🔬 physics

Adaptive Beam Selection for Efficient Scanning Probe Tomography

This paper proposes a novel sequential design method for X-ray tomography that directly identifies edge-aligned measurements from sinograms to bypass reconstruction, thereby improving computational efficiency and reconstruction quality while reducing radiation dose and measurement redundancy.

Original authors: San Dinh, Zichao Wendy Di, Matt Menickelly

Published 2026-06-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: San Dinh, Zichao Wendy Di, Matt Menickelly

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 figure out what a mysterious object looks like on the inside, but you can't touch it or take it apart. You have a powerful X-ray scanner, but there's a catch: every time you take a picture, it costs money, takes time, and can potentially damage delicate samples (like fragile biological tissues) with radiation.

The traditional way to solve this is to take thousands of pictures from every possible angle, like a camera spinning around the object. But the authors of this paper ask: "Do we really need to take every picture?"

Here is a simple breakdown of their solution, using everyday analogies.

The Problem: The "Blind" Scanner

Usually, to get a clear 3D picture (tomography), you need a lot of data. But taking too much data is slow and harmful.

The Old Way: Scientists try to be smart by guessing where the "edges" of the object are. They take a few pictures, build a rough 3D model, look at the model to see where the edges are, and then take more pictures there.
The Flaw: This is like trying to draw a map of a city by first drawing a blurry sketch, looking at the sketch to find the roads, and then trying to draw the roads more clearly. If your first sketch is wrong (which it often is with limited data), your next guesses will be wrong too. Plus, building that sketch takes a lot of computer power.

The Solution: Reading the "Shadow Map" Directly

The authors propose a clever shortcut. Instead of building a 3D model first, they look directly at the raw data coming from the scanner, which they call a sinogram.

The Analogy: The Shadow Puppet Show
Imagine the object is a hand puppet, and the X-rays are a light source. The "sinogram" is the shadow the puppet casts on the wall.

In a normal scan, you move the light around in a circle, taking a shadow picture every few degrees.
The authors realized that edges in the object create sharp, distinct lines in the shadow. You don't need to reconstruct the whole puppet to see where the sharp lines are in the shadow; you can just look at the shadow itself!

How Their Method Works

Their new method acts like a smart, adaptive detective that picks the next best "shadow" to look at, without ever needing to build the 3D puppet first.

The "Gaussian Process" (The Crystal Ball): They treat the raw shadow data like a smooth, continuous surface. They use a mathematical tool (a Gaussian Process) to guess what the shadow would look like in the spots they haven't scanned yet. It's like having a crystal ball that fills in the gaps between your measurements.
The "Edge Detector" (The Magnet): They know that the most important information is hidden in the sharp edges of the shadow. Their system calculates where the shadow changes most drastically (the edges) and treats those spots like magnets, pulling the scanner to focus there.
The "Exploration vs. Exploitation" Balance:
- Exploitation: "Hey, I see a sharp edge here! Let's scan right next to it to get more detail."
- Exploration: "I'm not sure what's happening over there; the data is fuzzy. Let's scan that area to clear up the confusion."
- The system balances these two, ensuring it doesn't just stare at one spot but also explores the unknown.

The Result: Smarter Scans

In their experiments, they tested this on five different "mystery objects" (ranging from simple shapes to complex biological samples).

The Comparison: They compared their "smart, adaptive" scanner against a "dumb, uniform" scanner that just takes pictures at regular intervals.
The Outcome: The smart scanner built a much clearer picture using the same number of shots as the dumb scanner.
Noise Handling: Even when they added "static" (noise) to the data to simulate a messy environment, their method still produced a clearer image than the traditional method.

The Bottom Line

This paper introduces a way to scan objects that is faster, cheaper, and safer for delicate samples. By looking directly at the raw "shadow" data to find edges, they skip the slow, error-prone step of building a 3D model first.

Important Note: The authors explicitly state that while their computer simulations show this works perfectly, the actual hardware to move the X-ray beam to these specific, random spots (instead of just spinning in a circle) doesn't exist yet. They are proving the idea works mathematically to encourage engineers to build the hardware in the future.

