Pith. sign in

REVIEW 4 major objections 5 minor 22 references

Deep Sylvester Posterior Inference for Adaptive Compressed Sensing in Ultrasound Imaging

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Adaptive scan-line selection with a Sylvester normalizing-flow posterior beats fixed ultrasound sampling by 15% and runs at 66 Hz.

desk verdict A genuinely real-time adaptive sampling method with an honest limitations section, but the headline 15% claim does not survive contact with Table I. read the letter →

arxiv 2501.03825 v1 pith:25ALMGVU submitted 2025-01-07 eess.IV cs.AIcs.CV

Simon W. Penninga , Hans van Gorp , Ruud J.G. van Sloun This is my paper · ORCID
classification eess.IVcs.AIcs.CV
keywords activeinferenceadaptivecompressedsensingultrasoundimagingSylvesternormalizingflowsvariationalmutualinformationscan-linesubsamplingfreeenergy
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that ultrasound scanners need not use a fixed pattern of scan lines: they can choose where to fire next based on what the current, partially observed frame suggests about the next frame. To make this practical, it replaces slow posterior sampling with a Sylvester normalizing-flow encoder that turns partial observations into a flexible Bayesian posterior in one network evaluation. The sampling policy then maximizes the mutual information between candidate scan-line masks and the next frame's predicted observations, implemented either by log-determinant or trace of the predicted covariance. Evaluated on cardiac ultrasound videos, the trace policy improves mean absolute reconstruction error by 15% over uniform, variable-density, and equispaced baselines under aggressive subsampling, while the full acquisition-decision step runs at roughly 66 Hz. The gain is largest when very few scan lines are used, so the method is aimed at exactly the regime where frame rate and field of view are hardest to balance.

What carries the argument

The load-bearing object is the approximate posterior $q_\phi(z_t|y_t)$, built as a variational encoder whose Gaussian latent is warped by $K=8$ orthogonal Sylvester normalizing-flow layers; this gives an expressive, multi-modal posterior at the cost of a single neural network evaluation. The policy side is the predictive covariance $\Sigma_{y_{t+1}|A_{t+1},\hat y_t}$, estimated from $N_S=3$ posterior samples decoded by the frozen generative model. Because choosing among all $\binom{N}{M}$ masks is intractable, the policy restricts itself to a random candidate set of size $S=10{,}000$ and scores candidates by $\log\det$ (covariance sampling) or by the trace (trace sampling), with a rule that forbids selecting neighboring scan lines. That combination is what converts the mutual-information objective into a real-time loop.

What would settle it

Rerun the trace-sampling policy on EchoNet videos whose frame intervals are doubled, making the heart move farther between frames, and compare L1 error to equispaced sampling; if the 15% advantage shrinks or reverses, the identity-transition assumption is carrying the reported gain.

Watch

Extended reading notes

Core claim

The central claim is that adaptive, information-driven scan-line selection in 2D ultrasound is feasible in real time and measurably better than static subsampling. The paper demonstrates this with a latent-variable generative model trained on full frames, a frozen decoder, and an inference network that outputs the parameters of a Gaussian latent plus eight Sylvester flow transformations from the partial observation alone. Sampling from that approximate posterior and pushing the samples through the decoder yields a predictive covariance for the next frame; the next mask is chosen from 10,000 random candidates by maximizing either the log determinant of that covariance or its trace, with neighboring scan lines explicitly excluded. On EchoNet, the trace policy outperforms uniform random, variable-density, and equispaced sampling at 5.4%, 8.0%, and 13.4% observation fractions, with the equispaced policy slightly ahead at one intermediate fraction. At 13.4% observation the active policy already approaches the reconstruction quality of the generative model under full observation.

Load-bearing premise

The sampling policy assumes the hidden image state does not change between the frame just observed and the frame being planned for, so the mask is chosen to reveal information about a state that may already be stale.

