Utilizes the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm to identify clusters of electron events.
- Parameters: eps (radius approximating a Bragg reflection) and min_samples (local electron count density).
- Process: Isolated intensity counts (noise) are treated as outliers and removed.
- Thresholding: A two-step thresholding process is applied to the remaining counts to generate a mask:
  - Soft Threshold: A sigmoid function creates a probability-like mask based on a persistence map (accumulated counts across the stack).
  - Hard Threshold: A binary cutoff is applied to the soft mask to finalize the diffraction mask.
- This approach is robust against low-frequency noise and preserves the physical location of Bragg reflections.
Decomposition-Based Denoising (Sparse PCA):
- Applies Sparse Principal Component Analysis to the stack, decomposing it into orthogonal components ordered by explained variance.
- Process: The data is reconstructed using only a single component ( $n_{comp}=1$ ). This acts as a data-driven sum where only pixels exhibiting consistent, structured variance are retained; random noise is suppressed to near-zero.
- Thresholding: A final low-intensity threshold (typically $10^{-2}$ to $1$) is applied to the reconstructed pattern.
- This method requires fewer input parameters (primarily an intensity threshold) and is simpler to implement.

Key Results

1. Performance on Simulated Data

Using simulated Au [001] diffraction patterns corrupted with Poisson noise:

PSNR Analysis: Both methods yielded higher Peak Signal-to-Noise Ratio (PSNR) than noisy reconstructions across almost the entire range of component patterns ( $n$ ).
Dose Efficiency: The DBSCAN method achieved the same PSNR as a noisy reconstruction using all 200 components when using only 31 components. This represents a ~16% reduction in exposure time (and dose) while maintaining comparable image quality.
Artifact Suppression: A soft-hard double-thresholding procedure was found necessary for DBSCAN to prevent discontinuities in the PSNR curve.

2. Virtual Imaging of Polycrystalline Gold

Applied to experimental data of a 2–3 nm polycrystalline Au film:

Contrast Improvement: Denoising effectively suppressed the substantial diffuse background signal inherent to gold, significantly improving the contrast of Virtual Dark Field (VDF) images.
Trade-off: Both methods resulted in a partial loss of diffraction disk structure, particularly for weak, high-angle reflections. DBSCAN retained intensities closer to raw data, while Sparse PCA offered more aggressive noise suppression.

3. Defect Mapping in Guanine Nanocrystals

Applied to guanine nanocrystals exhibiting crystallographic bending:

Sensitivity Enhancement: Denoising significantly improved the detection of subtle radial disk shifts caused by bending.
Quantitative Gain: The sensitivity to relative radial disk shift increased by 4.1×. Raw data showed peak-to-trough variations of ~0.9%, which increased to ~3.7% after denoising.
Mechanism: The DBSCAN method, with specific parameter tuning (lower eps), effectively captured intensity redistribution within the disk that was obscured by noise in the raw data.

4. Critical Dose Estimation and Exposure-Selective Imaging

The LTR-STEM acquisition strategy enabled the monitoring of beam-induced degradation in real-time:

Dose Tracking: By analyzing the integrated intensity of the (100) disk across the stack, the authors identified a monotonic decrease in signal due to beam damage.
Critical Dose: For guanine, a critical dose ( $D_C$ ) of approximately $1.66 \times 10^4$ e-/Å² was determined (where intensity drops to $1/e$ of the initial value).
Selective Imaging: The method allows for "exposure-selective imaging," where only frames prior to significant degradation (e.g., retaining $\ge$ 90% signal) are used for analysis, effectively discarding damaged data.

Significance and Claims

The paper claims that Event-Sparse Stack Denoising offers a practical, computationally efficient, and local strategy for low-dose 4D-STEM. Its significance lies in:

Dose Reduction: Achieving high-quality image reconstruction with significantly lower electron doses (~16% of standard exposure in simulations), which is critical for beam-sensitive materials.
Flexibility: Unlike non-local dimensionality reduction techniques, this method is independent of scan position, making it suitable for aperiodic, polycrystalline, and disordered materials.
Enhanced Sensitivity: It enables the detection of subtle structural features (e.g., crystallographic bending) that are otherwise obscured by noise.
Damage Characterization: The LTR-STEM acquisition mode facilitates the direct estimation of critical dose thresholds and enables exposure-selective imaging protocols.

Limitations

The authors acknowledge specific limitations:

Information Loss: There is an inherent trade-off between noise removal and the preservation of weak, high-angle diffraction disk structures.
Throughput Constraints: The method is currently limited by the data throughput of frame-based detectors (e.g., MerlinEM), where readout gap times restrict the number of patterns that can be captured per stack. The authors suggest that event-based detectors (e.g., TimePix3) with negligible gap times could resolve this bottleneck.