List of Papers Browse by Subject Areas Author List
Abstract
Deep neural networks (DNNs) have shown strong potential
for ultrasound computed tomography (USCT) reconstruction in ideal
noise-free environments, yet existing DNNs are vulnerable to the noisy
conditions in clinical practice, as they equally treat inputs that suffer
mild, moderate, or severe noise. More challenging, the distributions of
noise shift along with the environment, indicating the less effectiveness of
noise-aware training, which injects a specific noise distribution into the
training data. We rethink these challenges and observe that the DNN
models can become more robust to noise if we know the noise sources
and filter them out. This filtering operation is very alike the Multiplexers
(or MUX), a fundamental combinational circuit in digital logic design.
However, the challenge here is that noise can happen randomly during
inference; as a result, the manually predefined MUX cannot work. To ad-
dress these challenges, we propose MUX-USCT, a novel encoder-decoder
DNN architecture that encodes the known acoustic acquisition geometry
with an “adaptive MUX” that can automatically identify and filter noise,
where the attention mechanism is applied in reconstructing the speed-of-
sound map. On the OpenPros benchmark, MUX-USCT reaches 6.88 m/s
MAE with 17% fewer parameters than the leading baseline with 7.65 m/s
of MAE. Under simulated clinical noise, it remains stable across diverse
degradation types that cause geometry-agnostic baselines to fail. Results
show that the attention distributions in MUX-USCT reveal patterns that
shift in response to transducer signal degradation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5332_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/TheYuchen/mux-usct-miccai2026
Link to the Dataset(s)
OpenPros dataset: https://huggingface.co/datasets/ashynf/OpenPros
BibTex
@InProceedings{YuaYuc_MUXUSCT_MICCAI2026,
author = { Yuan, Yuchen AND Wu, Hanhan AND Li, Jinyang AND Wang, Hanchen AND Wu, Yixuan AND Lin, Youzuo AND Yang, Lei},
title = { { MUX-USCT: A Noise-Robust Neural Network for Ultrasound Computed Tomography } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16888},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes MUX-USCT, a geometry-aware deep neural network for robust ultrasound computed tomography (USCT) reconstruction. The key idea is to mimic a hardware multiplexer (MUX) in a learned manner: a path-aware encoder uses transducer attention to assess the information of structured acquisition channels, while a position-aware decoder uses per-pixel cross-attention to route path-region tokens back to spatial locations. Within the scope of the paper’s experiments, the method demonstrates strong robustness under several simulated clinical degradation scenarios, despite being trained only on clean data, and it achieves better clean-data accuracy with fewer parameters than the main baselines reported in the manuscript.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
6.1 A major strength of the paper is that it does not treat USCT inputs as a generic flattened feature tensor. Instead, it explicitly preserves path identity and transducer identity, and builds the network around the known acquisition geometry. This is a meaningful architectural contribution, because it aligns the model design with the physics of the sensing system rather than relying purely on generic image-network priors. The MUX analogy is conceptually clear and helps motivate why selective routing may improve robustness. 6.2 The quantitative and qualitative results are convincing within the simulated benchmark used in the paper. Table 2 shows that MUX-USCT outperforms InversionNet, ViT-Inversion, and Inv-FT on both clean data and the combined S4 degradation scenario, while using fewer parameters. Figures 4 and 5 further show that the method remains substantially more stable than the baselines across source noise, receiver noise, coupling loss, and combined degradation. This is an important empirical result, especially because Inv-FT is trained with noise augmentation whereas MUX-USCT is not. 6.3 The attention analysis in Figure 6 is one of the most interesting parts of the paper. The encoder attention maps suggest that corrupted transducers tend to receive lower weights, while the decoder visualizations indicate tissue- or location-dependent routing behavior. Even if these analyses are not yet sufficient to establish clinical diagnostics, they do provide useful insight into what the model may be learning and how the geometry-aware design differs from geometry-agnostic baselines. 6.4 Table 3 shows a meaningful separation of roles between the encoder and decoder attention modules. Removing encoder attention causes a modest clean-data drop but a large robustness loss under source corruption, while removing decoder cross-attention significantly harms clean reconstruction quality. This supports the claim that the encoder mainly contributes to channel selection robustness and the decoder mainly contributes to spatial reconstruction fidelity.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
7.1 Although the manuscript does not present real-world clinical validation, it makes several application-facing claims, including assisting specialists “in the field” and using attention weights for “clinical monitoring” and transducer health indication. Since all experiments are conducted on OpenPros FDTD simulations with simulated degradations, the current evidence more directly supports robustness on a realistic in-silico benchmark than validated phantom, exvivo, or in vivo clinical utility. I therefore recommend tempering the clinical-use wording, or alternatively adding at least phantom / ex vivo / in vivo validation to better support these application-oriented claims. 