Ultra-Sparse Sampling (USS)
- Ultra-Sparse Sampling (USS) is a regime that acquires only a tiny fraction of measurements and recovers missing data using structural priors, inverse methods, or learned models.
- It is applied in diverse fields such as ultrasound, photoacoustic tomography, CT, matrix completion, and video SCI to reduce scan time, radiation, and hardware complexity.
- Reconstruction methods in USS combine explicit priors and neural representations, showing rapid performance gains up to modality-specific sparsity thresholds before reaching diminishing returns.
Ultra-Sparse Sampling (USS) denotes acquisition regimes in which only a very small fraction of the measurements of a conventional dense system are collected, and the missing information is recovered through structural priors, inverse methods, or learned models. In the cited literature, USS appears as partial entry sampling of pre-beamformed ultrasound data, random receive-channel deactivation in ultrasound localization microscopy, spiral angle-wavelength interlacing in multispectral photoacoustic tomography, single-view to few-view cone-beam CT, one-hot temporal masks in video snapshot compressive imaging, and entrywise observation with probability in one-sided matrix completion (Zhang et al., 2018, Hardy et al., 2023, Zhong et al., 2024, Shen et al., 2021, Tan et al., 2024, Wang et al., 8 Feb 2026, Cao et al., 10 Sep 2025, Zhang et al., 18 Jan 2026, Liu et al., 24 Jun 2026). Across these formulations, the immediate objectives are to reduce data transfer, storage, hardware complexity, scan time, or radiation exposure while preserving the downstream task, such as B-mode reconstruction, microbubble localization, spectral unmixing, volumetric tomography, subspace recovery, or radio-map estimation.
1. Defining regimes and application domains
The literature does not use a single numeric threshold for USS. Instead, “ultra-sparse” is defined relative to the dense reference acquisition of each modality. In ultrasound signal reconstruction, sampling rates of or lower are studied; in multispectral PAT, the reported overall rate is approximately $1/30$; in CBCT, ultra-sparse-view settings include $1$, $2$, or $3$ views out of about $720$, as well as $23$ or $50$ views; in matrix completion, each entry is observed independently with probability ; in video SCI, each spatial location is assigned to exactly one sub-frame (Zhang et al., 2018, Zhong et al., 2024, Shen et al., 2021, Wang et al., 8 Feb 2026, Zhang et al., 18 Jan 2026, Cao et al., 10 Sep 2025).
| Domain | USS regime | Primary reconstruction target |
|---|---|---|
| Ultrasound / ULM | Partial RF entries, reduced receive channels | B-mode approximation, microbubble localization |
| PAT / CT / CBCT | Spiral interlacing, few-view angular sampling | Tomographic image or volume reconstruction |
| Matrix / radio map / video SCI | 0, ultra-low sensor masks, one-hot masks | Second moment 1, radio map, high-speed frames |
In ultrasound, the full pre-beamformed channel data matrix is written as 2, with measurements 3, where 4 keeps only sampled entries (Zhang et al., 2018). In SPARSE-ULM, full-array RF data 5 are reduced through a binary channel-selection matrix 6, yielding 7 (Hardy et al., 2023). In video SCI, the forward model is 8, but USS constrains the mask sequence so that 9 for every spatial location (Cao et al., 10 Sep 2025). In matrix completion, the regime is one-sided: exact recovery of the full matrix may be impossible, yet recovery of $1/30$0 remains feasible (Zhang et al., 18 Jan 2026). This suggests that USS is better understood as a family of acquisition constraints than as a single sampling doctrine.
2. Measurement models and structural assumptions
A common feature of USS formulations is that aggressive acquisition reduction is paired with a strong signal model. In ultrasound RF reconstruction, the signal is modeled as both low-rank and joint-sparse in a transform domain. With a partial Fourier matrix $1/30$1 and coefficient matrix $1/30$2, the factorization $1/30$3 is used, and reconstruction solves
$1/30$4
or equivalently the reduced problem in $1/30$5 (Zhang et al., 2018). The physical justification is explicit: adjacent channels are highly correlated, and ultrasound RF signals are bandlimited, so inter-channel redundancy and shared active frequencies can be exploited.
