List of Papers Browse by Subject Areas Author List
Abstract
Plane-wave ultrasound (PWUS) facilitates functional imaging through a high frame rate of a few thousand Hz. However, its application remains constrained due to the inferior B-mode image quality in comparison to conventional ultrasound imaging such as focused beam ultrasound (FBUS). In this paper, a data-driven approach is proposed through two steps to enhance the quality of PWUS images. In the first step, the unpaired neural Schrödinger bridge (UNSB) is employed to synthesize high-fidelity images that structurally correspond to the low-quality PWUS images. In the second step, our proposed model, R2B-WFC, is trained to reconstruct high-quality images from the PWUS radio frequency signals, incorporating a wavelet Fourier convolution (WFC) module. Multiple losses are also suggested, combining perceptual loss from a USNB pretrained model and a Markovian discriminator to preserve high-frequency detail more effectively. As a result, Fréchet Inception Distance (FID), Kernel Inception Distance (KID), Learned Perceptual Image Patch Similarity (LPIPS), Feature Similarity Index Measure (FSIM), Signal to noise ratio (SNR), and Contrast Ratio (CR) scores were 136.32, 0.0356, 0.1956, 0.9514, 41.18 dB, and 27.48 dB, respectively. Compared to image-to-image translation methods, R2B-WFC from RF signal-to-image also shows faster inference time.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0389_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{JeoHyu_R2BWFC_MICCAI2025,
author = { Jeong, Hyunsu and Yoon, Chiho and Sung, Minsik and Kim, Kiduk and Park, Dougho and Kim, Chulhong},
title = { { R2B-WFC Ultrasound Reconstruction: Wavelet Fourier Convolution-based Reconstruction from Radio Frequency to Image } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15975},
month = {September},
page = {453 -- 463}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors proposed low quality RF to high quality Bmode transformation. They used unpaird high and low quality data for training which is an advantage.
- 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.
- Novelty is suitable for MICCAI
- using unpaird data for training and employing unpaired neural Schrödinger bridge is a plus
- Ablation study is convincing enough
- The results are promising
- The choice of loss functions is interesting and suitable for this task
- 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- I think the main weakness is in the experiments section of other evaluated methods (not the main method). USNB CycleGan etc are all designed for image (Bmode) not RF. So adding a few convolution layers is not fair to be used for RF and we can see in Fig. 4 that they are not exactly showing the same anatomical structure. Therefore I believe that this comparison is not accurate. Even with this shortcomming, the method RF to Bmode (last row in Table 2) has better results than Bmode to Bmode (Table 1) which shows that using RF is advantageous. I suggest to merge the tables into one table. the proposed method employs RF and other methods employ Bmode.
2- Utlizing FFT and wavelet in this way is questionable to me. The RF data is not stationary so when FFT and wavelet are being used, the RF data should be divided into small patches to become stationary.
3- The method should also be used for anatomical structure not used in the training.
- 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 provide sufficient information for reproducibility.
- 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/2025/en/REVIEWER-GUIDELINES.html
Please see the weakness section.
- 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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I think the paper has enough contribution to be suitable for MICCAI . The main weakness was the way experiments were done for other compared methods.
- 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 #2
- Please describe the contribution of the paper
This manuscript presents a two-step deep learning approach to enhance plane-wave ultrasound imaging. It uses an unpaired neural Schrodinger bridge to generate high-fidelity targets and introduces R2B-WFC, which reconstructs high-quality images from RF data using a wavelet Fourier convolution module.
- 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 proposed training pipeline to create a model to reconstruct high-quality B-mode images from plane-wave RF data is novel.
- The presented qualitative and quantitative results visibly outperforms previous techniques doing the same task.
- 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.
- It is not clear from the presented results whether the technique’s performance could be generalized across diverse datasets acquired from different organs.
- A few crucial information are missing, making the current version of the manuscript hard to reproduce.
- 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 provide sufficient information for reproducibility.
- 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/2025/en/REVIEWER-GUIDELINES.html
- Abstract, last sentence: How much faster? Please provide quantitative information.
- Abstract: Please add a sentence to highlight the clinical potential of this work.
- Equation 4: How did you optimize the parameter values?
- Section: More information (e.g., ultrasound imaging parameters, train-test-split, training parameters, etc.) is needed.
- How do you tackle domain shift? For instance, if you train your framework on thyroid data, can it produce correct B-mode images from RF data acquired from liver?
- Fig. 3, Row 2: Why is the thyroid located at higher depth on the FBUS image?
- Reproducibility: Are you going to publish your data/code?
- 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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Major factors:
- The presented training approach for plane wave ultrasound reconstruction is appreciable
- The presented results are good.
- The reproducibility concern should be addressed in the rebuttal stage.
