List of Papers Browse by Subject Areas Author List
Abstract
Weakly supervised methods aim to reduce the burden of dense manual labeled annotations by using coarse supervision, yet they still rely on intensive human-provided labels, which remain particularly labor-intensive given the large number of blurred ultrasound frames. Moreover, recent coarse supervisions may further lead to performance degradation, especially for blurred lesion boundaries and motion artifacts. Motivated by the strong semantic understanding of recent vision–language models (VLMs) in medical imaging, we propose a novel weakly supervised method that effectively regards the VLM-embedded prior knowledge as the high-level supervisor to guide the segmentation model in autonomously exploring accurate tumor masks. Specifically, we propose a Guidance Conditional Diffusion Strategy to adaptively integrate multiple lesion cues for learning lesion characteristics. Then, inspired by the relative advantage comparison commonly used in reinforcement learning (RL), we propose an efficient Structure Advantage Learning strategy that encourages the model to actively explore optimal boundary details and overcome the limitations of weak supervision. We further collect a large-scale dataset with 433 annotated breast lesion videos to facilitate research on this topic. Extensive experiments on 3 video datasets demonstrate that our method consistently outperforms recent state-of-the-art methods while substantially reducing the manual annotation workload.
Our dataset and code will be released upon this publication.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1072_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{LiJia_Exploring_MICCAI2026,
author = { Li, Jialu AND Liu, Zekai AND Wang, Hongqiu AND Xu, Huihui AND Wang, Qiong AND Zhu, Lei},
title = { { Exploring Efficient Weakly Supervised Ultrasound Video Object Segmentation with VLM-Guided Structure Advantage Learning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents a novel weakly supervised approach for ultrasound video object segmentation that leverages VLMs as high-level supervisors to reduce annotation burden. The key contributions include: (1) a VLM-driven weak annotation framework that replaces dense pixel-level labels with coarse bounding box supervision requiring only ~40% manual correction. (2) A GCDS diffusion strategy that adaptively integrates multi-modal cues. (3) A Structure Advantage Learning mechanism inspired by reinforcement learning to progressively explore optimal lesion boundaries. (4) The authors also contribute a large-scale dataset of 433 annotated breast lesion videos, with experiments demonstrating consistent improvements over state-of-the-art methods across three benchmarks
- 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 proposed VLM-driven weak supervision framework substantially alleviates the labor-intensive burden of manual labeling in medical video segmentation. By leveraging frozen VLMs (Qwen3-VL-8B) to generate coarse bounding boxes with confidence scores, the method reduces manual annotation workload by approximately 60% compared to conventional box-supervised approaches.
2.The paper introduces an innovative combination of the GCDS and Structure Advantage Learning, which effectively addresses boundary ambiguity in weakly supervised settings. The Dirichlet-based adaptive weighting mechanism and RL-inspired relative quality quantification enable autonomous exploration of optimal lesion structures without requiring complex reinforcement learning pipelines.
- 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 method section is difficult to follow, with insufficient intuitive explanations of key components like GCDS and Structure Advantage Learning. Figure 1 is cluttered with excessive information and lacks clear visual hierarchy, making it challenging to understand the workflow and interactions between modules.
2.The approach heavily depends on Qwen3-VL-8B’s zero-shot capability for pseudo-box generation; however, this model is primarily trained on natural datasets, raising concerns about its reliability in ultrasound-specific scenarios with blurred boundaries and artifacts. Furthermore, the VLM operates frame-independently without explicit temporal modeling. Incorporating video-aware trackers like SAM 2 [1] could significantly enhance spatiotemporal consistency of pseudo-labels across frames, which is critical for video segmentation yet completely overlooked.
[1] Sam 2: Segment anything in images and videos, 2024.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper suffers from unclear methodological descriptions and over-reliance on VLM zero-shot capabilities.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper proposes a weakly supervised video object segmentation method for breast lesion ultrasound videos. Instead of dense pixel-level annotations, the method uses a frozen VLM (Qwen3-VL-8B) to generate bounding boxes with confidence scores, and only requires manual correction for low-confidence boxes. The segmentation backbone is a diffusion model that receives multiple guidance cues (spatial features, wavelet features, VLM embeddings) weighted by a Guidance Conditional Dirichlet Diffusion Strategy (GCDS), which samples candidate weighting strategies from a learned Dirichlet distribution. A Structure Advantage Learning module then scores these candidates using a spatial quality score (point-level agreement) and a temporal semantic score (VLM-based ranking across frames), selecting the best prediction via normalized advantage weights. The authors also contribute a new dataset (ULBV400+) with 433 annotated breast lesion ultrasound videos. Experiments on 3 datasets show consistent improvements over recent weakly supervised methods.