In SPARSE-ULM, the structural prior is not an explicit low-rank model of the RF data but a conventional ULM processing chain applied after channel subsampling. The framework assumes that SVD clutter filtering, Delay-and-Sum beamforming, PSF correlation, and sub-pixel Gaussian fitting remain effective even when most receive elements are inactive. Localization is obtained by solving
$1/30$6
on a local correlation map patch (Hardy et al., 2023).
Tomographic USS methods typically replace explicit sparse priors with geometry-aware or neural implicit representations. In U3S-PAT, the unknown image is embedded in a periodic-activation MLP $1/30$7, trained first on a fused prior image and then fine-tuned against ultra-sparse measurements using a self-supervised loss that couples neighboring spiral positions and wavelengths (Zhong et al., 2024). In GIIR for 3D CBCT, the forward model is $1/30$8, but known source-detector geometry is injected through a differentiable back-projection operator $1/30$9, separating projection synthesis from volumetric refinement (Shen et al., 2021). In CSDN, a Neural Attenuation Field $1$0 maps spatial coordinates to attenuation coefficients and synthesizes dense projections from ultra-sparse rays before diffusion refinement (Wang et al., 8 Feb 2026).
Outside biomedical imaging, the structural target can change substantially. In one-sided matrix completion, the estimand is
$1$1
rather than $1$2 itself, because when each row has only $1$3 entries and $1$4, accurate imputation of $1$5 is impossible (Zhang et al., 18 Jan 2026). In diffusion-based radio-map estimation, the observed matrix is $1$6, with sampling rate $1$7, and the problem is framed as non-linear matrix completion with side information (Liu et al., 24 Jun 2026).
3. Sampling-pattern design
USS sampling patterns range from random to task-learned to geometry-coupled. In SPARSE-ULM, each steering angle is associated with an independent random draw of $1$8 active elements, held constant over a buffer of $1$9 frames to allow SVD-based clutter filtering. A Bernoulli model with activation probability $2$0 is used, and although non-uniform laws favoring central or peripheral elements were tested, the uniform scheme (“Uni law”) was found to have only marginally different behavior while remaining simplest and nearly optimal (Hardy et al., 2023).
DPS generalizes USS pattern design into a learnable selection process. It introduces $2$1 categorical distributions over $2$2 candidate samples, parameterized by logits $2$3, and samples a binary mask $2$4 without replacement using the Gumbel-max trick and a straight-through Gumbel-Softmax estimator. The joint objective couples sampling design and downstream task performance. Once trained, the resulting sub-sampling patterns are fixed and directly implementable by non-uniform analog-to-digital conversion, sparse array design, or slow-time ultrasound pulsing schemes (Huijben et al., 2019).
Geometry-coupled USS appears prominently in PAT and CT. U3S-PAT uses a sparse ring-shaped transducer that rotates by $2$5 and translates by $2$6 whenever the laser switches wavelength, producing a discrete spiral trajectory
$2$7
with multispectral angle interlacing and effective data reduction $2$8 relative to dense $2$9 sampling (Zhong et al., 2024). MSDiff, by contrast, employs an equidistant angular mask so that selected views are as uniformly spaced as possible over the full angle range (Tan et al., 2024). In CBCT, ultra-sparse-view protocols are defined directly by the number of views, such as $3$0, $3$1, or $3$2 out of about $3$3, or $3$4 and $3$5 uniformly spaced angles (Shen et al., 2021, Wang et al., 8 Feb 2026).
Video SCI imposes a stricter combinatorial constraint. Under USS, for every pixel $3$6, exactly one of the $3$7 masks is $3$8 and all others are $3$9, so each pixel’s measurement is contributed by exactly one frame (Cao et al., 10 Sep 2025). In one-sided matrix completion, the sampling law is entrywise and independent: each $720$0 is observed with probability $720$1, which leads to approximately $720$2 total observations and on average $720$3 samples per row (Zhang et al., 18 Jan 2026). The surveyed literature therefore spans random Bernoulli selection, differentiable mask learning, equidistant view design, spiral interlacing, and strict one-hot masking, rather than a single canonical pattern.
