pith. sign in

arxiv: 2512.01433 · v2 · submitted 2025-12-01 · 📡 eess.SP

zea: A Toolbox for Cognitive Ultrasound Imaging

Tristan S.W. Stevens , Wessel L. van Nierop , Ben Luijten , Vincent van de Schaft , Ois\'in Nolan , Beatrice Federici , Louis D. van Harten , Simon W. Penninga
show 2 more authors
Noortje I.P. Schueler Ruud J.G. van Sloun
This is my paper

Pith reviewed 2026-05-17 03:22 UTC · model grok-4.3

classification 📡 eess.SP
keywords ultrasound imagingPython toolboxdeep learning integrationdifferentiable pipelineKeras 3signal processingcognitive imaging
0
0 comments X

The pith

Zea provides a modular Python pipeline for ultrasound data that supports custom reconstruction and direct integration of deep learning models.

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

The paper presents zea as a toolbox for cognitive ultrasound imaging. It supplies a flexible, modular, and differentiable pipeline along with pre-defined models for handling ultrasound images and signals. A high-level interface lets users build their own processing chains and insert deep learning components without low-level coding. The implementation rests on Keras 3, which automatically works with TensorFlow, PyTorch, and JAX backends. A sympathetic reader would care because the package lowers the effort needed to combine traditional ultrasound methods with modern machine-learning techniques in medical imaging.

Core claim

Zea offers a flexible, modular, and differentiable pipeline for ultrasound data processing together with a collection of pre-defined models for ultrasound image and signal processing, all accessible through a high-level interface that lets users define custom reconstruction pipelines and integrate deep learning models seamlessly while supporting TensorFlow, PyTorch, and JAX via Keras 3.

What carries the argument

The high-level interface built on Keras 3 that supplies differentiability, modularity, and automatic support for the three major deep-learning backends.

If this is right

  • Users gain ready-made models for common ultrasound image and signal tasks.
  • Custom reconstruction pipelines become straightforward to assemble and modify.
  • Machine-learning components plug directly into ultrasound workflows across frameworks.
  • The same code runs unchanged under TensorFlow, PyTorch, or JAX.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The design could shorten the time from idea to working prototype for adaptive, data-driven ultrasound systems.
  • Similar modular interfaces might be adopted in neighboring fields such as photoacoustic or magnetic-resonance imaging.
  • The differentiability property opens the door to end-to-end optimization of entire imaging chains rather than isolated stages.

Load-bearing premise

That the high-level interface and Keras 3 backend will let users create custom pipelines and add deep learning models without needing substantial extra implementation work.

What would settle it

A documented attempt by an external user to build and train a custom differentiable ultrasound pipeline with an inserted neural network that requires more than a few lines of code beyond the advertised interface.

Figures

Figures reproduced from arXiv: 2512.01433 by Beatrice Federici, Ben Luijten, Louis D. van Harten, Noortje I.P. Schueler, Ois\'in Nolan, Ruud J.G. van Sloun, Simon W. Penninga, Tristan S.W. Stevens, Vincent van de Schaft, Wessel L. van Nierop.

Figure 1
Figure 1. Figure 1: High-level overview of an ultrasound perception-action loop imple￾mented in zea. Statement of need The ultrasound research community has advanced significantly due to publically available high-quality software, including simulation tools such as Field II (Jensen, 2004) and k-wave (Treeby & Cox, 2010), as well as reconstruction and real-time processing libraries like USTB (Rodriguez-Molares et al., 2017), M… view at source ↗
Figure 2
Figure 2. Figure 2: Diffusion posterior samples of a model trained on the EchoNet￾dynamic dataset. surement models mapping from fully-observed to subsampled data. import keras import zea # (batch, height, width) data = ... # create a Greedy Entropy Minimization agent agent = zea.agent.selection.GreedyEntropy( n_actions=width // 8, n_possible_actions=width, img_width=width, img_height=height, ) # these would normally be sample… view at source ↗
Figure 3
Figure 3. Figure 3: Selected scan-lines as chosen by a Greedy Entropy Minimization agent. Availability, Development, and Documentation zea is available through PyPI via pip install zea, and the development version is available via GitHub. GitHub Actions manage continuous integra￾tion through automated code testing (PyTest), code linting and formatting (Ruff), and documentation generation (Sphinx). The documentation is hosted … view at source ↗
read the original abstract