7.2 The choice M=5 is currently fixed without sensitivity analysis, even though MMM directly controls the number of path tokens and therefore affects both representational capacity and computational cost. I therefore believe the paper would be stronger if it at least included a sensitivity study over MMM. In addition, because the method is presented as being parameterized by acquisition geometry and the conclusion suggests broader applicability to synthetic-aperture USCT systems, a brief scaling discussion — or, if feasible, a small pilot experiment—on changed array density / field of view would better support that broader claim. 7.3 The simulated degradation scenarios are clinically motivated and reasonable as a first step, but some are still simplified. In particular, S3 models coupling loss only as 10–50% amplitude attenuation, whereas real ultrasound propagation and coupling conditions can also involve phase aberration, frequency-dependent attenuation, and more complex wavefront distortion. In addition, the current study does not examine geometry perturbation, calibration drift, or slight probe deformation. These omissions do not undermine the main contribution, but they do mean that the present evidence is strongest for the four simulated scenarios in Table 1 rather than for a broader range of real-world perturbations. 7.4 The current comparison mainly shows that a baseline fine-tuned on S1-style source corruption transfers poorly to mismatched degradations such as S2–S4.This is a meaningful finding, but it is somewhat narrower than a general statement about the limits of noise-aware training as a whole. I therefore suggest either adding a stronger multi-noise-trained baseline, or tempering the claim to emphasize the limitation of single-distribution noise-aware fine-tuning under distribution shift. 7.5 The explanation in Section 2—that corrupted elements produce keys that fall outside the learned clean-signal subspace and therefore receive lower softmax weights—is intuitive and consistent with the empirical results, but it is still best understood as a heuristic interpretation rather than a formal guarantee. I therefore suggest presenting this mechanism a bit more cautiously. If feasible, the paper would be further strengthened by adding one stronger stress test, such as a more challenging out-of-distribution corruption setting or a case where a larger fraction of channels are jointly degraded, to better probe the boundary of this explanation. 7.6 The reported results are mostly point estimates. The paper would be stronger if it included standard deviations, confidence intervals, multiple random seeds, or statistical significance testing, especially for Table 2 and the ablation results in Table 3.This would help clarify whether the observed improvements are consistently reproducible. 7.7 A notable missing comparison is against simpler geometry-aware or rule-based channel-selection baselines. This omission is particularly important because the paper’s own motivation experiment in Fig. 2 shows that manual MUX-inspired reweighting already yields large gains under corrupted inputs. Given that result, the current evaluation does not fully disentangle how much of the improvement comes from the general idea of structured channel/path reweighting, and how much specifically comes from the proposed learned encoder-decoder attention design. The paper would therefore be stronger if it compared against at least one simple heuristic baseline, such as signal-quality-based channel masking/reweighting, path-level gating, or another lightweight structured routing method.
[R1] Wang, H. et al. OpenPros: A Large-Scale Dataset for Limited View Prostate Ultrasound Computed Tomography. arXiv, 2025. [R2] Wu, Y. , Lin, Y. InversionNet: An Efficient and Accurate Data-Driven Full Waveform Inversion. IEEE Transactions on Computational Imaging, 2020. [R3] Virieux, J. , Operto, S. An Overview of Full-Waveform Inversion in Exploration Geophysics. Geophysics, 2009. [R4] Jain, S. , Wallace, B. C. Attention is not Explanation. NAACL, 2019.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I rated this paper as Weak Accept because I found the core idea both meaningful and reasonably well supported within the paper’s experimental scope, but not yet strong enough for a clear accept. The main positive factor is the geometry-aware architectural design: instead of flattening USCT measurements into a generic tensor, the method explicitly preserves path and transducer identity and uses a path-aware encoder plus a position-aware decoder to perform learned selective routing. This is a well-motivated contribution for USCT, where acquisition geometry is physically meaningful rather than incidental. Empirically, the paper is also strong on its chosen benchmark: Table 2 shows that MUX-USCT improves over InversionNet, ViT-Inversion, and Inv-FT on both clean data and the combined degradation scenario S4, while also using fewer parameters; Fig. 4 and Fig. 5 further suggest substantially better robustness across source noise, receiver noise, coupling loss, and combined degradation. I also found the attention analysis and ablation study useful: Fig. 6 provides some insight into how the model may be downweighting corrupted channels and routing information spatially, while Table 3 supports a meaningful division of labor between encoder attention for robustness and decoder cross-attention for reconstruction fidelity. At the same time, several limitations kept my score at the borderline level. Most importantly, the evidence is still entirely based on OpenPros FDTD simulations with simulated degradations, so I do not think the current results fully justify some of the more application-facing claims about clinical monitoring, troubleshooting in the field, or broader practical utility. In addition, the robustness evidence is strongest only for the four simulated scenarios studied in Table 1, some degradation models remain simplified, and the comparison against noise-aware training is somewhat narrow because Inv-FT is fine-tuned only on S1-style source corruption. I also think the paper would be stronger with a sensitivity analysis for the fixed choice of M=5, some discussion or experiment on changed acquisition geometry, statistical reporting beyond single point estimates, and at least one simpler geometry-aware or rule-based channel-selection baseline, especially given the strong manual reweighting result already shown in Fig. 2.Overall, I think the paper is marginally above threshold because the method is well aligned with the sensing physics and the simulation results are promising, but my support would depend on the rebuttal appropriately narrowing the claims and clarifying these missing validations.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
