List of Papers Browse by Subject Areas Author List
Abstract
Guidance is essential in teaching the required skills for performing highly technical obstetric ultrasound examination. These skills include selecting the appropriate keyboard functions, making correct image adjustments, and positioning the probe effectively to acquire high quality anatomical views. Existing work has largely focused on assessing image quality, without identifying where errors occur within an ultrasound video or how a trainee should correct them. We introduce FOCUS, the Fetal Obstetric Corrective UltraSound instructor, a new multimodal framework designed to learn and deliver corrective guidance for trainee sonographers using both video and text. FOCUS analyses ultrasound video sequences and provides corrective feedback indicating specific actions needed to improve technical performance. To develop this framework, we collect obstetric ultrasound data captured by trainees, annotated with corrective instruction labels, anatomical labels, and accompanying expert textual feedback. Our experiments demonstrate that FOCUS effectively learns to recognise suboptimal imaging technique and generate targeted corrective feedback. We provide the code at \url{https://github.com/hlamdouar/focus-ultrasound-guidance}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1799_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/hlamdouar/focus-ultrasound-guidance
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LamHal_FOCUS_MICCAI2026,
author = { Lamdouar, Hala AND Wang, Angela Feixue AND Guo, Xiaoqing AND Men, Qianhui AND Lander, Jayne AND Papageorghiou, Aris T. AND Noble, J. Alison},
title = { { FOCUS: Towards Fetal Obstetric Corrective UltraSound Guidance } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16893},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper presents a framework for the automated generation of corrective feedback on ultrasound trainee performance during obstetric examinations.
- 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.
It is an interesting topic and not much needed to add to the paper.
There is a detailed introduction to Obstetric ultrasound.
- 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.
Could add a bit more background and references related to the “works explored language models for ultrasound” part. As that part is most closely related to the novelty of this contribution.
Typo: “the problem in from a Human–AI” should be “the problem from a Human–AI”
Final sentence of conclusion considers “real-time. “ Since real-time has not been discussed in the manuscript previously, it is not clear how this method could be applied to real-time, especially since many steps were involved in video processing (difficult to achieve in real-time), including: extracting videos with commentary, training with the 8 correction categories, using GPU. Is it that after training is completed, these steps would not be required for use in real-time application? Please expand how this method can be applied in real-time ultrasound training, or if not perhaps don’t mention real-time.
- 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.
(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?
Main sections of the paper are well organised, good description of the processing steps. Good contribution to a new proposed method of annotating training videos.
- 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 authors have addressed the majority of the review comments.
Review #2
- Please describe the contribution of the paper
This work introduces a clinically well-motivated novel task: generating corrective feedback from trainee performance videos. In addition to defining this problem, the authors propose a corresponding methodology to address it.
- 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 addresses a significant but often overlooked problem. While prior research has focused primarily on probe positioning, this work incorporates the optimization of ultrasound parameters, making the system much more aligned with actual clinical workflows.
- The feedback signals generated by the system are highly readable and provide practical, actionable guidance for trainees.
- A multi-task training framework is proposed to effectively handle the complexities of the 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.
- The dataset used in the study is relatively small. As a result, it is difficult to fully evaluate the system’s performance and reliability in a real-world clinical environment.
- Since the corrective feedback is generated from a finite set of possibilities, it appears that a classification-based approach might suffice. The authors should clarify the specific value or necessity of incorporating a language decoder in this context.
Minor Issues
- The specific sizes of the training and test sets are not clearly stated.
- All equations should be properly numbered for easier reference.
- Please ensure consistency in subscripts. Specifically, when subscripts represent words or abbreviations rather than variables, they should be set in roman (upright) font rather than italics (e.g., use $P_{\text{ref}}$ instead of $P_{ref}$).
- Ensure consistent capitalization for paragraph headings; please use either Title Case or sentence case throughout the manuscript.
- 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?
My recommendation of Weak Accept is based on a balance between the paper’s strong clinical motivation and some remaining concerns regarding the experimental validation scale and methodological necessity.
- 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.
N/A
Review #3
- Please describe the contribution of the paper
This paper presents a pipeline for automatic feedback generation to support training in fetal ultrasound scanning. The authors propose a novel model architecture that jointly learns anatomical understanding and expert feedback, and integrates a language module to translate model outputs into human-interpretable feedback cues. The approach is evaluated on an in-house dataset consisting of seven fetal ultrasound scans.