Editorial extensions

If this is right

  • At 13.4% observation the trace policy's L1 error (0.058) is close to the generative model's full-observation bound (0.053), so further quality gains will have to come from richer priors or better policies rather than denser masks.
  • The active advantage is largest at the most aggressive subsampling ratios, meaning adaptive sampling is most valuable exactly when hardware constraints on frame rate or field of view are tightest.
  • Because the 66 Hz acquisition-decision loop is unoptimized, there is headroom to apply the same active scheme in real-time 2D ultrasound systems.
  • Since all compared methods share the same generative model, the reported reconstruction differences isolate the effect of the sampling strategy itself.
  • Application to 3D ultrasound is the paper's stated next direction, where the volume-rate versus resolution trade-off is substantially harder than in 2D.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • This suggests that a transition model anticipating cardiac phase could extend the gains to faster-moving anatomy, because the paper's policy plans with the current latent rather than a predicted next latent.
  • A testable extension is to treat the explicit ban on neighboring scan lines as a tunable proxy for the imaging system's point-spread function and optimize that spacing for the actual resolution cell.
  • The random 10,000-mask candidate set means policy quality depends partly on how luckily the random masks cover informative columns; a learned or differentiable mask proposer could sharpen the mutual-information estimate.
  • The 15% figure is tied to EchoNet and the tested subsampling ratios, so testing on other anatomies, transducer geometries, and noise levels would reveal whether the adaptive advantage is general or dataset-specific.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes an adaptive scan-line subsampling method for 2D ultrasound video. A variational autoencoder with a Sylvester normalizing-flow encoder infers an approximate posterior over a latent state from partial observations, and a sampling policy selects the next subsampling mask by maximizing an estimate of the mutual information between future observations and the latent state, using either a full covariance or a trace approximation. The method is evaluated on the EchoNet cardiac ultrasound dataset against uniform random, variable-density random, and equispaced baselines. The authors report that trace sampling improves mean absolute reconstruction error by 15% and runs at 66 Hz, and they attribute the gains to the active sampling strategy. The central empirical claim is not consistently supported by Table I, and the comparison is confounded by policy-specific inference models.

Significance. If the headline claim held, this would be a useful contribution toward real-time adaptive ultrasound acquisition: the flow-based posterior encoder provides fast inference under partial observations, the closed-loop training procedure is a sensible design choice, and the reported 0.015 s acquisition step is practically relevant. However, the empirical evidence as presented does not establish the claimed advantage over equispaced sampling, and the lack of uncertainty quantification makes it difficult to assess whether the method improves over baselines at all. The core methodological idea remains interesting and publishable pending a more careful and controlled evaluation.

major comments (4)
  1. [Abstract; Section III.B, Table I] The abstract's claim of a 15% mean absolute error improvement over baselines including equispaced scan-lines is not supported by Table I. Compared with equispaced sampling, trace sampling reduces L1 loss by about 4% at l=6 (0.070 vs 0.073), about 3% at l=9 (0.062 vs 0.064), is worse at l=12 (0.061 vs 0.060), and equal at l=15 (0.058 vs 0.058). The roughly 15% figure is only obtained relative to uniform random or variable-density random sampling, and only at some subsampling rates. The abstract and Section III.B should be revised so that the reported improvement is consistent with the table.
  2. [Table I; Section III.B] No uncertainty quantification is provided for any reported L1, SSIM, or PSNR value. The table gives single-point estimates only, with no standard deviations, confidence intervals, multiple training seeds, or paired significance tests. Because the differences between trace and equispaced sampling are only 0.003 and 0.002 in L1 loss (about 3-4%) and reverse sign at l=12, the measured gains cannot be distinguished from noise without such intervals.
  3. [Section III.A; Algorithm 1] The statement that all methods share the same generative model and therefore differences are attributable to sampling strategy only is misleading, because each policy is paired with its own inference model trained on observations generated under that policy. The comparison therefore conflates the sampling policy with the quality of the policy-specific inference model. A controlled comparison would use a fixed inference model across policies, or retrain all inference models from multiple seeds, before the claim that the sampling strategy is the source of the gains can be made.
  4. [Section II.C, Eq. (5)] The policy assumes an identity transition z_{t+1}=z_t when computing the predictive covariance for the next frame. In cardiac ultrasound, the latent state changes between frames, so the chosen scan-lines maximize information about the current (stale) state rather than the actual next frame. This assumption is stated but not validated; at minimum the paper should compare against a one-frame-delayed or constant policy to measure the sensitivity of the results to this modeling choice.
minor comments (5)
  1. [Section II.B, Eq. (2)] The term labeled 'Likelihood zK' in the equation appears to be the log prior log p(zK), not a likelihood; the label should be corrected.
  2. [Section III.A] The text says the IWAE algorithm is used, but Eq. (2) is written as the standard ELBO; the relationship between the optimized objective and IWAE should be made explicit.
  3. [Section III.A] The neighbor-exclusion mechanism is described only by example; a short formal definition or pseudocode would make the trace-sampling policy reproducible.
  4. [Section III.B] The paragraph reports that trace sampling outperforms at l=6, l=9, and l=15, but the table shows a tie at l=15 in L1-loss; the text should be aligned with the table.
  5. [Section III.A] The claim that increasing NS and S beyond 3 and 10,000 gives 'minimal performance improvement' is not supported by any data; a small sensitivity study would strengthen this statement.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the active-sampling comparison is evaluated on held-out EchoNet test frames with a frozen shared generative model, and no reported quantity reduces to the training objective by construction.