Technical Summary: Adaptive Beam Selection for Efficient Scanning Probe Tomography

Problem Statement
In X-ray tomography, particularly at synchrotron facilities, high-quality reconstruction typically requires a large number of projections. However, increasing the number of projections leads to higher experiment costs, longer acquisition times, and increased radiation doses, which can damage delicate or radiation-sensitive samples. While sequential experimental design offers a mitigation strategy by selecting measurements based on previously acquired data to maximize information gain, existing approaches face significant challenges. Most current methods rely on reconstructing the sample at each iteration to identify "edge-aligned" beams (projections intersecting sample boundaries). This reliance on intermediate reconstructions introduces computational bottlenecks, as the dimension of the covariance matrix scales with image resolution, and creates susceptibility to reconstruction errors. Furthermore, many existing sequential approaches optimize only projection angles for raster scans, failing to leverage the full potential of adaptive scanning positions in scanning probe tomography.

Methodology
The authors propose a novel sequential design method that performs adaptive beam selection directly in the sinogram domain, bypassing the need for intermediate image reconstruction. The core methodology involves the following steps:

Gaussian Process Modeling: The sinogram is modeled as a zero-mean Gaussian Process (GP) with a Matérn-3/2 kernel. This allows for the interpolation of unmeasured beam values and the quantification of uncertainty via confidence intervals.
Edge-Alignment in Sinogram Space: Instead of detecting edges in a reconstructed image, the method defines an "edge-mapping function" ( $\hat{E}$ ) directly on the sinogram. This function calculates the magnitude of the gradient of the sinogram with respect to angular and spatial perturbations, identifying regions where the beam is likely to intersect object edges.
Acquisition Function: A batch of $n_r$ $n_{r}$ new measurements is selected by maximizing an acquisition function $\Psi$ $Ψ$ . This function is the product of two terms:
- Uncertainty ( $\hat{\Omega}_k$ ): The width of the confidence interval derived from the GP posterior covariance, promoting exploration in regions with sparse data.
- Edge-Alignment ( $\hat{E}$ ): The edge-mapping term, promoting exploitation by prioritizing beams likely to intersect sample boundaries.
Greedy Batch Selection with Clustering Prevention: The algorithm exhaustively evaluates all candidate beams, sorts them by their acquisition values, and selects the top $n_r$ beams. To prevent the clustering of beams in high-value regions, a minimum distance threshold ( $\delta_\psi$ ) is applied as a post-processing step, removing beams that are too close to already selected ones.

Key Contributions

Reconstruction-Free Edge Detection: The primary contribution is the ability to identify edge-aligned measurements directly from the sinogram using a Gaussian Process, eliminating the computational cost and error propagation associated with iterative image reconstruction.
Joint Optimization: The framework simultaneously selects both projection angles and scanning positions, offering a more flexible adaptive strategy than methods that only optimize angles for raster scans.
Noise Robustness: By modeling the sinogram as a GP, the method inherently acts as a filter for measurement noise, utilizing the predictive mean to improve reconstruction accuracy even in noisy conditions.

Results
The proposed framework was evaluated on five representative samples (including Shepp–Logan phantoms, biological samples, and Siemens Star patterns) with a resolution of $128 \times 128$ pixels. The performance was measured using Root Mean Square Error (RMSE) against ground truth, comparing the adaptive method against a baseline of evenly spaced raster scan angles.

Reconstruction Quality: The adaptive method achieved lower RMSE than the baseline across all samples, particularly in the early iterations with fewer measurement beams. The selected beams were observed to concentrate along the edges of the objects without prior knowledge of the sample structure.
Noise Tolerance: Experiments with added measurement noise (up to 5% of the maximum signal) demonstrated that the proposed approach maintained higher reconstruction accuracy compared to the baseline, validating the noise-filtering capability of the GP model.
Beam Distribution: The inclusion of the distance threshold successfully prevented beam clustering, ensuring a more uniform distribution of informative measurements.

Significance and Claims
The paper claims that this approach offers a computationally efficient alternative to Bayesian sequential design methods that rely on full image reconstruction. By operating directly in the sinogram domain, the method reduces the susceptibility to reconstruction errors and lowers the computational overhead per iteration. While the authors acknowledge that independent beam positioning (scanning probe tomography) introduces experimental overhead not currently standard in hardware, they assert that the simulation results demonstrate the potential benefits of such adaptive strategies. The work serves as a motivation for future hardware implementations of adaptive scanning and establishes a framework for dose-efficient, high-quality tomographic reconstruction.