- 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.Timely and clinically relevant problem The paper addresses automated feedback generation for medical training, which is an emerging and important area with clear clinical value. In particular, the focus on supporting training in low-resource settings highlights a meaningful and practical motivation.
2.Novel application domain The use of fetal ultrasound as the target application is both novel and impactful. The emphasis on generating comprehensive feedback, rather than focusing on a single artifact or detection task, distinguishes this work from more narrowly scoped approaches.
3.Proposed model architecture with some supporting exploration of model variants The authors introduce a custom architecture that jointly models anatomical understanding and expert feedback, coupled with a language module for interpretability. The inclusion of studies on model variants helps support some of the design choices and provides useful insight into the model components.
4.Clear motivation and generally strong presentation The manuscript is well organized and effectively communicates the motivation and broader context of the work. The writing is clear, with only minor typographical issues.
- 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.Very limited dataset size and unclear composition The evaluation is based on only seven fetal ultrasound scans, which raises concerns about robustness and generalizability. In addition, important details about the dataset are missing, including:
- the level of expertise of the practitioners (e.g., novice vs. expert),
- whether multiple operators or patients are included,
- whether scans were collected under consistent conditions. These details are necessary to properly interpret both task difficulty and model performance.
2.Potential data leakage between training, validation, and test sets Clips used for training, validation, and testing appear to be sampled from the same underlying videos. This introduces a risk of overlap in visual content across splits, which may lead to artificially inflated performance.
3.Unclear and potentially inconsistent annotation process The annotation pipeline requires further clarification. Clips are defined based on synchronized expert audio feedback, yet reported durations range from 0.04 seconds to 60 seconds. This wide variation is difficult to reconcile with realistic feedback delivery and raises concerns about how clips are segmented and temporally aligned.
4.Lack of annotation validation It is not specified whether annotations were reviewed or validated for correctness prior to being used for training and evaluation. Without such validation, it is difficult to assess the reliability of the labels and, by extension, the reported performance.
5.Class imbalance not adequately addressed in evaluation The dataset appears to be imbalanced, but the evaluation relies primarily on aggregate metrics that do not account for this. Reporting balanced accuracy and/or per-class metrics would provide a more complete picture of model performance.
6.Selective qualitative results The qualitative examples presented focus only on successful cases. Including failure cases would provide a more balanced and realistic view of model behavior and limitations.
7.Limited comparison to alternative methods The evaluation focuses only on variants of the proposed model. While this is somewhat understandable given the novelty of the task, including comparisons to related approaches where possible would help contextualize the results.
- 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
1.There is a formatting error on the final page where text extends into the margin.
2.In Figure 4, the choice of colors for highlighting words may be unintentionally misleading, as green and red are often interpreted as correct versus incorrect. A more neutral color scheme may reduce potential confusion.
3.It would also be helpful to define all evaluation metrics in the Methods section rather than introducing them in table captions, and to present them in a consistent order aligned with their appearance in the results Table 1.
- 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 work tackles an interesting and clinically relevant problem and proposes a novel methodological approach that could be of interest to the MICCAI community. However, the current evaluation is limited by the very small dataset and several ambiguities in the annotation and data collection processes. These issues make it difficult to assess the validity and generalizability of the reported results. Additionally, the presentation of qualitative results and limited discussion of potential failure modes further reduce confidence in the conclusions. Addressing these concerns would substantially strengthen the work.
- Reviewer confidence
Very confident (4)
- [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.
While the dataset is small and limits the generalizability of results, the authors have satisified my concerns regarding the standardization and validation of the annotations that were used for training.
Author Feedback
We thank the reviewers for their time and insightful feedback. We are particularly grateful that all reviewers recognized the relevance, novelty, and strong clinical motivation of our work. We address the main questions below. Typos and minor concerns will be addressed in the revised version.
[R1] on real-time: Video/comment pairs are only required to generate labeled ground truth. Once training is complete, the model operates solely on ultrasound video at inference time. Although we do not claim real-time performance of our model, the proposed framework can be deployed in such a setting, where a trained model runs in the background during scanning and provides instructive feedback.
