Adaptive Scan Strategies: Methods & Applications

• Adaptive scan strategies are advanced methods that iteratively determine optimal data acquisition locations using algorithmic, statistical, and deep learning approaches.
• They enhance performance by maximizing task-specific utility—such as improved image fidelity and localization—while reducing resource consumption in complex, high-dimensional settings.
• Applications in MRI, STEM, robotics, and Bayesian inference showcase improvements in efficiency, accuracy, and computational cost compared to traditional static scanning methods.

Adaptive scan strategies refer to methodologies—algorithmic, statistical, or deep learning-based—that iteratively and adaptively determine where to sample, sense, or acquire data in a domain (spatial, temporal, frequency, or combinatorial) in response to ongoing measurements, uncertainty, or task requirements. These strategies aim to maximize task-relevant utility (e.g., information gain, image fidelity, localization power) while minimizing required resources (e.g., dose, scan time, sample count, computation), making adaptivity essential in high-dimensional, underdetermined, or resource-constrained regimes. Adaptive scan strategies underpin advances across computational imaging, Bayesian inference, robotics, surveying, and more, encompassing reinforcement learning for sequential scanning, multiscale hypothesis testing, scan-adaptive architectures in deep vision models, and real-time sampling optimization in clinical imaging.

1. Mathematical Formulations and Theoretical Guarantees

Adaptive scan strategies are often formalized using sequential decision, optimization, or statistical inference frameworks.

• Markov Decision Processes (MDP): In transmission electron microscopy (STEM), adaptive scanning is modeled as a finite-horizon MDP with the state comprising the mask of already-sampled pixels and their intensities; actions represent the selection of the next scan segment and direction, and transitions deterministically yield new observations. The reward at each step is the marginal reduction in mean squared error (MSE) or structural similarity index (SSIM) upon hypothetical immediate reconstruction, explicitly: $$R_t = PSNR(\hat I_t, I^*) - PSNR(\hat I_{t-1}, I^*)$$ where $\hat I_t=Q_\phi(Y_t, M_t)$ is the current CNN reconstruction (Ede, 2020).
• Multiscale Scan Statistics: For submatrix localization in data mining, the adaptive scan statistic maximizes a size-normalized sum over all matrix blocks $(I,J)$, subtracting a combinatorial penalty that depends on block size:

$$S(X) = \max_{I, J} \left\{ \frac{1}{\sqrt{|I||J|}} \sum_{(i, j) \in I \times J} X_{ij} - \frac{\lambda_{|I|,|J|}}{\sqrt{|I||J|}} \right\}$$

with

$$\lambda_{m,n}=\sqrt{2\log(MN\binom{M}{m}\binom{N}{n})}$$

providing oracle-tight recovery thresholds for the signal (Liu et al., 2019).

• Information-theoretic and Bayesian Criteria: In free-space optical acquisition, expected time to detection or coverage is analyzed using mixture models and geometric sums, supporting provable acquisition-time improvements for region-prioritized adaptive spirals or randomized "shotgun" scans (Bashir et al., 2020). Bayesian range-finder models employ mixture likelihoods and variational inference for environment-driven adaptivity (Laet et al., 2014).

Adaptive scan strategies come with explicit theoretical results:

• Provable recovery thresholds matching minimax-optimal rates up to constants (as in adaptive scan statistics for submatrix recovery) (Liu et al., 2019).
• Convergence and semi-global practical asymptotic stability (SGPAS) of control-theoretic adaptive scan rate adjustment in atomic force microscopy (Wang et al., 2019).
• Integrated autocorrelation time and time-optimality criteria in adaptive scan Gibbs sampling (Smolyakov et al., 2018).

