List of Papers Browse by Subject Areas Author List
Abstract
Preterm birth (PTB) prediction can enable targeted surveillance and timely intervention, yet most ultrasound-based models use a single selected transvaginal ultrasound (TVUS) frame per patient despite routine exams acquiring multiple cervical images. We formulate PTB prediction as a multiple instance learning (MIL) problem, representing each patient as a variable-sized bag of TVUS images with a single outcome label. To move beyond standard MIL aggregators that collapse a bag into a point estimate, we propose a Gaussian Mixture Model (GMM) pooling, which summarizes all images in a bag into a fixed-length representation by modeling their feature distribution. This design captures intra-patient variability. We evaluate the method on a private clinical cohort and on a public lymph node metastasis benchmark. For PTB prediction, GMM pooling improves over the instance-based model PR-AUC from 0.44 to 0.56. On the lymph node benchmark, it achieves state-of-the-art performance with 0.91 F1-score and 0.89 ROC-AUC for classification and 0.18 MAE for regression. The code is publicly available at https://github.com/HussainAlasmawi/GMM_Pooling.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4491_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/HussainAlasmawi/GMM_Pooling
Link to the Dataset(s)
https://drive.google.com/drive/folders/1HLqLkf-U7u6OrTyc1xB0jpWze-zWUs-J
BibTex
@InProceedings{AlaHus_From_MICCAI2026,
author = { Alasmawi, Hussain AND Saeed, Numan AND Said, Soha AND Yaqub, Mohammad},
title = { { From Point Estimates to Distributions: GMM Pooling for MIL in Preterm Birth Prediction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16896},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose a new method for using a Gaussian Mixture Model (GMM) to pool features from a variable number of instances (images) to create a fix-sized representation to feed into a classifier. Unlike previous approaches to this problem, the GMM approach has the advantage that it is able to capture inter-relationships between features. The method is validated on two tasks: pre-term birth prediction on a private trans-vaginal ultrasound and a public histopathology dataset.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The proposed pooling mechanism is well motivated and addresses an important theoretical limitation of other pooling methods. Min/mean/max pooling reduce instances to a single point estimate, which lacks expressiveness. Other attempts at representing a distribution over instances usually consider only marginals for each feature dimension. To my knowledge, this is the first attempt to explicitly formulate a trainable neural network architecture that captures a multidimensional distribution over instances. It seems reasonable to suppose that this would be important in solving certain types of pooling 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.
Despite its potential theoretical advantages, the results are not especially compelling. On the preterm birth task, the task used to motivate the paper, the GMM pooling performs worse than much simpler baselines (max pooling, attention pooling), depending on the metric. Slight improvements are seen on the lymph node dataset classification task (but not the more challenging regression task), though this dataset is treated like an afterthought in the paper. Though I appreciate the general technical motivation for more advanced pooling mechanisms, it is not really clear to me why this is important in the PTB application. My understanding is that in clinical practice, cervical measurements are taken and this is used to predict the risk of pre-term birth. The authors do not adequately justify why automating the cervical measurement process would be insufficient and a more sophisticated (but less interpretable) classifier is necessary, and the vital comparison to using cervical measurements to predict risk (current practice) is not evaluated experimentally. The paper is also muddied by the inclusion of the segmentation part of the model, which is quite tangential to the main methodological contribution. The use of the features from the center of a UNet is rather unconventional, and is not the natural choice for evaluating the pooling methodology. I don’t understand why the choice of pooling mechanism for the intermediate features would have any effect on the segmentation branch, so it’s not clear to me what I can draw from the “Segmentation Dice” column of Table 1.Generally, the segmentation part of the paper is discussed little and does not add much to the paper.
- 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 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
This may be a little tangential, but I am not convinced that it is correct or useful to describe this problem as multiple instance learning. It shares similarities with MIL in that it involves learning from a variably-sized bag in a permutation invariant way using a bag-level label. However, my understanding is that MIL has further assumptions: that each individual instance has a label and the label of the bag is that “ANY” operator over instance labels. Clearly when you need to model a full multidimensional distribution over feature vectors, the problem does not have these characteristics of MIL because all instances contribute to the label. This is not a critique of the formulation or the approach, merely a comment on terminology. I am also aware that others have stretched the definition of MIL (including reference [16] and going right back to ABMIL), so maybe it is too late. But I would suggest the authors consider whether this is the best terminology.
Is “K” in Figure 3 meant to be “P”?
- 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?
Major factors are lack of performance benefit of the method over simpler baselines, and innapprorpriate matching of method to the clinical applications.
Nevertheless, I would encourage the authors to pursue their interesting methodological developments.