Inspired by multiplexers in digital logic, the paper starts from the observation that, under signal degradation, manually filtering out noisy acquisition paths can substantially improve prostate USCT speed-of-sound (SoS) reconstruction. Building on this intuition, the authors propose MUX-USCT, an encoder-decoder framework that explicitly encodes the physical acquisition geometry and learns to identify and suppress noisy channels. Unlike conventional networks that treat all input channels equally, MUX-USCT uses a path-aware encoder in which attention over transducer elements acts like learned MUX select signals, and compresses each path into regional tokens that capture different acoustic subregions. A position-aware decoder then allows each spatial location to attend to all path-region tokens according to its geometry. Trained on paired clean USCT signals and corresponding ground-truth SoS maps, the model shows stronger robustness across several simulated noise scenarios while also providing interpretable attention maps.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper proposes a well-motivated architecture that explicitly incorporates physical acquisition geometry into the reconstruction framework, together with a position-aware decoder for spatially selective path routing. This is a meaningful departure from geometry-agnostic designs and improves both interpretability and reconstruction quality. 2.The model is trained on paired clean USCT signals and ground-truth SoS maps, and uses the gap between intact-signal statistics and corrupted channels to naturally suppress noisy features at inference time. This gives it stronger robustness across several simulated noise settings without requiring noise-augmented training. 3.The attention visualizations are a notable strength, as they reveal transducer quality estimates and tissue-dependent spatial routing, which could help USCT specialists troubleshoot acquisition issues in practice.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The clinical relevance remains limited because the open dataset is collected and used for training only on phantoms. Although several clinically motivated noise types are simulated, real in vivo conditions are much more complex, involving multiple scattering, reflection, diffraction, and other effects. Even without in vivo ground truth, the paper should at least show in vivo reconstructions and compare them with a classical method such as FWI. 2.The claim of training on “clean” data is somewhat overstated. Since the data are acquired on a real system, they inevitably contain some native noise; the main difference is that the simulated degradations are stronger than the usual system noise. More importantly, it is unclear whether this training strategy would transfer well to realistic settings where the acquisition itself is already highly noisy. The paper does not show that the method can be trained on noisy data while retaining denoising ability. 3.In Table 1, the phrase “10–50% amplitude attenuation” is not precise enough. It should be clarified whether this attenuation is defined through path, transducer, or by some other rule. 4.In Fig. 4, for S1 and S2, the proposed method sometimes appears to perform better as the noise level increases, which is counterintuitive. This raises the possibility that the model is better at detecting severe corruption than moderate corruption. The paper should provide a clearer explanation for this behavior and discuss it explicitly as a limitation of the method. 5.The paper does not provide sufficient visualization or discussion of the bone-related results, and the corresponding training strategy is also unclear.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Weak accept. The main reason is that the paper presents a technically interesting architecture that explicitly incorporates acquisition geometry into USCT reconstruction, achieves strong robustness under several simulated degradation settings without noise-augmented training, and provides interpretable attention patterns with potential practical value. However, the validation is still limited to phantom data, the notion of “clean-data training” is not entirely convincing, and several experimental details and behaviors remain insufficiently explained. Overall, the core idea is promising enough for acceptance, but the empirical support is still incomplete.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The paper proposed a deep neural network with geometry-aware and path-aware encoder and decoder, which achieves superior imaging performance and noise robustness.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper analyse the acoustic paths and degradation and the effect for image reconstruction, and design an interesting neural network to consider the acoustic path.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The ablation experiements may be insufficient to justify the author’s designs.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper propose to improve the robustness by considering the acoustic paths and achieve superior performance. Yet need more ablation experiements.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank all reviewers. All three recognized the core contribution: a geometry-aware architecture aligned with USCT sensing physics (R1-6.1, R2-S1), strong robustness without noise augmentation (R1-6.2, R2-S2), interpretable attention (R1-6.3, R2-S3), and a well-supported encoder-decoder division of labor (R1-6.4). We address the major concerns below.
