List of Papers Browse by Subject Areas Author List
Abstract
Tuberculosis (TB) remains the leading cause of death from an infectious agent, disproportionately affecting low-resource settings where diagnostic infrastructure is limited. Point-of-care lung ultrasound (POCUS) offers a portable, affordable modality for TB screening, but clinical diagnosis requires integrating heterogeneous evidence across multiple lung zones into a single patient-level decision, a challenge not addressed by existing single-video AI systems. We propose Hierarchical Multi-Video Multiple Instance Learning (HMV-MIL), enabling patient-level TB detection directly from routine multi-video ultrasound acquisitions under weak supervision. HMV-MIL aggregates diagnostically informative information within and across videos to produce clinically meaningful patient-level predictions, while auxiliary pathology supervision aligns learned features with recognizable pulmonary findings. On a prospective cohort of 504 patients acquired under a 14-site POCUS protocol, HMV-MIL achieves 0.89 AUC, 0.82 sensitivity, and 0.75 specificity, outperforming conventional spatiotemporal baselines by 0.15–0.35 AUC and approaching WHO triage performance targets. These results demonstrate the feasibility of automated TB triage using frontline-acquired ultrasound data, supporting decentralized diagnostic assessment in settings where conventional testing remains inaccessible.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3381_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/tbrokowski/Ultr-AI-HMV-MIL/tree/main
Link to the Dataset(s)
N/A
BibTex
@InProceedings{BroTre_Hierarchical_MICCAI2026,
author = { Brokowski, Trevor AND Barbiere, Mattia Nino AND Suttels, Véronique AND Onofrey, John A. AND Hartley, Mary-Anne},
title = { { Hierarchical Multi-Video Multiple Instance Learning: AI-Interpreted Point-of-Care Ultrasound for TB Detection } },
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 evaluates the efficacy of a Hierarchical Multi-Video Multiple Instance Learning (HMV-MIL) framework for Tuberculosis (TB) detection. The authors utilize an internal dataset of 14-site bilateral point-of-care lung ultrasound (POCUS) videos, demonstrating how a multi-video aggregation strategy can improve diagnostic accuracy in a screening context.
- 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 study addresses TB screening using accessible handheld POCUS technology.
- Problem Formulation: The authors successfully frame the 14-site POCUS protocol as a multi-video multiple instance learning problem, which naturally aligns with the clinical workflow of examining multiple anatomical zones.
- Methodological Integration: A notable strength is the demonstration of performance gains achieved by aggregating auxiliary pathology signatures, which suggests the model is capturing relevant clinical features beyond simple global labels.
- Performance: The HMV-MIL approach shows superior performance over single-video baseline methods in TB classification tasks.
- 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.
Major Comments The conceptual novelty of the Hierarchical Multi-Video Multiple Instance Learning (HMV-MIL) architecture is not sufficiently established. The proposed framework bears significant similarities to existing multi-view evaluation models, such as the view-primed vision-language systems used in echocardiography [1]. The authors should explicitly clarify what distinguishes HMV-MIL from these prior multi-view architectures. The comparison between the proposed multi-video model and single-video baselines lacks clarity. It is unclear which specific video from the 14-site protocol was provided to the single-view models, or if they were trained on all videos independently. This ambiguity makes it difficult to determine if the performance gain is due to the architecture or simply the volume of data processed. There is a concerning discrepancy regarding the reported results. The manuscript claims a 0.15 AUC improvement, yet this is not supported by the data in Table 1.Furthermore, the closest performing model, DINOv3-Bert, is notably absent from the discussion in Section 5.1 (Architecture Comparison) and the abstract. Minor Comments In Section 3.1, the authors define the video dimensions as (Ti×3×224×224). It is unclear why three channels are maintained for monochromatic ultrasound data. The description of the HMV-MIL architecture in Section 3.1 is presented as a dense block of text, making the technical workflow difficult to follow. This section would benefit from a more structured explanation and connecting the text with Fig. 1. The performance gap between the baselines and the proposed model (92M parameters) may be attributed to model capacity rather than architectural design. The authors used ResNet-18 as a baseline; comparing HMV-MIL against higher-capacity models (e. g. , ResNet-50 or ResNet-101) would provide a more rigorous evaluation. [1] Vukadinovic, M. , Chiu, IM. , Tang, X. et al. Comprehensive echocardiogram evaluation with view primed vision language AI. Nature 650, 970–977 (2026).
