Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Point-of-care lung ultrasound has emerged as a valuable bedside tool for diagnosing pulmonary congestion, which is often identified by the presence of hyperechoic reverberation artifacts called B-lines. However, manual B-line quantification is limited by substantial inter-operator variability and the requirement for specialized technical expertise. To address these challenges, we propose a two-stage segmentation-guided transformer framework to automate B-line severity scoring. This approach provides a consistent and reproducible assessment of pulmonary pathology directly at the point of care. On a curated clinical dataset, the proposed method significantly outperforms the baseline, improving case-level quadratic weighted kappa from 0.57 to 0.74 (difference = 0.16; 95% CI: 0.04–0.30; p=0.003). The F1 score similarly increases from 0.59 to 0.75 (p=0.003). The segmentation module achieves a pixel-level F1 score of 63.1% and a frame-level B-line counting F1 score of 66.7%. The model improves the accuracy and consistency of automated B-line scoring, an important step towards standardized bedside assessment of pulmonary congestion.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5697_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Fahim-F/AI-Driven-B-Line-Detection

Link to the Dataset(s)

https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/HFTV7D

BibTex

@InProceedings{FooFah_AIDriven_MICCAI2026,
        author = { Fooladgar, Fahimeh AND Krishnaswamy, Deepa AND Ungi, Tamas AND Dinh, Viet AND Jin, Mike AND Alves, Matheus AND Schissel, Caroline AND Puducheri, Shreyas AND Misawa, Shuhei AND Duhaime, Erik AND Hallisey, Stephen AND Duran Mendicuti, Alejandra AND Duggan, Nicole AND Abolmaesumi, Purang AND Harrison, Nicholas AND Goldsmith, Andrew AND Kapur, Tina},
        title = { { AI-Driven Pulmonary Congestion Assessment for Lung Ultrasound via Segmentation-Guided Transformers } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16886},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes a two-stage framework that combines frame-level B-line segmentation with transformer-based temporal aggregation to predict clip-level pulmonary congestion severity from lung ultrasound videos.

  • 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 problem is clinically relevant, focusing on automated B‑line assessment for pulmonary congestion. 2.The overall pipeline is complete, covering segmentation, temporal modeling, and case-level prediction. 3.Consistent improvements over the provided baselines are demonstrated, with ablation studies and qualitative visualizations supporting the design.

  • 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 novelty is limited, as the method mainly combines existing components (segmentation + transformer) in an incremental manner. 2.The experimental comparison is not strictly fair, as the baseline uses max pooling over counts while the proposed method employs feature-level learned aggregation. 3.The experimental comparison is not fully aligned with the claimed task. Given that the primary goal is clip‑ or case‑level pulmonary congestion severity scoring, comparisons against segmentation baselines alone are insufficient. More appropriate baselines would include prior clip‑level or spatiotemporal models for B‑line assessment. 4.The problem formulation is somewhat unclear. It is not well justified why a transformer‑based aggregation is compared against a max‑based clinical heuristic, nor whether this reflects a meaningful modeling improvement.

  • 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?

    While the paper addresses an important clinical problem and demonstrates performance improvements, the experimental design makes it difficult to clearly attribute the gains to the proposed method. Due to limited novelty and the lack of well-controlled comparisons, it is uncertain whether the improvements stem from the transformer architecture itself or from the change in aggregation strategy. In addition, the absence of task-level state-of-the-art baselines further limits the strength of the conclusions.

  • 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

    The paper proposes a two-stage, segmentation-guided transformer framework for automated B-line severity scoring in lung ultrasound, combining spatial localization and temporal modeling. Specifically, a SegFormer-based segmentation model enhanced with Explicit Visual Prompting (EVP) generates frame-level B-line masks and embeddings, which are then aggregated via a transformer to capture inter-frame temporal dynamics and predict clip- and case-level severity scores. This design addresses the common trade-off between spatial interpretability (segmentation-based methods) and temporal reasoning (clip-level models), and demonstrates improved performance over prior approaches across multiple evaluation levels, highlighting the benefit of integrating segmentation-derived features with sequence modeling for clinically relevant ultrasound analysis.