[1] Scope of Claims (R1-7.1, R1-7.5, R2-W1)
R1 and R2 note that clinical-facing language exceeds the simulation-only evidence. We agree. In the camera-ready, we will reframe “clinical monitoring” and “troubleshooting in the field” as potential future applications, and clarify that the attention filtering is an empirical observation, not a guarantee.
On in vivo validation (R2-W1): in vivo USCT data with ground-truth SoS maps is extremely scarce; OpenPros [17] is the largest publicly available prostate USCT benchmark. In vivo validation is an important next step.
[2] Clarifications on Data and Experimental Details (R2-W2, W3, W4, W5)
R2-W2 (“clean data” overstated): While OpenPros [17] phantoms are derived from real clinical MRI/CT and ex vivo specimens, the waveform data (network input) is generated via FDTD simulation and contains no sensor or system noise. “Clean-data training” accurately describes our setup. R2 also asks whether the method transfers when training data itself is noisy. MUX-USCT’s robustness comes from structural design (attention learns to weight channels by informativeness) rather than from specific properties of the training noise distribution. Training on noisy data is a relevant direction, but clean-data training is the standard protocol on this benchmark.
R2-W3 (S3 definition): Coupling loss is applied per-source element. Each degraded source’s waveform amplitude is scaled by a random factor in [0.5, 0.9] (i.e., 10-50% attenuation as stated in Table 1).
R2-W4 (Fig. 4, counterintuitive behavior): This occurs in S1/S2 where the x-axis is noise intensity, not the count of corrupted channels. Stronger noise on a fixed set of channels makes them easier to detect and suppress (Eq. 1), while the pool of clean channels remains unchanged. S3’s x-axis is the number of degraded sources, so increasing it reduces usable information regardless of detection. Moderate corruption remains harder to detect, which we acknowledge as a limitation.
R2-W5 (bone): Fig. 6b directly shows tissue-dependent routing: bone relies more on cross-probe paths (TA->TR, TR->TA) while prostate draws from same-probe reflection (TR->TR). Fig. 5 confirms bone boundary preservation across all scenarios. No tissue-specific training strategy is used.
[3] Baseline Comparisons (R1-7.4, R1-7.7)
R1-7.4 (Inv-FT): We agree that Inv-FT’s S1-only fine-tuning makes its poor S2-S4 transfer partly expected, and we will narrow the claim accordingly. More broadly, real-world noise types and distributions are unpredictable, so exhaustive noise-augmented fine-tuning is not a scalable solution. This motivates MUX-USCT’s design: the model should be structurally adaptive rather than reliant on anticipating every degradation at training time.
R1-7.7 (simpler baselines): The manual reweighting in Fig. 2 requires oracle knowledge of corrupted transducers, which is unavailable in practice. It motivates the approach but is not deployable; MUX-USCT’s contribution is making this selection learned and adaptive.
[4] Additional Analyses (R1-7.2, R1-7.6, R1-7.3, R3)
We appreciate these suggestions. On M (R1-7.2): M=5 already achieves SOTA on both clean and noisy data with 17% fewer parameters (Table 2); a sensitivity study would be a useful follow-up. On statistics (R1-7.6): we will extend evaluation to the full test set and report standard deviations in the camera-ready. On degradation complexity (R1-7.3) and additional ablations (R3): these are natural extensions we intend to pursue. Finally, we intend to release our code upon acceptance to support reproducibility.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
Given that all three reviewers have recommended acceptance, I am pleased to congratulate the authors on the acceptance of their paper at this phase. Despite a consensus among three positive reviews, two of the reviewers have explicitly stated that their decisions of weak acceptance are conditional. As such, the authors are invited to response after provisional acceptance to clarify the missing details in the experimental validations raised by R1 and R2.Given the limited space, the authors are advised to prioritize the most critical points and allocate their response strategically to maximize the effectiveness of their reply.