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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?
The recommendation is primarily driven by two factors: a lack of distinct methodological innovation over existing multi-view frameworks and significant inconsistencies between the text claims and the reported results in the tables. While the application is clinically valuable, the paper lacks rigorous comparisons and convincing analysis.
- 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.
The authors addressed the key concerns and comments in the rebuttal
Review #2
- Please describe the contribution of the paper
The paper proposes a hierarchical multi video approach to TB detection using point of care ultrasound (POCUS) data. This is the first approach to handle POCUS data with 10-15 videos for a single patient which goes beyond the conventional single video / image approach. Leveraging POCUS data for TB detection with vision transformer based approach is notable contribution
- 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.
To achieve TB detection, the paper proposes a novel intra video attention and inter video aggregation which is the core novelty. The multi head MIL (multi Instance Learning) approach captures spatial heterogeneity of the dataset. The hierarchical design genuinely matches clinical practice. TB diagnosis requires synthesizing evidence across bilateral thoracic zones, and no prior system addressed this from raw video. The auxiliary pathology heads add interpretability that most end-to-end systems lack. Fine tuning of vision transformer for the task is also commendable.
- 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.
Single-site evaluation (one hospital in Benin) is a serious limitation — the authors acknowledge it but the paper’s claims about decentralized, global TB screening are difficult to support without multi-country validation. The robustness and performance of the model to unseen data is not validated since the test data is held out set from the training data.There is no analysis on false positive, false negative and TB cases that the model misses. TB varies significantly with co-morbidities, malnutrition, age, demographic and geography, however, the impact of these parameters on the accuracy of the model is not analysed. The paper claims that the system is developed for “resource limited settings” however there are no discussions related the deployment of the model in 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 submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
Currently, the performance of fine tuned ViT is compared against CNN based models and video transformers. However, to understand the significance of fine tuning, the performance has to be compared with vision foundational models for medical domain for TB. Ultrasound is notoriously operator independent, the impact in performance of model when image with poor quality is being exposed to the model has to be analysed. Further, what is minimum skill level of operator required when stated as - “minimally trained operator” has to be elaborated more.
- 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 novelty of the approach to use POCUS data with intra video attention and inter video aggregation for TB detection. This technique will inspire more research in this direction which will subsequently enable TB detection in resource limited environments.
- 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
The paper presents a model architecture for patient-level tuberculosis (TB) classification using multiple lung ultrasound videos within a multi-instance learning framework. The proposed approach first learns pathology-related representations from individual standardized scan videos and subsequently aggregates these representations to produce a patient-level classification.
- 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.Clear organization and presentation The manuscript is well structured and clearly written, making it easy to follow the proposed methodology and experimental design.
2.Novel and timely application The use of lung ultrasound for TB classification is an emerging research area, with most prior work focusing on chest radiographs. The proposed patient-level classification framework, which aggregates information across multiple standardized ultrasound videos, appears to be a novel contribution in this space.
3.Thoughtful validation strategy The authors evaluate their framework across different backbone architectures and include ablation studies to assess key components of the model. This provides some insight into the behavior of the proposed approach and demonstrates an effort to systematically analyze design choices.
- 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.Unclear benefit relative to simpler approaches While the proposed framework is novel, it is not yet clear whether the added complexity is justified. Prior work on TB classification from individual ultrasound frames appears to achieve stronger performance, in some cases surpassing the WHO-reported benchmarks cited by the authors. Including a direct comparison to simpler 2D models would help clarify whether the multi-instance, multi-video framework provides a meaningful advantage.