2. Algorithmic Realizations and Architectures

Adaptive scanning is algorithmically realized via a range of methods:

• Reinforcement Learning Scan Policies: In STEM, a recurrent deterministic policy gradient network parameterizes scan segment selection, observing a mask plus partial intensity image, using a 3-layer CNN encoder fused with 2-layer GRUs to select actions maximizing reward (PSNR/SSIM gains). Reconstruction networks (U-Net style) are trained jointly (Ede, 2020).
• Iterative Coordinate Descent (ICD) and Greedy Selection: In MRI scan-adaptive undersampling (e.g., SUNO, dSUNO, neighborhood clustering), ICD is used to optimize a scan-specific Cartesian mask: at each step, a sampled k-space line is swapped for another if it reduces reconstruction loss (via a co-trained deep reconstructor). At inference, nearest-neighbor selection identifies the most appropriate optimized mask given low-frequency data (Gautam et al., 16 Jan 2025, Gautam et al., 15 Feb 2026, Gautam et al., 2023).
• Alternating Optimization and Mask Prediction Networks: Neural architectures such as the MNet sampling network generate scan-adaptive masks from input navigators, supervised by ICD-generated masks and trained jointly with the reconstructor in an alternated regime (Dhar et al., 21 Sep 2025). Deep scan-adaptive networks have improved MR image fidelity versus variable-density random or population-trained masks.
• Statistical Hill-Climbing and Efficient Search: Adaptive scan statistics for anomaly localization employ multistage hill-climbing and golden-section search across submatrix sizes for computationally viable approaches to NP-hard global scan optimization (Liu et al., 2019).
• Graph-based Perception in Robotics and Surveying: Adaptive region-based scan selection in robotics leverages geometric predicates—frontier and bridging scans—composed in star-convex submaps supporting efficient, sequentially composed controllers with theoretical safety and completeness guarantees (Latha et al., 2024). In underwater surveying, view-graphs and graph neural network architectures, combined with deep Q-learning, select informative next scan-angles for multi-view active perception (Sethuraman et al., 2024).
• Vision SSMs with Learnable Scan Orders: Adaptive scan order is integrated directly into modern state-space model (SSM) vision backbones (DAMamba). Feature map patches are offset via a CNN (OPN), producing per-image scan/adjacency trajectories; adaptive bilinear sampling aligns sampling order to salient content while maintaining global receptive fields (Li et al., 18 Feb 2025). In ASM-UNet, per-volume scan scores blend group-level priors and instance-level predictors to dynamically reorder spatial tokens fed to SSM/Mamba modules, enabling fine-grained, structure-aware medical segmentation (Wang et al., 10 Aug 2025).

3. Application Domains and Representative Results

Adaptive scan strategies have demonstrated impact across diverse modalities:

Area	Adaptive Scan Function	Quantitative Gain
STEM imaging	RL-driven boundary-hunting partial scans	+2.3 dB PSNR vs. best static at 5% dose
MRI undersampling	Mask ICD/sampler + reconstructor alternation, NN mask selection	+2-3 dB PSNR, +0.03–0.06 SSIM over VDRS
X-ray fly-scan microscopy	Adaptive path via anchor optimization, iterative scan-completion loop	Typical dose savings >30%, PSNR/SSIM boost
Atomic force microscopy	Extremum-seeking scan-rate controller, online metric adaptation	50% reduction in edge Z-error vs. fixed
Robotics (map, plan, control)	Scan-graph with frontier/bridge regions, region sequencing	+25% coverage efficiency, provable safety
Large-scale Bayesian inference	Adaptive batch-size scan Gibbs sampling	15–40% lower MSE/perplexity vs. collapsed
Vision SSM architectures	Dynamically learned patch scan order (DAS, ASM)	+1.2% ImageNet accuracy (DAMamba), +0.7% Dice in FGS segmentation
Side-scan sonar AUV surveying	Angular view-graph, GNN-based NBV policy	0.96 accuracy, highest classification efficiency

4. Advantages, Limitations, and Design Guidelines

Advantages:

• Resource efficiency: Adaptive scans concentrate measurements/dose in information-rich or high-impact regions, sharply exceeding static patterns for a fixed budget (Ede, 2020, Gautam et al., 16 Jan 2025, Gautam et al., 15 Feb 2026).
• Specimen/task awareness: Sequential adaptation enables sample-driven focus (edges, defects), and integration of contextual or historical scan data (robotic star-convex regions, learned scan order) (Latha et al., 2024, Li et al., 18 Feb 2025).
• Joint learning: End-to-end alternation (scan mask + reconstructor), or joint tuning of policy and completion, enables optimal trade-offs in observed metrics (PSNR, structural similarity, localization error) (Dhar et al., 21 Sep 2025, Sugathan et al., 2015).

