Multi-Iteration Fly-Scan Framework

• The paper introduces a feedback-driven x-ray microscopy technique that recomputes scan paths iteratively to focus on regions of interest.
• It employs repeated cycles of anchor-point selection, gradient-based optimization with ADAM, and nearest-neighbor path planning to significantly improve reconstruction quality.
• The method balances exploration and exploitation by combining continuous fly-scan acquisition with inverse-distance interpolation, reducing scan time and x-ray exposure while boosting PSNR.

The multi-iteration fly-scan framework is an adaptive acquisition strategy for x-ray microscopy in which a continuous fly-scan path is recomputed over multiple iterations so that measurements concentrate on regions of interest and unscanned regions are reconstructed by image completion. In the formulation introduced in "Optimizing Paths for Adaptive Fly-Scan Microscopy: An Extended Version" (Lu et al., 2 Sep 2025), each experiment proceeds through repeated cycles of anchor-point selection, anchor-point optimization, path generation, and inverse-distance-weighted completion, with the stated aims of shortening the scanning process and potentially decreasing x-ray exposure dose while maintaining high-quality and detailed information in critical regions.

1. Problem setting and conceptual rationale

Traditional raster-scanning techniques in x-ray microscopy acquire a microscopic image as a series of step-scans, whereas a fly-scan moves the x-ray probe along a continuous path. The latter reduces scan time and increases scan efficiency, but the paper emphasizes that not all regions of an image are equally important. Currently used fly-scan methods, as described there, do not adapt to the characteristics of the sample during the scan and therefore often waste time in uniform, uninteresting regions (Lu et al., 2 Sep 2025).

The same source contrasts this setting with a step-scan alternative in which deep learning techniques are used to select a shorter optimal scan path and reconstruct the entire image from partially scanned data. That alternative is described as heavily dependent on the quality of the initial sampling, requiring a large dataset for training, and incurring high computational costs. The multi-iteration fly-scan framework is positioned as a different response: it retains continuous-path acquisition, but redirects the path toward regions of interest and reconstructs non-scanned areas by image completion rather than by a training-intensive end-to-end model.

A common misconception is to equate adaptive fly-scanning with uniform global path shortening alone. In this framework, adaptation is sample-dependent and iteration-dependent: the current reconstruction influences where the next scan will go. This suggests a shift from fixed coverage toward progressive, feedback-driven allocation of the sampling budget.

2. Iterative acquisition pipeline

The framework is organized as a $K$-iteration pipeline. Iteration $k$, starting at $k=0$, contains four steps. First, given the current reconstruction $f^k$, with $f^0$ obtained from a short random initial scan, the method computes $N$ candidate anchor points $\Omega_k^0$ that are likely to lie in regions of interest. Second, starting from $\Omega_k^0$, it runs $S$ steps of a gradient-based optimizer, specifically ADAM, to minimize a loss $\mathcal L(f^k,\Omega_k)$, producing optimized anchor points $k$0. Third, it treats $k$1 as nodes in a graph with Euclidean edge lengths and applies a nearest-neighbor heuristic to approximate the shortest closed path $k$2, along which the fly-scan is performed. Fourth, it fills in all unscanned pixels by inverse-distance-weighted interpolation, yielding the updated reconstruction $k$3 (Lu et al., 2 Sep 2025).

After $k$4 iterations, or once a stopping criterion is met, the final output is

$k$5

The paper also notes an alternative stopping strategy based on anchor stability: one could stop early if $k$6 falls below a threshold. In the reported experiments, however, the iteration count $k$7 was fixed.

This pipeline combines exploration and exploitation in an explicit cycle. The score function proposes where to look, the objective function refines those proposals, the fly-scan collects measurements along a continuous trajectory, and the completion stage prepares the state from which the next cycle begins.

3. Formal components: score function and objective function

The score function $k$8 operates on the current reconstruction $k$9. Its starting point is the discrete gradient magnitude at pixel $k=0$0: $k=0$1 A discrete probability is then assigned to each pixel in proportion to that gradient: $k=0$2 The initial anchor set is defined by

$k=0$3

which simply draws the top-$k=0$4 pixels according to $k=0$5 (Lu et al., 2 Sep 2025).

The subsequent optimization stage defines a candidate anchor set $k=0$6 and minimizes

$k=0$7

The first component is an uncertainty term, denoted EWUF: $k=0$8 Here,

$k=0$9

is the normalized squared distance from anchor $f^k$0 to its $f^k$1-th nearest previously optimized point $f^k$2, and $f^k$3 is the softmax over these distances. The constants $f^k$4 and $f^k$5 are the noise standard deviation and length-scale. The second component is the gradient-magnitude term,

$f^k$6

evaluated at the anchors and encouraging them to lie in regions of large image gradient.

Optimization is performed through $f^k$7 iterations of

$f^k$8

implemented with ADAM and step-size $f^k$9, producing $f^0$0 (Lu et al., 2 Sep 2025). The paper explicitly states that the loss balances uncertainty, used to explore under-sampled areas, and image gradient magnitude, used to exploit high-detail regions.

4. Path planning, continuous acquisition, and image completion

Once the optimized anchors $f^0$1 have been obtained, the framework formulates path construction as an approximation to the shortest closed tour through those anchors. Although one could solve the full Traveling-Salesman Problem to obtain the exact shortest closed path, the method instead uses a greedy nearest-neighbor rule. Starting from an arbitrary $f^0$2, the next point is chosen by

$f^0$3

This process continues until all $f^0$4 anchors are visited, yielding $f^0$5. The paper states that the resulting path length is within a small factor of the TSP optimum, at far lower computation cost (Lu et al., 2 Sep 2025).