2.Missing or incomplete citations Several methods referenced in the introduction, including qXR and CAD4TB, are not accompanied by citations. In addition, the WHO guidelines referenced in the paper should be properly cited to support the reported benchmarks and provide appropriate context.
3.Interpretation of ablation results may be overstated The conclusions drawn from the ablation study do not appear to be fully supported by the reported results. In particular:
- Variants using attention-based key frame selection, uniform sampling, and mean pooling show relatively small differences in performance, suggesting that the attention mechanism may not play as strong a role as claimed.
- Increasing the number of selected key frames appears to degrade performance, which does not clearly support the interpretation that the attention mechanism is effectively identifying the most informative frames.
- Mean pooling performs comparably or better than key-frame selection, which may indicate that simpler aggregation strategies are sufficient.
- Removing pathology supervision results in only a minimal performance drop, raising further questions about the necessity of this component.
Overall, these findings suggest that some components of the proposed architecture may not contribute as substantially as implied, and the interpretation of these results could be revisited for greater balance.
- 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.
- 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
There appear to be issues with several references in the bibliography. Some cited works either do not seem to exist or have inconsistencies between author lists and titles. The references with potential issues include [2, 3, 4, 11, 18], and reference [5] appears to duplicate reference [4]. It would be important to carefully verify and correct the reference list to ensure accuracy and credibility.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses an interesting and emerging problem and proposes a novel framework. However, there are several concerns that limit confidence in the current version of the work. In particular, the lack of comparison to simpler baseline methods makes it difficult to assess whether the added model complexity is warranted. Additionally, the interpretation of the ablation results appears to emphasize conclusions that are not fully supported by the data. Finally, the presence of questionable or potentially incorrect references raises concerns about the reliability of the manuscript. Addressing these issues 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.
Reject
- [Post rebuttal] Please justify your final decision from above.
The authors have addressed many of the comments by promising to make changes for the final manuscript but the number of changes would need to be significant and it is unlikely that they have the space to address them all sufficiently.
Author Feedback
We thank R1, R2, R3, and Meta for their careful comments; all clarifications will be incorporated in the final manuscript.
Problem formulation and multi-view comparison (Meta, R1):TB caused 1.23 million deaths in 2024, and AI-POCUS offers a promising path to point-of-care detection. Yet no prior work has built a deployable TB detection system from raw, multi-site POCUS video. HMV-MIL’s novelty lies in defining this previously unformulated problem space and constructing an architecture purpose-built for it. Existing multi-view AI-systems, particularly in echocardiography, rely on standardized acquisition planes, canonical views, and structured exams. TB lung POCUS has none of these: no canonical TB views, no frame- or site-level labels, and no single sufficient ultrasound sign. Diagnostic evidence is weak, sparse, and distributed across bilateral sites. HMV-MIL is the first framework to formulate this as hierarchical multi-video MIL, intra-video attention identifies informative frames under temporal noise; inter-video gated MIL integrates anatomically distributed findings under patient-only weak supervision. This is structurally and operationally distinct from view-primed architectures.
AUC discrepancy and DINOv3-BERT (Meta, R1, R2, R3): We will clarify the paper: the 0.15-0.35 AUC improvement refers exclusively to conventional spatiotemporal baselines. We will revise Section 5.1 to discuss DINOv3-BERT (418.6M parameters) achieves 0.85 AUC, 0.04 below HMV-MIL (4.5x smaller). A larger backbone does not match our purpose-built design, disproving the capacity hypothesis.
Baseline protocol (Meta, R1): All baselines used full patient-level data across all 14 sites. Section 4.2 will be revised to remove ambiguity around “single-video” terminology. Comparison to prior work (R3). The prior 0.93 AUC (Suttels et al.) requires trained sonographers to curate frames before inference, selecting the clearest, most diagnostically informative images, simplifying the input distribution and presupposing the expert knowledge the system is meant to replace, making it undeployable at scale in resource-limited settings. HMV-MIL achieves 0.89 AUC end-to-end on raw, operator-acquired video, eliminating expert curation entirely. The residual gap represents the cost of removing the expert from the loop, a deliberate trade-off for a fully autonomous system.