We present zea (pronounced ze-yah), a Python package for cognitive ultrasound imaging that offers a flexible, modular, and differentiable pipeline for ultrasound data processing. Additionally, it includes a collection of pre-defined models for ultrasound image and signal processing. The toolbox is designed to be easy to use, with a high-level interface that enables users to define custom ultrasound reconstruction pipelines and integrate deep learning models seamlessly. Built on top of Keras 3, it supports all three major deep learning backends: TensorFlow, PyTorch, and JAX, making it straightforward to incorporate custom ultrasound processing pipelines into machine learning workflows. Documentation is available at https://zea.readthedocs.io/.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The manuscript introduces zea, a Python package for cognitive ultrasound imaging. It claims to deliver a flexible, modular, and differentiable pipeline for ultrasound data processing built on Keras 3, with support for TensorFlow, PyTorch, and JAX backends. The toolbox includes pre-defined models for image and signal processing and a high-level interface allowing users to define custom reconstruction pipelines and integrate deep learning models.

Significance. If the differentiability and modularity claims hold with working implementations, the toolbox would enable end-to-end gradient flow from deep learning models through ultrasound-specific steps such as beamforming and envelope detection. This could lower the barrier for research on adaptive, learning-driven ultrasound systems and support reproducible workflows across major deep-learning frameworks.

major comments (1)
  1. [Abstract and pipeline-description section] The central claim that custom pipelines remain end-to-end differentiable when users compose pre-defined models or insert custom blocks is asserted in the abstract and introduction but is not accompanied by any code examples, gradient-verification tests, or explicit treatment of classically non-differentiable operations (Hilbert transform, log compression, thresholding, or certain apodization windows). Without such evidence the seamless-integration guarantee for cognitive-imaging use cases cannot be evaluated.
minor comments (1)
  1. The documentation link is provided but no usage examples or benchmark results appear in the manuscript itself; adding a short code snippet illustrating a differentiable pipeline would improve clarity.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive comments on the manuscript. We respond to the major comment below and outline the planned revisions.

read point-by-point responses

  1. Referee: [Abstract and pipeline-description section] The central claim that custom pipelines remain end-to-end differentiable when users compose pre-defined models or insert custom blocks is asserted in the abstract and introduction but is not accompanied by any code examples, gradient-verification tests, or explicit treatment of classically non-differentiable operations (Hilbert transform, log compression, thresholding, or certain apodization windows). Without such evidence the seamless-integration guarantee for cognitive-imaging use cases cannot be evaluated.

    Authors: We agree that the differentiability claim would be strengthened by explicit supporting material. The implementation relies on Keras 3 primitives to enable automatic differentiation where the underlying operations permit it, but the manuscript does not currently demonstrate this with examples or tests. In the revised version we will add code examples of pipeline composition and custom block insertion, numerical gradient-verification tests across the three backends, and a dedicated discussion of non-differentiable steps: the Hilbert transform is realized via FFT (differentiable), log compression will be treated with a differentiable approximation or gradient detachment where appropriate, and thresholding/apodization will be addressed with straight-through estimators or explicit notes on gradient flow. These additions will appear in an expanded pipeline-description section. revision: yes

Circularity Check

0 steps flagged

Software toolbox description contains no derivations or predictions

full rationale

The paper is a description of the zea Python package for cognitive ultrasound imaging. It presents a high-level interface, pre-defined models, and Keras 3 backend support but contains no mathematical derivations, equations, fitted parameters, predictions, or uniqueness theorems. Claims about differentiability and modularity are implementation assertions rather than results derived from prior steps within the paper, so no load-bearing step reduces to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities are introduced; the work consists of a software implementation that depends on existing libraries such as Keras 3.

pith-pipeline@v0.9.0 · 5458 in / 1198 out tokens · 71011 ms · 2026-05-17T03:22:21.647851+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages · 1 internal anchor

  1. [1]

    G., Steiner, B., Tucker, P

    Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., Devin, M., Ghemawat, S., Irving, G., Isard, M., Kudlur, M., Levenberg, J., Monga, R., Moore, S., Murray, D. G., Steiner, B., Tucker, P. A., Vasudevan, V., Warden, P., Zheng, X. (2016). T ensor F low: A S ystem for L arge-scale M achine L earning. In K. Keeton & T. Roscoe (Eds.), 12th USENIX s...

    work page 2016

  2. [2]

    J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., VanderPlas, J., Wanderman-Milne, S., & Zhang, Q