full rationale

The paper's central claim is empirical: active trace/covariance sampling is compared against uniform, variable-density, and equispaced baselines on held-out EchoNet test videos, with L1/SSIM/PSNR computed from posterior-mean reconstructions versus ground-truth frames (Table I). The generative model p_theta is pre-trained on full observations and frozen, after which the inference model q_phi is trained on partial observations via the ELBO in Eq. (2); the sampling objective and the evaluation metric are therefore not definitionally linked. Eqs. (5)-(7) estimate the covariance of future observations from posterior samples and use it to select A_{t+1}, which is a standard information-gain policy rather than a fitted parameter renamed as a prediction. The closest self-referential element is Algorithm 1, where A_{t+1} is chosen using q_phi and then used to acquire the next training observation; the authors explicitly acknowledge this in Section IV: 'because the sampling policy generates the observations on which the inference model is trained, and the inference model in turn affects the sampling policy, their optimization becomes intertwined. This may lead to collapse.' That is a training-dynamics or stability limitation, not a circular reduction of the reported results to their inputs. Self-citations [5,8,10] are contextual related work and are not load-bearing for the empirical claim. The abstract's 15% improvement is not fully supported by Table I (trace is worse than equispaced at l=12 and tied at l=15), and no error bars are given, but this is an evidence-quality issue, not circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim depends on a handful of hyperparameters and the structural assumptions of the generative model and the identity transition. No new physical entities are introduced.

free parameters (6)
  • beta (ELBO weight) = 1e-4
    Set in Section III.A; weights reconstruction likelihood against KL divergence, and the authors note the beta trade-off affects posterior accuracy and sampling quality.
  • NS (posterior samples) = 3
    Section III.A; used in Eq. (5) to estimate the predictive covariance; with only 3 samples the entropy estimate is noisy.
  • S (candidate action set size) = 10000
    Section III.A; random subset of the binomial action space; larger S improves the policy but costs compute.
  • K (flow steps) = 8
    Section III.A; number of Sylvester flow layers in the posterior encoder.
  • latent dimension = 512
    Section III.A; dimensionality of the latent variable z.
  • neighbor exclusion distance = not specified
    Section II.C; the trace policy excludes neighboring scan-lines, but the exact distance is only shown by example.
assumptions (4)
  • domain assumption The generative model p_theta(x|z) and prior p(z) capture the true signal distribution p(x).
    Section II.A; the decoder is pre-trained on full observations and frozen, so the entire sampling policy relies on this prior being adequate.
  • domain assumption Latent state transition is identity: z_{t+1} = z_t.
    Section II.C; 'assuming an identity transition z_{t+1}|z_t'; in cardiac video the heart moves, so the latent state changes between frames.
  • standard math The conditional entropy H(y_{t+1}|z_{t+1},A_{t+1}) equals the noise entropy and is independent of A_{t+1}.
    Section II.C; with y = A x + n and x = g(z), the observation entropy given z is the entropy of n.
  • ad hoc to paper The empirical covariance from NS=3 posterior samples approximates the true predictive covariance.
    Section II.C, Eq. (5); the mutual information estimate rests on this small-sample approximation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Sylvester Posterior Inference for Adaptive Compressed Sensing in Ultrasound Imaging." pith.science (2026). https://pith.science/paper/25ALMGVU