Limitations:

• Discrete or restricted scan segmentations may introduce overhead or miss highly oblique or complex features; vectorized/continuous action spaces are active research (Ede, 2020).
• Initial random sparse scans or poor mask initialization may lead to unrecoverable missing detail; joint or online retraining mitigates this (Dhar et al., 21 Sep 2025, Gautam et al., 16 Jan 2025).
• Some strategies require significant offline optimization (mask database, anchor path planning) and storage for large-scale dictionaries or lookup tables (Gautam et al., 15 Feb 2026, Gautam et al., 2023).
• Physical implementation constraints: Adaptive spiral search in FSO requires higher pointing accuracy than the shotgun approach, which is more robust (Bashir et al., 2020).

Design Guidelines:

• Initialization and candidate set size should match the scale and noise properties of the system. Orientation and uncertainty-driven sampling probabilities are more robust than uniform or random.
• In iterative frameworks, key parameters (e.g., mask size, dither amplitude, scan step size, regularization weights) should be selected to balance computation per step and global convergence (Wang et al., 2019, Dhar et al., 21 Sep 2025, Gautam et al., 16 Jan 2025).
• For vision transformer-analogous SSMs, adaptive scan order should blend local and global priorities (group and instance components, bilinear sampling) to retain both receptive field and anatomical fidelity (Wang et al., 10 Aug 2025, Li et al., 18 Feb 2025).

5. Practical Implementations and Impact

• Data-driven MRI: Clinical MRI protocols incorporating scan-adaptive strategies achieve higher-fidelity images at fixed scan time, lower patient exposure, and sharper diagnostic features. Nearest-neighbor mask selection from pre-optimized libraries (SUNO, dSUNO) is already demonstrated on large fastMRI datasets (Gautam et al., 16 Jan 2025, Gautam et al., 15 Feb 2026).
• STEM and Microscopy: RL-based scanning agents for STEM directly exploit specimen boundaries, yielding sharply improved performance for low-dose or high-throughput acquisition (Ede, 2020). Iterative anchor optimization and path planning in x-ray fly-scan minimizes unnecessary exposure (Lu et al., 2 Sep 2025).
• Robotics and Autonomous Sensing: Region-based scan selection algorithms guarantee finite-time, safe, and efficient environment exploration, with application in autonomous delivery, inspection, and mapping (Latha et al., 2024). In AUV sonar, GNNs and active perception policies enable informed, time-optimal, multi-view reconnaissance (Sethuraman et al., 2024).
• Statistical Inference: Adaptive batch schemes for large-scale MCMC (e.g., mini-batch scan Gibbs) achieve improved posterior mixing and faster convergence in big-data Bayesian inference (Smolyakov et al., 2018).
• Computer Vision: State space model (SSM) backbones with dynamic adaptive scan orders outperform fixed-scan architectures and match or surpass leading CNN/ViT models on classification, detection, and segmentation (Li et al., 18 Feb 2025).

Overall, adaptive scan strategies unify principles from control, decision theory, combinatorial optimization, and deep learning to enable measurement processes that optimize task performance in complex, high-dimensional, or resource-constrained environments.

6. Selected References and Further Reading

• Adaptive RL-driven STEM: (Ede, 2020)
• Scan-adaptive MRI (SUNO, dSUNO, MNet): (Gautam et al., 16 Jan 2025, Gautam et al., 15 Feb 2026, Dhar et al., 21 Sep 2025, Gautam et al., 2023)
• Adaptive scan statistical inference: (Liu et al., 2019, Smolyakov et al., 2018)
• Atomic force microscopy adaptive scanning: (Wang et al., 2019)
• Vision SSM adaptive scan: (Li et al., 18 Feb 2025, Wang et al., 10 Aug 2025)
• Region-based scan and control in mobile robotics: (Latha et al., 2024)
• Adaptive acquisition in FSO: (Bashir et al., 2020)
• Active perception with side scan sonar: (Sethuraman et al., 2024)
• Variable-pixel adaptive image scan: (Sugathan et al., 2015)

These works provide detailed mathematical frameworks, algorithms, empirical benchmarks, and practical guidelines for designing adaptive scan strategies tailored to the modality and application domain.