The actual acquisition stage is a continuous fly-scan along $f^0$6, producing true intensity values $f^0$7. Reconstruction of unscanned pixels then proceeds by inverse-distance weighting. For an unscanned pixel $f^0$8,

$f^0$9

In practice, only the nearest $N$0 neighbors of each query point are collected in order to keep the cost linear in $N$1 per pixel.

Two clarifications are important. First, the framework does not claim exact global route optimality; it uses a nearest-neighbor heuristic for computational efficiency. Second, the completion stage is not a learned image generator in the reported implementation; it is an analytical IDW interpolation scheme. The paper also remarks that the nearest-neighbor heuristic can produce sharp turns and that a future extension could add a curvature penalty to $N$2 to enforce smoother motor trajectories (Lu et al., 2 Sep 2025).

5. Quantitative behavior, hyperparameters, and operating trade-offs

The reported benchmarks cover four images: synthetic shapes $N$3, Shepp-Logan phantom $N$4, Cameraman $N$5, and a WSe$N$6 nanoflake $N$7. All tests used less than $N$8 of full raster-scan samples, and the paper summarizes average results over 10 runs for three conditions: an initial random $N$9 scan, a random-anchor baseline with the same total number of points, and the proposed method (Lu et al., 2 Sep 2025).

Image	Sampling %	Initial / Random / Proposed PSNR-SSIM
Synthetic	21.4%	7.3/0.62, 15.1/0.74, 23.7/0.91
Phantom	16.2%	12.8/0.70, 17.6/0.82, 22.4/0.88
Cameraman	22.4%	18.6/0.63, 24.9/0.71, 30.1/0.87
WSe$\Omega_k^0$0 Flake	23.5%	16.9/0.52, 21.3/0.69, 28.3/0.85

Qualitatively, the adaptive paths are described as concentrating on edges and fine features, whereas raster paths waste effort in featureless regions. The paper further states that the final PSNR gains over random selection exceed 6–8 dB. Repeated runs over 10 trials showed less than 1 dB standard deviation in PSNR.

The iteration schedule and hyperparameters were fixed in the experiments. For $\Omega_k^0$1 images, $\Omega_k^0$2; for the $\Omega_k^0$3 image, $\Omega_k^0$4. This ensured that the total sampled points remained below approximately $\Omega_k^0$5 of a full raster scan. The anchor count was $\Omega_k^0$6 per iteration, except $\Omega_k^0$7 for the WSe$\Omega_k^0$8 image. Each iteration used $\Omega_k^0$9 ADAM steps with learning rate $\Omega_k^0$0. The uncertainty weight was set to $\Omega_k^0$1, with parameter analysis reported as stable for $\Omega_k^0$2. The length-scale was set to $\Omega_k^0$3, and $\Omega_k^0$4 yielded less than $\Omega_k^0$5 dB PSNR variation. Experiments were run on a MacBook Air M3 with simulated fly-scan speed $\Omega_k^0$6, exposure $\Omega_k^0$7, and dead time $\Omega_k^0$8.

These results define the principal trade-off strategy of the framework. By choosing $\Omega_k^0$9 and $S$0, the practitioner directly trades total scan time and x-ray dose, proportional to the total sampled points, against reconstruction quality as measured by PSNR and SSIM.

6. Relation to broader fly-scan systems and adjacent reconstruction frameworks

Within the supplied literature, the phrase "multi-iteration fly-scan framework" refers to related but non-identical levels of the fly-scan stack. In the adaptive microscopy setting of (Lu et al., 2 Sep 2025), the emphasis is on where to scan next and how to reconstruct unmeasured regions. In "Progress and outlook on advanced fly scans based on Mamba" (Li et al., 2023), the emphasis is on how repeated fly-scan fragments are orchestrated at beamline scale through motors/mechanics, a sequencer such as PandABox or the software-only fallback Bubo, detectors/data processing, double-buffering, and online tuning. In "Adorym: A multi-platform generic x-ray image reconstruction framework based on automatic differentiation" (Du et al., 2020), the emphasis is on iterative reconstruction for fly-scan ptychographic tomography through a general forward model, automatic differentiation, refinable probe positions, and distributed optimization.

The Mamba-based work describes each fly-scan iteration as a collaboration of three loosely coupled subsystems and gives a generic online-tuning loop in which parameters are adjusted between fragments after online analysis. It also provides concrete throughput and latency budgets, including an architectural discussion of $S$1 kHz fly scans, the distinction between PandABox and Bubo in jitter performance, and the need for continuous-drive motors on the flying axes at those rates (Li et al., 2023). A plausible implication is that the adaptive path-planning strategy of (Lu et al., 2 Sep 2025) could be embedded into such a fragment-wise control architecture, although that integration is not claimed in the supplied text.

Adorym, by contrast, models continuous-motion fly-scan blur through multiple incoherent probe submodes displaced along the motion trajectory during an exposure, and optimizes object, probe, and position corrections through automatic differentiation. Its reported use cases include 2D fly-scan ptychography, sparse multislice ptychography, and joint ptychotomography, with both data-parallel and distributed-object modes (Du et al., 2020). This places the adaptive anchor-point method of (Lu et al., 2 Sep 2025) in a different methodological category: it is a path-selection and completion framework rather than a general physics-based inverse solver.

Taken together, these works show that multi-iteration fly-scan methodology spans at least three layers: adaptive sampling policy, beamline execution and online retuning, and iterative physical reconstruction. The term therefore should not be treated as denoting a single canonical algorithm. In the specific sense established by (Lu et al., 2 Sep 2025), however, it denotes a feedback-driven procedure in which analytical scoring, differentiable anchor optimization, approximate route planning, and IDW completion are repeated until a stopping condition or fixed iteration budget is reached.