4. Reconstruction methodologies
Reconstruction under USS is dominated by methods that compensate for severe ill-posedness through explicit priors or learned inductive bias. In low-rank and joint-sparse ultrasound reconstruction, the solver is a Simultaneous Direction Method of Multipliers (SDMM) operating on the coefficient matrix $720$4. The algorithm alternates a quadratic $720$5-update, singular-value soft-thresholding for the nuclear norm, row-wise $720$6 soft-thresholding for the $720$7 term, and data-consistency enforcement on observed entries (Zhang et al., 2018).
SPARSE-ULM retains a largely conventional signal-processing workflow after acquisition sparsification. Each block of compound frames undergoes SVD clutter filtering on slow-time data, Delay-and-Sum beamforming on the reduced $720$8 RF channels onto a 2D grid, correlation with a pre-computed single-microbubble PSF, and sub-pixel localization by 2D Gaussian fitting around thresholded peaks (Hardy et al., 2023). The point is not to alter the localization model fundamentally, but to test how far receive-channel sparsification can be pushed before detection degrades.
U3S-PAT uses a self-supervised image reconstruction strategy tailored to the spiral scan. A 4-layer, width-512, SIREN-style MLP first fits a dense-coverage prior image fused from neighboring angles and wavelengths, then fine-tunes against the ultra-sparse raw measurements using a joint loss that includes data fidelity at the current slice and weighted penalties from neighboring slices and wavelengths, with $720$9 (Zhong et al., 2024). GIIR also decouples the problem: a 2D network synthesizes missing projections, a non-learned geometric back-projection preserves exact system geometry, and a Y-shaped 3D U-Net variant refines the resulting geometry-preserving images (Shen et al., 2021).
Diffusion-based USS reconstruction has diversified into several architectures. MSDiff trains two score-based diffusion models in the projection domain: a Full-view Diffusion Model on complete sinograms and a Sparse-view Diffusion Model on masked sinograms. During inference, the method alternates sparse refinement, merge, global denoising, and a closed-form data-consistency projection (Tan et al., 2024). CSDN begins with a Neural Attenuation Field trained from sparse rays, synthesizes dense projections, decomposes them into sinogram and digital-radiography domains, applies residual diffusion in both pathways, and fuses the corresponding FDK reconstructions voxel-wise through the Dual-Projection Reconstruction Fusion module (Wang et al., 8 Feb 2026).
In video SCI, the mismatch between DMD and CCD breaks the ideal decomposition of a USS measurement into independent sub-measurements. BSTFormer addresses this by building a sparse-aware transformer with Local Block Attention, Global Sparse Attention, and Global Temporal Attention, applied after a mask-normalized initialization based on $23$0 (Cao et al., 10 Sep 2025). In one-sided matrix completion, recovery of $23$1 proceeds through a Hajek-style unbiased estimator on observed co-occurrences, followed by nonconvex gradient descent on a factorization $23$2 with an incoherence-encouraging regularizer (Zhang et al., 18 Jan 2026).
5. Quantitative trade-offs and reported performance
Ultrasound studies emphasize that USS can preserve reconstruction quality surprisingly far into the sparse regime, but not without task-dependent degradation. In low-rank and joint-sparse ultrasound reconstruction, results on an in-vivo cardiac RF volume show that at $23$3 there are severe artifacts and CNR drops by approximately $23$4, whereas at $23$5 and above images are visually almost indistinguishable from the reference and CNR is nearly recovered; pure sparsity-only methods are reported as limited to $23$6 (Zhang et al., 2018). In SPARSE-ULM simulation over $23$7 frames, the False Positive Rate rises from $23$8 for $23$9 channels in receive and $50$0 steered angles to $50$1 for $50$2 channels and $50$3 angles, while the average localization accuracy changes from approximately $50$4 for $50$5 channels and $50$6 angles to approximately $50$7 for $50$8 channels and $50$9 angles; reducing 0 channels cuts data by 1 (Hardy et al., 2023). DPS reports that for channel sub-sampling, learned sampling plus a task network outperforms uniform sampling plus a network by up to 2 lower MSE and is on par with hand-designed sum-coarray arrays when half the channels are used (Huijben et al., 2019).