- 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.New dataset. A 433-video breast lesion ultrasound dataset is a meaningful contribution. Ultrasound video segmentation datasets are scarce, and this fills a real gap. The promise to release the dataset and code adds value. 2.Evaluation is solid. Three datasets (ULBV400+, ERUS-10K, DPSTT), all with standard deviations reported. This is a good statistical practice to prove it’s not a chance. Consistent improvements across all three datasets strengthen the claims. 3.Practical annotation reduction. Table 5 shows the method saves ~60–70% of manual annotation effort compared to existing methods that require 100% manual bounding boxes. This is a concrete, measurable practical benefit. 4.Interesting core idea. Using a VLM as both an annotation generator and a quality evaluator for candidate predictions is a creative approach. The idea of sampling multiple candidate strategies and letting the VLM rank them to guide learning is genuinely different from standard pseudo-label or consistency-based methods.
- 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 RL framing is misleading. The paper repeatedly draws connections to reinforcement learning, but what’s actually happening is just score normalization of quality scores across candidate predictions (Eq. 9). There’s no policy, no environment, no reward signal in the RL sense. The advantage weight A^n is a static score normalization, not a learned advantage function. Calling this “Structure Advantage Learning” and framing it as RL-inspired overstates what the method actually does. The contribution is still valid, it’s just candidate ranking with VLM feedback, but the framing should be more honest. 2.Computational cost of VLM-in-the-loop training is not addressed. The temporal semantic score (Eq. 8) requires prompting Qwen3-VL-8B to compare N candidate predictions per frame, across T frames, during training. This is only a training-time cost; inference doesn’t need the VLM; but it still seems expensive and no training time or cost comparison with baselines is reported. For a paper that emphasizes “efficiency” in its title, this should be quantified. 3.The method is very complex with many moving parts. Diffusion backbone + wavelet features + BioMedCLIP embeddings + Dirichlet distribution sampling + spatial quality scoring + VLM temporal semantic scoring + advantage-weighted loss. It’s hard to tell what’s actually necessary. The ablation (Table 4) adds components progressively (M1→M2→M3→M4→Ours) which is helpful, but doesn’t test whether simpler alternatives to individual components (e.g., learned attention instead of GCDS) could achieve comparable results. A simpler method might get most of the gains. 4.No comparison with fully supervised methods. The paper only compares against other weakly supervised methods. A fully supervised upper bound using the same backbone and full pixel-level annotations would help contextualize how much performance is being left on the table. Without this, it’s hard to judge whether the remaining gap justifies the complexity of the approach. Minor Weaknesses 5.The paper claims compatibility with “different diffusion models and vision language models” (Section 3.1, italicized in the paper) but only tests one of each. This claim is unsupported without experiments. 6.Dataset details (ethical review, hospital information) are deferred to after review. While understandable for anonymity, this makes it harder to assess data quality, patient diversity, and annotation protocol.
Clarity and Presentation The paper is reasonably well-written but the method section is dense. Figure 1 tries to show the entire pipeline in one diagram and ends up being hard to parse, the text in the figure is too small to read at print resolution. The math is mostly clear, though the notation gets heavy with the Dirichlet formulation. The RL terminology adds confusion rather than clarity, as discussed in 1.The experimental section is clean and the tables are easy to read.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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
1.What is the training time compared to the baselines? What is the inference time per frame compared to baselines? 2.How does the method compare to a fully supervised baseline using the same diffusion backbone with pixel-level annotations? 3.How sensitive is performance to the confidence threshold (7) for manual correction? 4.Would a simple learned attention over the three feature streams replace GCDS with comparable performance? 5.How does the method perform with a different VLM (e.g., a smaller or larger model)?
- 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 makes two genuine contributions: a new ultrasound video dataset (433 videos) and a creative approach to weakly supervised segmentation that uses a VLM as both annotator and quality evaluator. The evaluation is above average, three datasets, standard deviations reported, consistent improvements, and a practical annotation reduction analysis. The core idea of sampling candidate strategies and using VLM feedback to select the best one is interesting and different from standard approaches. However, the RL framing is overstated, the method is complex with many components whose individual necessity isn’t fully justified, and the computational cost of VLM-in-the-loop training is never addressed despite the paper’s emphasis on efficiency. The missing fully supervised comparison also makes it hard to contextualize the results. On balance, the dataset contribution, the solid evaluation, and the novelty of the VLM-guided approach outweigh the concerns. I lean toward acceptance, conditional on the authors clarifying computational costs and toning down the RL claims in the rebuttal.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
My pre-rebuttal recommendation was Weak Accept. The rebuttal strengthens the case: the RL framing concern is acknowledged and toned down, the efficiency claim is clarified as annotation efficiency (with concrete numbers), and the method complexity concern is addressed with a GCDS-vs-attention comparison showing the module contributes meaningfully. The fully supervised upper bound puts the results in proper context, the method is reasonably close while using ~40% of manual annotations. Combined with the original strengths, this paper makes a clear contribution to weakly supervised medical video segmentation. I maintain my recommendation to accept.