@misc{pith2026250103825,
  author       = {Pith},
  title        = {Pith review of: Deep Sylvester Posterior Inference for Adaptive Compressed Sensing in Ultrasound Imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/25ALMGVU}},
  note         = {Machine review of arXiv:2501.03825}
}
read the original abstract

Ultrasound images are commonly formed by sequential acquisition of beam-steered scan-lines. Minimizing the number of required scan-lines can significantly enhance frame rate, field of view, energy efficiency, and data transfer speeds. Existing approaches typically use static subsampling schemes in combination with sparsity-based or, more recently, deep-learning-based recovery. In this work, we introduce an adaptive subsampling method that maximizes intrinsic information gain in-situ, employing a Sylvester Normalizing Flow encoder to infer an approximate Bayesian posterior under partial observation in real-time. Using the Bayesian posterior and a deep generative model for future observations, we determine the subsampling scheme that maximizes the mutual information between the subsampled observations, and the next frame of the video. We evaluate our approach using the EchoNet cardiac ultrasound video dataset and demonstrate that our active sampling method outperforms competitive baselines, including uniform and variable-density random sampling, as well as equidistantly spaced scan-lines, improving mean absolute reconstruction error by 15%. Moreover, posterior inference and the sampling scheme generation are performed in just 0.015 seconds (66Hz), making it fast enough for real-time 2D ultrasound imaging applications.

Figures

Figures reproduced from arXiv: 2501.03825 by the authors.

Figure 1
Figure 1. Most related to our approach, van de Camp et al. recently proposed the use of deep generative latent variable models for adaptive subsampling designs [10]. While effective, the method relied on Markov Chain Monte-Carlo methods for generating samples from the posterior, rendering inference prohibitively slow for time-sensitive applications such as ultrasound imaging. Moreover, the scene was considered static, and obs… view at source ↗
Figure 1
Figure 1. Schematic overview of the active sampling loop of a single video frame. Partial observations of the full frame are used to estimate the latent [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Reconstruction results for three consecutive frames [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

  1. [1]

    Compressed sensing,

    D. L. Donoho, “Compressed sensing,” IEEE Transactions on Information Theory, vol. 52, no. 4, pp. 1289–1306, 4 2006

    work page 2006

  2. [2]

    Dehazing Ultrasound using Diffusion Models,

    T. S. Stevens, F. C. Meral, J. Yu, I. Z. Apostolakis, J.-L. Robert, and R. J. Van Sloun, “Dehazing Ultrasound using Diffusion Models,” IEEE Transactions on Medical Imaging , pp. 1–1, 2024

    work page 2024

  3. [3]

    Denoising Plane Wave Ultrasound Images Using Diffusion Probabilistic Models,

    H. Asgariandehkordi, S. Goudarzi, M. Sharifzadeh, A. Basarab, and H. Rivaz, “Denoising Plane Wave Ultrasound Images Using Diffusion Probabilistic Models,” IEEE Transactions on Ultra- sonics, Ferroelectrics, and Frequency Control , 2024

    work page 2024

  4. [4]

    Deep Learning-Based Optimal Spatial Subsampling in Ultrasound Nondestructive Testing,

    H. Wang, E. P ´erez, and F. R ¨omer, “Deep Learning-Based Optimal Spatial Subsampling in Ultrasound Nondestructive Testing,” in 2023 31st European Signal Processing Conference (EUSIPCO). IEEE, 9 2023, pp. 1863–1867

    work page 2023

  5. [5]

    Learning Sub-Sampling and Signal Recovery with Applications in Ultrasound Imaging,

    I. A. Huijben, B. S. Veeling, K. Janse, M. Mischi, and R. J. Van Sloun, “Learning Sub-Sampling and Signal Recovery with Applications in Ultrasound Imaging,” IEEE Transactions on Medical Imaging, vol. 39, no. 12, pp. 3955–3966, 12 2020

    work page 2020

  6. [6]

    Efficient B-Mode Ultrasound Image Reconstruction From Sub-Sampled RF Data Using Deep Learning,

    Y . H. Yoon, S. Khan, J. Huh, and J. C. Ye, “Efficient B-Mode Ultrasound Image Reconstruction From Sub-Sampled RF Data Using Deep Learning,” IEEE Transactions on Medical Imaging, vol. 38, no. 2, pp. 325–336, 2 2019

    work page 2019

  7. [7]