[R2-R3] on dataset scale and reliability in real-world clinical environments: Our dataset scale reflects the practical challenges of obtaining expert annotations in obstetric ultrasound. We will explicitly acknowledge this limitation in the revised paper. Our dataset has been carefully curated and captures a variety of challenging subtasks encountered by sonographer trainees, enabling meaningful evaluation. We agree that larger-scale validation remains important and plan to extend this work by collecting a broader set of expert corrective commentary. However, we emphasize that this work is not intended for immediate clinical deployment, but rather represents an important milestone toward that goal.
[R2] On the value of incorporating a language decoder: Our motivation is twofold. First, the language model attends to visual, anatomical, and instruction embeddings, capturing richer clinical reasoning and causal context from expert feedback beyond discrete labels (e.g. L1 in Fig4, CorLM identified shadowing caused by fetal anatomy, which classification alone cannot capture). Second, based on feedback from sonographers, there is a clear preference for human-readable outputs that provide instructive guidance.
[R3] More on the dataset: 7 scans were acquired by a newly qualified obstetric sonographer across 4 participants in 2nd trimester, using a GE Voluson E8 system with C2-9D transducer, over 3 examination sessions spanning 9 months. Expert commentary was provided retrospectively by a sonographer with extensive experience in training. The videos are segmented into 228 training, 32 validation, and 64 test clips, capturing a range of challenging subtasks encountered by trainees and enabling meaningful evaluation across clinically relevant scenarios.
[R3] Annotation consistency/validation: We used WhisperX for initial transcription and timestamp, then addressed temporal misalignment due to expert reaction lag during fast probe movement. A trained researcher manually validated corrected transcripts and clip boundaries by jointly inspecting video and audio to ensure expert comments align to the correct anatomical views being described. The final dataset includes 324 curated clips with verified text–time alignment.
[R3] Risk of data leakage: The train/val/test splits were constructed manually to minimize the risk of data leakage due to overlapping visual content. Video-level splitting was infeasible as scans do not cover all anatomies and correction types; instead, subsets were balanced to include all views and categories for comprehensive evaluation.
[R3] Per-class metrics: Following reviewer’s suggestion and without introducing additional results, we will include per-class breakdown of the results reported in Table1 e.g. Ours-L1 achieves performance including GV 62.5, HG 33.3, HM 22.2, LM 81.2, SH 66.3.The potential impact of class imbalance will be added to limitations.
[R3] Failure modes: We agree. We will add incorrect anatomy/instruction examples to Fig4. [R3] Understandably limited comparison to alternative methods given novelty of the task: While we did test existing comment generation approaches, we found that they only describe anatomical content, rather than generate actionable instructions, which is the central objective of our work.
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.
The paper sits in a borderline zone: one clear accept, one weak accept, and one weak reject. All reviewers agree the problem is important and the application is novel, with strong clinical relevance.
The main concern is the dataset and evaluation. The very small number of scans, potential data leakage between splits, and unclear annotation process significantly weaken confidence in the results. These are central issues and must be addressed clearly.
There are also questions about whether the language component is necessary given the limited output space, and whether claims around real-time applicability are justified.
That said, the idea is interesting and timely, and the framework could be impactful if validated properly. A rebuttal is warranted.
The authors should focus on clarifying the dataset composition, how splits were constructed to avoid leakage, how annotations were created and validated, and why the chosen modeling approach is necessary. If these points are convincingly addressed, the paper could move into the accept range.
- 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.
The main limitation is the evaluation. The dataset contains only 7 scans from 4 participants, producing 324 clips, and this is too small to support the claims made about model performance and generalizability. The rebuttal clarifies the dataset composition and annotation process, but also confirms that train, validation, and test clips were sampled from the same small underlying scan set. Since video-level splitting was not possible, the risk of visual overlap and leakage remains a central concern.
The annotation process is better explained in the rebuttal, but still relies on retrospective commentary, automatic transcription, manual correction, and clip boundary validation by a researcher rather than independent expert validation. The language-generation component is also not fully justified given the small and finite feedback space, and the real-time applicability remains aspirational rather than demonstrated.
Meta-review #2
- 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.
The rebuttal addressed reviewers’ concerns. All three reviewers moved to accept post-rebuttal, and given the strong clinical motivation and convincing clarifications, the recommendation is Accept.
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.
This paper proposes a method for generating corrective feedback from trainee performance videos for fetal ultrasound scanning. The task is clinically well-motivated and very relevant. The main concerns of the reviewers were about the dataset size and experimental setup. These concerns have been successfully addressed by the author’s rebuttal and all reviewers agree on the acceptance of this work.