  • 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.
    • Integration of spatial and temporal modeling: Combines segmentation (fine-grained localization) with a transformer-based temporal aggregation, addressing a known limitation of prior works that treat these aspects separately.
    • Segmentation-guided representation learning: the method uses embeddings derived from segmentation outputs rather than raw images, which is a meaningful design choice to enforce clinically interpretable features.
    • Efficient domain adaptation via EVP: the paper introduces Explicit Visual Prompting to adapt a pretrained transformer to ultrasound data without full fine-tuning, reducing trainable parameters while capturing high-frequency B-line patterns.
    • Comprehensive multi-level evaluation: the authors evaluate performance at pixel, frame, clip, and case-levels, providing a thorough validation of the method’s effectiveness.
    • Strong empirical performance with statistical validation: this work demonstrates significant improvements over baselines (e.g., QWK +0.16 with p=0.003) and includes confidence intervals and statistical testing.
    • Clinically relevant formulation: Aligns prediction targets (max B-line count per clip, case-level aggregation) with real clinical workflows.
  • 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.
    • Limited methodological novelty relative to prior hybrid approaches: Combining segmentation with temporal modeling has been explored in related forms (e.g., spatiotemporal attention models for LUS [Kerdegari et al., 2021] and transformer-based segmentation like TransUNet/TransBound-UNet [Abbasi et al., 2025]), making the overall framework more incremental than fundamentally new.
    • Dependence on dense annotations: The approach relies on pixel-level B-line annotations, which are costly and limit scalability, a known issue compared to detection-based methods (e.g., YOLO-based approaches [Bottino et al., 2025]).
    • Marginal gains from temporal modeling not deeply analyzed: while improvements are shown, the contribution of the transformer over simpler aggregation strategies (e.g., max/mean pooling) is not fully studied beyond ablations.
    • Segmentation performance remains moderate: Pixel-level F1 (~63%) suggests imperfect localization, which may propagate errors to downstream classification.
    • Lack of comparison with recent foundation-model-based approaches: Methods like SAM-based adaptations (e.g., SAM-Adapter [Chen et al., 2023]) are included, but broader comparison with emerging large-scale medical vision models is limited.
    • Single-dataset evaluation limits generalizability: experiments are conducted on a single dataset from one healthcare network, raising concerns about robustness across devices, institutions, and acquisition protocols.
  • 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 paper addresses a clinically relevant problem and presents a well-designed, coherent framework that effectively combines segmentation and temporal modeling, leading to consistent and statistically significant improvements over strong baselines across multiple evaluation levels. The use of segmentation-guided embeddings and efficient domain adaptation via explicit visual prompting are sensible and practically motivated design choices, and the evaluation is relatively thorough. However, the overall methodological novelty is somewhat limited, as the combination of segmentation and temporal modeling builds on existing ideas, and the gains, while meaningful, are not transformative. Additionally, the reliance on a single dataset, and moderate segmentation performance, weaken generalizability. These strengths and limitations place the work slightly above the acceptance threshold.

  • 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 proposes a novel two-stage deep learning framework for automated B-line severity scoring in point-of-care lung ultrasound (LUS). Stage I utilizes a SegFormer-B4 backbone enhanced for ultrasound domain with Explicit Visual Prompting (EVP) and summed with high-frequency prompts via FFT to segment B-lines. Stage II employs a Transformer Encoder to aggregate temporal features across clips to predict clinical severity scores (0 to 3+). Evaluated on the multi-center MLSC-BEDLUS dataset, the method demonstrates a significant improvement in case-level Quadratic Weighted Kappa (QWK) from 0.57 to 0.74 compared to the non-temporal baseline.

  • 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.

    Using a novel approach that combines temporal modeling with frame level B lines segmentation for predicting severity score which led to a more accurate and clinically interpretable model.Technical innovation specifically focuses on re-projecting pre-trained ImageNet embedding to ultrasound domain summed with the High-Frequency Prompts via FFT which is a highly principled choice. Since B-lines are linear reverberation artifacts, explicitly steering the model to focus on high-frequency components is a “physics-informed” approach. Additionally use of a transformer to jointly predict severity score aligned with clinical scoring system for pulmonary congestion assessment. The authors also provide a multi-level evaluation (pixel, frame, clip, and case) and compare with 5 other methods. The use of patient-level splits and bootstrap resampling for confidence intervals ensures that the reported gains are statistically robust

  • 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 study shows that performance peaks at T=8 frames and declines for longer sequences (T=16). Given that a typical respiratory cycle in ultrasound lasts much longer than 8 frames, the model may only be capturing “micro-dynamics” rather than the full evolution of B-lines during breathing. While the model outperforms baselines, the case-level F1 (74.1%) and QWK (0.71) are lower than the clip-level counterparts. This suggests that the aggregation of multiple zones into a single patient-level score remains a challenge, likely due to the “variability in scanned lung zones” mentioned in the limitations. In Table 1, the precision/recall for some baselines appears very low compared to the proposed method. While the improvement is clear, more discussion on why standard SOTA models like AdapterSAM or Att-UNet struggled significantly on this specific dataset would be beneficial.

  • 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

    For a standard 20-30 fps ultrasound, 8 frames represent <0.5 seconds. The authors should clarify if the model is intended to capture motion or simply to provide a “temporal vote” to reduce noise. With a pixel-level F1 of 63.1%, it would be helpful to see a figure illustrating where the model fails

  • 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.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The technical innovation such as the use of Explicit Visual Prompting (EVP) with FFT high-frequency masking is domain-specific adaptation combined with transformer based temporal modeling, validated on a multi-center dataset and multi-level validation of the method. The significant boost in model performance over baselines, interpretability and clinical utility of the developed pipeline a greta contribution to the field along with detailed explanation of implementation for reproducibility.

  • 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

N/A



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.

    The paper addresses an important clinical problem in automated pulmonary congestion assessment from lung ultrasound. It provides a clear two-stage pipeline, uses patient-level splits, evaluates performance at multiple levels, and reports statistical testing. Although the method is not highly novel, MICCAI contributions do not need to be limited to entirely new algorithms; clinically meaningful applications and well-supported empirical insights can also be valuable. However, the authors should better justify the temporal modeling component, clarify generalizability beyond the current dataset, and address remaining concerns about baseline fairness.



back to top