    Randomized channel subsampling method for efficient ultrafast ultrasound imaging,

    J. Yu, X. Guo, S. Yan, Q. Le, V . Hingot, D. Ta, O. Couture, and K. Xu, “Randomized channel subsampling method for efficient ultrafast ultrasound imaging,” Measurement Science and Technology, vol. 34, no. 8, p. 084005, 8 2023

    work page 2023

  8. [8]

    Active Deep Probabilistic Subsampling,

    H. Van Gorp, I. Huijben, B. S. Veeling, N. Pezzotti, and R. J. G. Van Sloun, “Active Deep Probabilistic Subsampling,” in Proceedings of the 38th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, M. Meila and T. Zhang, Eds., vol. 139. PMLR, 8 2021, pp. 10 509–10 518

    work page 2021

Show all 22 references
  1. [9]

    Sylvester Normalizing Flows for Variational Inference,

    R. v. d. Berg, L. Hasenclever, J. M. Tomczak, and M. Welling, “Sylvester Normalizing Flows for Variational Inference,” 3 2018

    2018

  2. [10]

    Active Subsampling Using Deep Generative Models by Maximizing Expected Information Gain,

    K. C. van de Camp, H. Joudeh, D. J. Antunes, and R. J. G. van Sloun, “Active Subsampling Using Deep Generative Models by Maximizing Expected Information Gain,” in ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 6 202...

    2023

  3. [11]

    Auto-Encoding Variational Bayes,

    D. P. Kingma and M. Welling, “Auto-Encoding Variational Bayes,” University of Amsterdam, Amsterdam, Tech. Rep., 12 2013

    2013

  4. [12]

    Variational Inference with Normalizing Flows,

    D. J. Rezende and S. Mohamed, “Variational Inference with Normalizing Flows,” Proceedings of the 32nd International Conference on Machine Learning , pp. 1530–1538, 5 2015

    2015

  5. [13]

    Bayesian Inference and the Principle of Maximum Entropy,

    D. K. Foley and E. Scharfenaker, “Bayesian Inference and the Principle of Maximum Entropy,” 7 2024

    2024

  6. [14]

    Video-based AI for beat-to-beat assessment of cardiac function,

    D. Ouyang, B. He, A. Ghorbani, N. Yuan, J. Ebinger, C. P. Langlotz, P. A. Heidenreich, R. A. Harrington, D. H. Liang, E. A. Ashley, and J. Y . Zou, “Video-based AI for beat-to-beat assessment of cardiac function,” Nature, vol. 580, no. 7802, pp. 252–256, 4 2020

    2020

  7. [15]

    Language Modeling with Gated Convolutional Networks,

    Y . N. Dauphin, A. Fan, M. Auli, and D. Grangier, “Language Modeling with Gated Convolutional Networks,” 12 2016

    2016

  8. [16]

    Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift,

    S. Ioffe and C. Szegedy, “Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift,” 2 2015

    2015

  9. [17]

    Gaussian Error Linear Units (GELUs),

    D. Hendrycks and K. Gimpel, “Gaussian Error Linear Units (GELUs),” 6 2016

    2016

  10. [18]

    Importance Weighted Autoencoders,

    Y . Burda, R. Grosse, and R. Salakhutdinov, “Importance Weighted Autoencoders,” 9 2015

    2015

  11. [19]

    PyTorch: An Imperative Style, High-Performance Deep Learning Library,

    A. Paszke, S. Gross, F. Massa, and et al., “PyTorch: An Imperative Style, High-Performance Deep Learning Library,” 12 2019

    2019

  12. [20]

    Attention Is All You Need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention Is All You Need,” Google Brain, Tech. Rep., 2017

    2017

  13. [21]

    xLSTM: Extended Long Short-Term Memory,

    M. Beck, K. P ¨oppel, M. Spanring, A. Auer, O. Prudnikova, M. Kopp, G. Klambauer, J. Brandstetter, and S. Hochreiter, “xLSTM: Extended Long Short-Term Memory,” 5 2024

    2024

  14. [22]

    Very Deep V AEs generalize Autoregressive Models and can outperform them on images,

    R. Child, “Very Deep V AEs generalize Autoregressive Models and can outperform them on images,” ICLR, pp. 1–17, 2021

    2021

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.