Review #3
- Please describe the contribution of the paper
The authors proposed a novel weakly supervised learning framework that integrates Vision-Language Models (VLMs), relative advantage comparison, and the Mixed Dirichlet distribution to significantly reduce manual annotation efforts. They promised to release a new large-scale dataset, ULBV400+ (433 videos), which will positively contribute to the medical video analysis community.
- 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.Adapting the concept of relative advantage comparison (conceptually from RL/GRPO) into a weakly supervised segmentation pipeline is a innovative strategy. 2.The application of the Mixed Dirichlet distribution aligns well with the adaptive fusion of combinatorial guidance cues. 3.The framework achieves a substantial and practically valuable reduction in the manual annotation burden for ultrasound video segmentation. Experimental validation is extensive across multiple diverse datasets.
- 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.Figure 1 is overly dense and hard to map to formulas. Several variables are undefined or inconsistent (e.g., $n$ in Eq. 2, $\epsilon$ in Eq. 9). The extraction process for the three Guidance Cues ($F, F^W, F^{VLM}$) lacks technical detail. 2.Table 5 raises concerns regarding its practical significance. To my knowledge, methods like BBTP, FSRM, and MonoBox are purely weakly supervised segmentation algorithms. I would be much more interested to see how these baseline models perform when fed the exact same annotations generated by VLM (partially manual refined). Additionally, the authors may consider exploring how performance fluctuates as the volume of manual annotation (dictated by the required annotation threshold) varies. 3.Were the weakly supervised baseline methods trained using the original ground-truth annotations provided by the dataset or the coarse annotations generated by the VLM? If the former, does the proposed ULBV400+ dataset also contain 100% dense manual annotations to support this? This may need be confirmed.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 integration of VLM evaluation, relative advantage learning, and Dirichlet-based fusion is highly novel and practically reduces clinical annotation costs. However, the manuscript requires revisions to clarify the baseline training data and fix the mathematical notations. Additionally, a further exploration of how varying ratios of required manual annotation impact model performance would significantly strengthen the work. If these evaluation and presentation issues are resolved, I recommend an acceptance for this paper.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The paper presents a meaningful dataset contribution and a novel weakly supervised segmentation framework that combines VLM-guided annotation, candidate quality ranking, and adaptive multi-cue fusion. The experimental results are generally strong, showing consistent improvements across three datasets together with a substantial reduction in manual annotation effort.
My main concerns were related to method clarity, the fairness of the supervision/comparison protocol, the reliance on VLM-generated annotations, and the somewhat overstated RL-inspired framing. The rebuttal addressed these points reasonably well by clarifying the role of the VLM supervision, providing additional explanation of threshold sensitivity and model compatibility, and refining the intended interpretation of efficiency and relative-quality learning. While the manuscript would still benefit from clearer presentation and a more explicit description of the supervision protocol, I believe the overall contribution is sufficient for acceptance.
Author Feedback
We appreciate all reviewers’ (R2,R3,R4) favorable comments on our method novelty (R2,R3,R4), and the large-scale dataset contribution (R2,R3,R4). Below, we clarify all the reviewer issues. R2 1[Clarity]Thanks for your advice. We have simplified Fig.1 by reducing redundant text and reorganizing the workflow with a clearer visual hierarchy. We also revised our description with a step-by-step explanation for VLM-assisted annotation, GCDS generating multiple candidate masks, and SAL relative-quality ranking among candidates. 2[Capability]Thanks, our method does not directly use raw Qwen-BBox as supervision, but uses 2 protocols to balance robustness and annotation cost, reducing ~60% manual workload while outperforming fully manual-BBox SOTAs: 1) \lambda (Eq.1) controls a reliable sample scale from Qwen-BBox for reliable point supervision, rather than directly training on Qwen-BBox; 2) Only high-confidence BBox (~60%,score>=7) are used, while low-confidence BBox need further manual correction. We additionally evaluated SAM2-style strategy (MedSAM2), yielding 67.88/78.44/81.71/81.45(Jac/Dice/Prec/Rec) on ULBV400, showing the flexibility for different pseudo-label approaches. R3 1[Framing]Thanks, we have carefully revised the descriptions as VLM-guided relative quality ranking to avoid misleading. 2[Efficiency]Thanks, we clarify that the term “efficient” refers to VLM annotation efficiency, i.e., reducing ~60% manual annotation cost while achieving competitive performance, rather than model speed. The average train/infer time (s/frame) of Ours, FSRM, MonoBox, BBTP, WeakMedSAM, HCL-HRL, Box2Mask are 2.448/0.047, 1.357/0.031, 2.896/0.061, 0.780/0.031, 0.445/0.040, 0.758/0.034, 0.757/0.044.3[Complexity]We have replaced GCDS with standard attention on Tab.4 DPSTT dataset, the Jac/Dice/Prec/Rec are 68.94/78.66/84.37/77.97, showing the effectiveness of our module. 