Ablation interpretation (Meta, R2, R3): Modest 0.01-0.02 AUC differences are expected near clinically relevant performance thresholds, where gains are increasingly constrained by noise and inter-patient heterogeneity, these results reflect robustness, not insensitivity to design, all ablations degraded performance, confirming each component contributes. Pathology supervision’s primary value is clinical interpretability, per-site, human-readable outputs for auditing, and we will frame it as such.
Deployment and operator training (Meta, R2): “Minimally trained operator” refers to frontline workers trained over ~20 hours across 4 days. HMV-MIL has been converted to TensorFlow.js (344 MB) for consumer-grade hardware. We will add discussion of acquisition quality sensitivity and quality-control thresholds, and will temper global deployment claims pending multi-site validation.
False positives/negatives (Meta, R2): At 82% sensitivity/75% specificity, the model’s FN rate (~18%) reflects the cost of removing expert curation. For screening deployment, the operating threshold can be shifted toward higher sensitivity (e.g., ≥90% sensitivity is achievable at reduced specificity, per the ROC curve). We will add a confusion matrix and per-fold FP/FN breakdown to the supplementary.
References and citations (R1, R3): We recognize that reference errors undermine manuscript credibility and will have verified and addressed all concerns in the final manuscript. We will also restructure Section 3.1 for clarity. Code and full implementation details will be released upon acceptance.
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 presents an ultrasound-based approach for TB screening in which the 14-site (as in sites on the patient anatomy) image acquisition protocol is posed as a multi-video multiple instance learning problem. The work has many strengths including the use of intra-video attention and inter-video aggregation design, which matches clinical practice. It is recommended that the authors clarify these points in order to improve the quality of the manuscript: Explicitly clarify what distinguishes the HMV-MIL architecture from existing multi-view frameworks, such as vision-language systems used in echocardiography. Address the major concerning discrepancy where the claimed 0.15 AUC improvement is not supported by Table 1, and explain why the closest performing model (DINOv3-Bert) is omitted from the discussion. Clarify the comparison with single-video baselines, specifically identifying which video from the 14-site protocol was used for single-view training. Elaborate on the meaning of “minimally trained operator” and provide discussion related to the model’s real-time deployment in “resource limited settings.” Clarify the absence of or add the analysis on false positives, false negatives on model accuracy.
- 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 authors have answered the key questions satisfactorily.
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.
This paper addresses an important and clinically relevant problem by formulating multi-site point-of-care ultrasound for tuberculosis detection as a hierarchical multi-video multiple instance learning problem. The proposed framework is well aligned with clinical practice, combining intra-video attention and inter-video aggregation to integrate evidence across multiple anatomical sites. The work is timely, particularly given the need for scalable TB screening solutions in resource-limited settings.
The reviewers highlighted the novelty of the problem formulation, clinically meaningful design of the hierarchical aggregation strategy, and strong empirical performance. The rebuttal successfully clarified key concerns regarding the comparison protocol, reported performance gains, distinction from existing multi-view architectures, and intended deployment setting. Some limitations remain, including single-site validation, limited analysis of failure cases and subgroup performance, and need for stronger external validation. Overall, the paper presents a meaningful and well-executed contribution that is likely to stimulate further research in ultrasound-based TB screening.
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.
The rebuttal clarified the novelty of the hierarchical multi-video MIL formulation, resolved questions regarding reported performance, baseline comparisons, deployment considerations, and error analysis. The remaining concerns primarily relate to presentation and scope rather than fundamental methodological flaws. Given the clinical importance of the problem, the strong empirical results, and the positive reviewer consensus after rebuttal, I recommend acceptance.