    Bradbury, J., Frostig, R., Hawkins, P., Johnson, M. J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., VanderPlas, J., Wanderman-Milne, S., & Zhang, Q. (2018). JAX : Composable transformations of P ython+ N um P y programs (Version 0.3.13). http://github.com/jax-ml/jax

    work page 2018

  3. [3]

    Brevett, T. (2024). QUPS: A MATLAB T oolbox for R apid P rototyping of U ltrasound B eamforming and I maging T echniques. J. Open Source Softw., 9(101), 6772. https://doi.org/10.21105/JOSS.06772

    work page doi:10.21105/joss.06772 2024

  4. [4]

    Chollet, F., & others. (2015). K eras . https://keras.io

    work page 2015

  5. [5]

    Federici, B., Sloun, R. J. G. van, & Mischi, M. (2024). A ctive I nference for C losed-loop transmit beamsteering in F etal D oppler U ltrasound. CoRR, abs/2410.04869. https://doi.org/10.48550/ARXIV.2410.04869

    work page doi:10.48550/arxiv.2410.04869 2024

  6. [6]

    Garcia, D. (2021). M ake the most of M U S T , an open-source M atlab U ltra S ound T oolbox. 2021 IEEE International Ultrasonics Symposium (IUS), 1--4. https://doi.org/10.1109/IUS52206.2021.9593605

    work page doi:10.1109/ius52206.2021.9593605 2021

  7. [7]

    Jarosik, P., & others. (2020). ARRUS : Research/remote ultrasound platform . https://github.com/us4useu/arrus

    work page 2020

  8. [8]

    Jensen, J. A. (2004). S imulation of advanced ultrasound systems using F ield I I . 2004 2nd IEEE International Symposium on Biomedical Imaging: Nano to Macro (IEEE Cat No. 04EX821), 1, 636--639. https://doi.org/10.1109/ISBI.2004.1398618

    work page doi:10.1109/isbi.2004.1398618 2004

  9. [9]

    Low Propagation Loss X-Band Impedance Matched Lamb Mode Delay Lines in 30% Scandium Aluminum Nitride,

    Kvalevåg, M. D., Vrålstad, A. E., Rindal, O. M. H., Bjåstad, T. G., Denarie, B., Kristoffersen, K., Måsøy, S.-E., & Løvstakken, L. (2023). Vbeam: A F ast and D ifferentiable B eamformer for O ptimizing U ltrasound I maging. 2023 IEEE International Ultrasonics Symposium (IUS), 1--4. https://doi.org/10.1109/IUS51837.2023.10307255

    work page doi:10.1109/ius51837.2023.10307255 2023

  10. [10]

    Leclerc, S., Smistad, E., Pedrosa, J., Østvik, A., Cervenansky, F., Espinosa, F., Espeland, T., Berg, E. A. R., Jodoin, P.-M., Grenier, T., Lartizien, C., D'hooge, J., Løvstakken, L., & Bernard, O. (2019). D eep L earning for S egmentation U sing an O pen L arge-scale D ataset in 2 D E chocardiography. IEEE Trans. Medical Imaging , 38(9), 2198--2210. http...

    work page doi:10.1109/tmi.2019.2900516 2019

  11. [11]

    A., & Bernard, O

    Liebgott, H., Rodriguez-Molares, A., Cervenansky, F., Jensen, J. A., & Bernard, O. (2016). P lane-wave I maging C hallenge in M edical U ltrasound. 2016 IEEE International Ultrasonics Symposium (IUS), 1--4. https://doi.org/10.1109/ULTSYM.2016.7728908

    work page doi:10.1109/ultsym.2016.7728908 2016

  12. [12]

    Luijten, B., Cohen, R., Bruijn, F. J. de, Schmeitz, H. A. W., Mischi, M., Eldar, Y. C., & Sloun, R. J. G. van. (2020). A daptive U ltrasound B eamforming U sing D eep L earning. IEEE Trans. Medical Imaging , 39(12), 3967--3978. https://doi.org/10.1109/TMI.2020.3008537

    work page doi:10.1109/tmi.2020.3008537 2020

  13. [13]

    Nolan, O., Stevens, T. S. W., Nierop, W. L. van, & Sloun, R. van. (2025). A ctive D iffusion S ubsampling. Trans. Mach. Learn. Res., 2025. https://openreview.net/forum?id=OGifiton47

    work page 2025

  14. [14]

    P., Ashley, E

    Ouyang, D., He, B., Ghorbani, A., Lungren, M. P., Ashley, E. A., Liang, D. H., & Zou, J. Y. (2019). E chonet-dynamic: A large new cardiac motion video data resource for medical machine learning. NeurIPS ML4H Workshop: Vancouver, BC, Canada, 5

    work page 2019

  15. [15]