PAT and tomographic USS papers report similarly strong but thresholded trade-offs. U3S-PAT states that with 3, 4, and 5, the overall rate is approximately 6, and that even at this rate the method achieves similar reconstruction and spectral unmixing accuracy as non-spiral dense sampling; in the reported spectral-unmixing experiment for 7, full U3S-PAT attains 8 PSNR, 9 SSIM, 00 Hb-Dice, and 01 HbO02-Dice, while RMSE curves show rapid PSNR/SSIM gain until 03, beyond which returns diminish (Zhong et al., 2024). In GIIR, averaged over 04 test cases, the single-view setting gives NRMSE 05, SSIM 06, and PSNR 07, the two-view setting gives NRMSE 08, SSIM 09, and PSNR 10, and the three-view setting gives NRMSE 11, SSIM 12, and PSNR 13 (Shen et al., 2021). MSDiff reports, on AAPM data with 14 views, PSNR 15 and SSIM 16, outperforming FBP, U-Net, FBPConvNet, and GMSD; ablation shows that the combined FDM+SDM model improves over either component alone (Tan et al., 2024). Under 17-view CBCT on the L067 phantom, CSDN reports 18 PSNR and 19 SSIM versus 20 and 21 for NAF, and under the more challenging 22-view condition it still exceeds the strongest alternative by over 23 in PSNR and 24 in SSIM (Wang et al., 8 Feb 2026).
In non-tomographic settings, the quantitative message changes from direct image fidelity to subspace or frame recovery. In one-sided matrix completion, the proposed method reduces bias by 25 relative to baseline estimators on three MovieLens datasets, reduces the recovery error of 26 by 27 and of 28 by 29 on an Amazon reviews dataset with sparsity 30, and yields a 31 reduction in 32-error in the two-observations-per-row regime relative to a nuclear-norm baseline (Zhang et al., 18 Jan 2026). In video SCI, BSTFormer reports average PSNR 33 and SSIM 34 on simulated data at compression ratio 35, compared with 36 and 37 for the previous best EfficientSCI-B; the same work reports that USS retains full 38 dynamic range per sub-frame, whereas RS integrates approximately 39 frames per pixel and yields per-frame range approximately 40 (Cao et al., 10 Sep 2025).
6. Theoretical limits, misconceptions, and open directions
A recurring misconception is that USS necessarily seeks exact recovery of the fully sampled signal. Several papers state the opposite. In one-sided matrix completion, when each row contains only 41 observed entries and 42, accurate imputation of the full matrix 43 is impossible; the tractable objective is instead recovery of the averaged second-moment matrix 44 or the underlying row span (Zhang et al., 18 Jan 2026). In SPARSE-ULM, reducing the number of active receive elements deteriorates signal-to-noise ratio and can lead to false microbubble detections, even though localization accuracy remains nearly invariant over the reported range (Hardy et al., 2023). In U3S-PAT, the discussion states that below 45 quality collapses, and that the current validation is virtual on a fixed-ring system, with a true rotating/translating prototype yet to be built (Zhong et al., 2024).
Theoretical work on radio-map estimation makes the dependence on sparsity explicit. The minimum achievable error of a diffusion model is lower-bounded by the discrepancy between the deployment distribution and the true underlying propagation law, and the actual error under ultra-low sampling rate satisfies a bound of the form
46
with a critical sampling-rate threshold
47
above which further increases in 48 yield only marginal MSE gains (Liu et al., 24 Jun 2026). This formalizes a point that is otherwise empirical in the imaging papers: USS performance often improves rapidly up to a modality-specific threshold and then exhibits diminishing returns.
Open directions are stated directly in several works. SPARSE-ULM proposes extension to 49D ULM with sparse-matrix arrays, incorporation of compressed-sensing or deep-learning priors for even smaller 50, and optimization of angle-to-element assignment to suppress side-lobes further (Hardy et al., 2023). GIIR suggests that geometric priors can be embedded in other modalities, including PET, MRI, photoacoustic imaging, and ultrasound tomography (Shen et al., 2021). U3S-PAT notes that adaptation to linear or spherical arrays requires geometry-specific forward models (Zhong et al., 2024). Video SCI identifies USS as a good choice for a complete system on chip because of fixed exposure time and mentions emerging global-shutter CMOS arrays as an implementation path (Cao et al., 10 Sep 2025). Taken together, the literature indicates that USS is not merely an aggressive downsampling heuristic; it is a design regime in which sensing, prior modeling, and task definition must be co-specified to remain well posed.