4[Upper]The ULBV400 upper bound (Jac/Dice/Prec/Rec) is 72.57/82.84/86.90/81.64.Our method is closer to the upper bound than SOTAs. 5[Compatibility]We validated the compatibility on ULBV400.Qwen3-VL-4B achieved 66.29/76.90/80.72/80.06, the larger 30B achieved 69.04/79.49/83.75/81.02 (Jac/Dice/Prec/Rec). Replacing the current backbone with MedSegDiff yielded 66.97/77.50/80.88/81.19, supporting our flexibility and compatibility. 6[Dataset]Videos are annotated with ITK-SNAP by 3 experienced clinicians. Each frame was independently annotated 3 times, and GT was generated via pixel-wise majority voting to reduce observer variation and enhance reliability. Due to the double-blind policy, detailed hospital and ethical information will be released upon publication. 7[Clarity]We have detailedly revised the method section and simplified Fig.1 to provide a more intuitive, step-by-step explanation and enhance the readability. We also clarified the relative quality ranking operation rather than a full RL pipeline. 8[Threshold]On ULBV400, the Jac/Dice/Prec/Rec with confidence threshold 5 are 66.98/77.33/79.96/80.89, while threshold 9 achieves 70.03/80.25/84.78/81.13.Ours(7) could balance the annotation cost and supervision quality. R4 1[Clarity] 1)Thanks for your advice. We have simplified Fig.1 and revised the equations to improve readability and variable consistency. 2) F is the used origin TBGDiff backbone encoder feature; F^W is generated by standard Haar wavelet high-frequency components; F^VLM is obtained from frozen BioMedCLIP embeddings and aligned the dimension with F through convolution. 2[Annotation]BBTP, FSRM, MonoBox achieved 45.65/56.58/63.61/61.67, 46.36/59.09/58.94/73.21, 52.43/63.87/76.24/62.21 (Jac/Dice/Prec/Rec), showing performance drops compared with fully manual BBox. On ULBV400, confidence threshold 5 achieved 66.98/77.33/79.96/80.89, while threshold 9 achieved 70.03/80.25/84.78/81.13.Higher-confidence pseudo labels provide better performance. 3[Supervision]Yes, all comparisons are trained using GT-BBox. Our dataset contains 100% pixel-level expert annotation.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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.
This paper received mixed review scores of weak reject and weak accept. The major strengths summarized by reviewers include the dataset contribution, the annotation cost reduction by using frozen VLMs, two module designs based on RL and mixed Dirichlet distribution. The critical concerns raised by reviewers include unclear method description, the over-reliance and sensitivity on VLM, complex method without full justification for each module, incomplete validation and comparisons, along with other comments detailed in individual reviews.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After the rebuttal, three reviewers gave the acceptance decision. The AC concurs with this recommendation and the final submission should be revised based on the reviews and rebuttal accordingly.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
I recommend rejection. The main weakness is that the paper does not convincingly justify why a language or vision-language model is necessary for this problem. The proposed VLM is mainly used to generate bounding boxes/confidence scores and to rank candidate masks, but these are essentially visual localization and mask-quality estimation tasks. The paper does not establish that language reasoning provides a principled advantage over simpler vision-only alternatives, such as a detector/segmenter teacher, SAM/MedSAM-style promptable model, temporal tracking, visual foundation features, or conventional boundary/temporal consistency scores.
The experimental evidence also does not isolate the contribution of language. The ablation only shows incremental gains from adding VLM-related features and the proposed loss, but lacks key controls: an image-only foundation encoder, a non-language pseudo-box generator, a learned visual quality scorer, prompt-free VLM features, different prompts/models, or expert-validated mask ranking. Therefore, the reported improvement may come from using a strong black-box pseudo-labeler rather than from any necessary language-based reasoning.
The VLM-based temporal semantic score is particularly under-validated. The paper asks the VLM to compare candidate masks and provide scores/reasoning, but does not show that these scores are calibrated, reproducible, robust to prompt variation, or consistently correlated with expert mask quality. In a medical segmentation setting, using such unverified language-generated scores as training supervision is not sufficiently reliable.
Overall, while the dataset and empirical results are potentially useful, the core methodological claim is not adequately supported. A stronger paper would need to clearly demonstrate why language/VLM supervision is needed for ultrasound video segmentation and provide rigorous comparisons against simpler visual or geometric supervision mechanisms.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
I have read the manuscript, review comments, rebuttal letter. All three reviewers recommend acceptance (after rebuttal). The responses addressed the concerns about method description.