- 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 #3
- Please describe the contribution of the paper
This paper propsed a data-centeric US reconstruction method which directly transform RF signal to B model image. This method uses (unpaired neural Schrödinger bridge) UNSB to generate high-quality ultrasound images and trains the R2B-WFC model with these images. The approach not only achieves high-quality ultrasound imaging but also accelerates the reconstruction speed.
- 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) This method innovatively applies the UNSB approach to the generation of high-quality ultrasound images. (2) The method designs the R2B-WFC model, which innovatively uses a wavelet Fourier convolution module to directly convert RF signals into high-quality ultrasound images.
- 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.
When obtaining high-quality image pairs, the author used Verasonics devices and GE to generate low-quality and high-quality images respectively, which to some extent led to insufficient diversity of data samples. Therefore, it is necessary to consider increasing the generalization of experimental proof methods.
- 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 provide sufficient information for reproducibility.
- 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/2025/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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The author innovatively combines the UNSB method with the wavelet Fourier convolution method, proposing a technique to directly generate high-quality images from RF signals. This approach demonstrates advantages in speed, but its generalization across different devices still requires experimental validation.
- 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 sincerely thank all reviewers for their valuable comments. As suggested, we will revise our submission by adding key information and further elaborating on the evaluation of generalization performance.
Q1. Discussion on generalizability (R1, R2, R3). We fully agree that the generalization performance needs to be addressed by applying the model on anatomical structure not used in the training. In this study, we collected thyroid, carotid artery, and musculoskeletal (MSK) using a linear probe. To further validate the generalizability of our model across different anatomies accessible with a linear probe, we plan to include additional superficial structures such as the breast, peripheral vasculature, and lymph nodes in future work. In contrast, the liver is typically imaged using a convex probe, which operates at a lower frequency, allowing deeper tissue penetration and a wider field of view. Therefore, we also plan to evaluate our model on liver data acquired with a convex probe to assess its performance under different probe settings and imaging characteristics.
Q2. Detailed implementation descriptions (R2, R3). We very thank for the comments about implementation details. We empirically determined the parameter values in Equation 4 based on preliminary experiments on the validation set. Specifically, we performed a grid search over a plausible range of values and selected those that achieved stable convergence and consistent performance across multiple runs. However, we fully agree that it is necessary to show how the parameter that determines the importance of individual losses affects the final result. We will address these results in the future work. In ultrasound setting, both imaging techniques such as GE and Versonics employed harmonic imaging with the demodulation frequency of 12 MHz. For training, the dataset was split into training, validation, and test sets with a 6:2:2 ratio, corresponding to 309, 104, and 104 samples, respectively. The models were trained for 3000 epochs using a batch size of 2, a learning rate of 1e-3, and the Adam optimizer.
Q3. Methodology about the RF patch size (R1). We appreciate the reviewer’s insight regarding the instability of RF signals and the suitability of small patches, especially when training frequency-domain modules like FFT- or wavelet-based convolutions. While small patches may capture local signal variations more effectively, we empirically found that overly small patches led to degraded performance due to the lack of contextual information, which is essential for reconstructing anatomically meaningful B-mode images. In particular, Fourier-based modules needs sufficiently large spatial support to extract meaningful global frequency patterns. To balance signal locality and structural context, we selected an intermediate patch size that maintains sufficient spatial support while preserving the ability to model local frequency patterns. We will prove this argument by conducting ablation study about the patch size for the future.
Q4. Comparison the focused beam ultrasound (FBUS) image with plane wave ultrasound (PWUS) image (R2). Fig. 3, Row 2 demonstrates that the FBUS thyroid shows higher depth than PWUS. FBUS concentrates acoustic energy at a specific focal depth, resulting in higher signal-to-noise ratio (SNR) and better resolution in deeper tissues. In contrast, Plane Wave Ultrasound (PWUS) distributes energy more uniformly, which leads to reduced SNR and image quality at greater depths.
Q5. Clinical potential in this study (R2). Plane-wave ultrasound (PWUS) provides microvascular flow information. The results demonstrate that R2B-WFC shows promising clinical utility in the quantification of microvascular blood flow and the assessment of early vascular disorders.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
All reviewers are in agreement that the proposed method makes a substantial contribution in several key aspects. Specifically, it introduces a novel integration of deep learning modules—namely, the unpaired neural Schrödinger bridge and wavelet Fourier convolution—to enable training an ultrasound (US) image reconstructor using unpaired datasets for RF signal-to-image transformation, which was rarely addressed in other researches. The method is further validated by its ability to generate high-quality US images and demonstrates the potential to accelerate the reconstruction process [R3]. Based on these merits, we recommend early acceptance and believe the innovation and impact of this work will significantly benefit the ultrasound research community.
Even with this recommendation, the AC would like to highlight that all reviewers have raised minor but consistent concerns that should be addressed in the final version. These include: (1) improving reproducibility by either releasing the source code or providing more detailed implementation descriptions, and (2) including a necessary discussion on the model’s generalizability—particularly how it performs when applied to data from different vendors or anatomical regions beyond the training set.