    Z., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Chintala, S

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Köpf, A., Yang, E. Z., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Chintala, S. (2019). P y T orch: A n I mperative S tyle, H igh-performance D eep L earning L ibrary. In H. M. Wallach, H....

    work page 2019

  16. [16]

    Data-driven process- ing using parametric neural network for improved Bluetooth Channel Sounding distance estimation,

    Penninga, S. W., Gorp, H. V., & Sloun, R. J. G. van. (2025). D eep S ylvester P osterior I nference for A daptive C ompressed S ensing in U ltrasound I maging. 2025 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2025, Hyderabad, India, April 6-11, 2025 , 1--5. https://doi.org/10.1109/ICASSP49660.2025.10888253

    work page doi:10.1109/icassp49660.2025.10888253 2025

  17. [17]

    Rodriguez-Molares, A., Rindal, O. M. H., Bernard, O., Nair, A., Lediju Bell, M. A., Liebgott, H., Austeng, A., & LØvstakken, L. (2017). T he U ltra S ound T ool B ox. 2017 IEEE International Ultrasonics Symposium (IUS), 1--4. https://doi.org/10.1109/ULTSYM.2017.8092389

    work page doi:10.1109/ultsym.2017.8092389 2017

  18. [18]

    van de, Nolan, O., & Sloun, R

    Schaft, V. van de, Nolan, O., & Sloun, R. J. G. van. (2025). O ff-grid U ltrasound I maging by S tochastic O ptimization. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 72, 1245--1255. https://doi.org/10.1109/TUFFC.2025.3586377

    work page doi:10.1109/tuffc.2025.3586377 2025

  19. [19]

    Sloun, R. J. G. van. (2024). Active inference and deep generative modeling for cognitive ultrasound. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 71(11), 1478--1490. https://doi.org/10.1109/TUFFC.2024.3466290

    work page doi:10.1109/tuffc.2024.3466290 2024

  20. [20]

    Smistad, E. (2021). FAST: A framework for high-performance medical image computing and visualization. In S. McIntosh-Smith (Ed.), IWOCL'21: International workshop on OpenCL, munich germany, april, 2021 (pp. 14:1--14:2). ACM . https://doi.org/10.1145/3456669.3456717

    work page doi:10.1145/3456669.3456717 2021

  21. [21]

    Stevens, T. S. W., Meral, F. C., Yu, J., Apostolakis, I. Z., Robert, J.-L., & Sloun, R. J. G. van. (2024). D ehazing U ltrasound U sing D iffusion M odels. IEEE Trans. Medical Imaging , 43(10), 3546--3558. https://doi.org/10.1109/TMI.2024.3363460

    work page doi:10.1109/tmi.2024.3363460 2024

  22. [22]

    Stevens, T. S. W., Nolan, O., Robert, J.-L., & Sloun, R. J. G. van. (2025). S equential P osterior S ampling with D iffusion M odels. 2025 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2025, Hyderabad, India, April 6-11, 2025 , 1--5. https://doi.org/10.1109/ICASSP49660.2025.10889752

    work page doi:10.1109/icassp49660.2025.10889752 2025

  23. [23]

    Stevens, T. S. W., Nolan, O., Somphone, O., Robert, J.-L., & Sloun, R. J. G. van. (2025). H igh V olume R ate 3 D U ltrasound R econstruction with D iffusion M odels. CoRR, abs/2505.22090. https://doi.org/10.48550/ARXIV.2505.22090

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.22090 2025

  24. [24]

    Stevens, T. S. W., Overdevest, J., Nolan, O., Nierop, W. L. van, Sloun, R. J. G. van, & Eldar, Y. C. (2025). D eep generative models for B ayesian inference on high-rate sensor data: Applications in automotive radar and medical imaging. Philosophical Transactions A, 383(2299), 20240327. https://doi.org/10.1098/rsta.2024.0327

    work page doi:10.1098/rsta.2024.0327 2025

  25. [25]

    E., & Cox, B

    Treeby, B. E., & Cox, B. T. (2010). K- W ave: M A T L A B toolbox for the simulation and reconstruction of photoacoustic wave fields. Journal of Biomedical Optics, 15(2), 021314--021314. https://doi.org/10.1117/1.3360308 CSLReferences document

    work page doi:10.1117/1.3360308 2010