- 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 understood the feedback from the reviewers, but their response does not change my opinion of the paper, which is that this is a technically interesting direction that I think will make a strong paper with further work, but that this particular presentation is muddled as it conflates application with method and involves extraneous components (the segmentation part), and does not convince me of the value of the approach empirically.
Review #2
- Please describe the contribution of the paper
This paper formulates PTB prediction as a multiple instance learning (MIL) problem to address the limitation of insufficient information from a single image. The authors propose a GMM pooling module to overcome the drawbacks of conventional point-estimate pooling methods. Experiments on a private PTB dataset and a public lymph node dataset demonstrate that the proposed model achieves competitive performance.
- 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.
This paper formulates PTB as a MIL problem, which better conforms to clinical practice and has strong clinical utility. The study is conducted on a cohort of clinically defined high-risk pregnant women, making the task more challenging and better reflecting the potential value of the model in real clinical settings. Experiments on multiple datasets demonstrate the generality and robustness of the proposed distribution-based pooling strategy.
- 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) This paper clarifies that density-based pooling does not explicitly model cross-dimensional dependencies in the feature space. Given that the authors propose GMM pooling to address this issue, how exactly does GMM pooling explicitly capture such dependencies? (2) The study uses a small PTB dataset and therefore applies 5-fold cross-validation. However, it is unclear whether the data splitting was performed randomly and whether an independent test set was used. (3) The differences in PR-AUC between GMM pooling, max pooling, and attention pooling are very small (below 0.01). Although the authors claim that GMM pooling is more stable, the paired t-test yields a p-value of 0.873, indicating no statistically significant difference. Thus, the observed lower standard deviation might merely reflect random fluctuation. Furthermore, GMM pooling involves two sets of attention weights: instance importance weights and soft instance-to-component responsibilities. Without clear performance gains, is attention pooling not simpler to implement and more suitable for real-world deployment? (4) This paper lacks comprehensive ablation studies to verify the contribution of key components, such as the instance importance weights, soft instance-to-component responsibilities, and learnable probes. (5) This study employs U-Net to segmentation, but what is the role of segmentation in the PTB task?
- 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 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
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?
This paper addresses a meaningful clinical task and presents a reasonable density-based MIL approach. However, experimental validation, and motivation of the proposed method need improvement.
- 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 author have made efforts to address the concerns, and I maintain my original decision.
Review #3
- Please describe the contribution of the paper
This paper attempt to improve how instance aggregation is performed in multiple instance learning (MIL) problem. Instead of collapsing the embedded features of each instance in a bag into a single-point with simple aggregator (max, mean, etc), the author proposed estimating their joint distribution with Gaussian Mixture Model (GMM) prior to that.
- 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 proposal of aggregating MIL features with GMM pooling is interesting. Conventional approaches aggregate feature of the instances in a bag into a single point. According to the author, this “create an information bottleneck by collapsing within-bag variability that may be informative for bag-level prediction”, to which I agree. Distribution-based MIL pooling aims at resolving this issue, but doesn’t capture inter-feature dependencies, which is addressed by the author in this work.
- 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 author noted that “all patients were prospectively classified as high-risk for PTB by the treating clinicians”, even though only 44 out of the 182 cases were eventually PTB. It would be nice to evaluate the method also on the low-risk cases, and I am concerned about whether there’s a sample selection bias here.
- the author noted that “the comparison between GMM and max pooling showed no significant difference” on the preterm birth dataset, and did not report similar comparison on the lymph node dataset. i am curious about the advantage of using GMM pooling over the simpler methods.
- 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
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?
While there are minor weaknesses about the paper, the proposed method remains interesting, and may find wider applications across different MIL problems within the community.
- 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.
The authors had addressed my concerns sufficiently, with a minor suggestion to rephrase the notion “all patients were prospectively classified as high-risk for PTB by the treating clinicians” to something along the line of “all patients were identified as high-risk for PTB following
by the treating clinicians". I am keeping my opinion that GMM pooling is an interesting approach and may find broader applications within our community. Therefore I would recommend accepting the paper.
Author Feedback
We sincerely thank reviewers for their comments and suggestions. We appreciate that R1 highlighted the novelty and strong theoretical motivation of GMM pooling, particularly its ability to capture feature distributions and inter-feature relationships compared with standard pooling approaches. We also appreciate that R2 noted the methodological contribution and clinical relevance, highlighting that the formulation aligns with practice and shows consistent behavior across datasets. We also note that R4 found the approach interesting and valuable, especially for mitigating information loss in conventional MIL pooling and its potential applicability to broader problems. Lack of performance advantage R1,R2,R4: We agree that GMM’s performance gains are marginal. However, GMM pooling addresses a representation limitation rather than optimizing only for predictive performance. Standard pooling collapses a bag to a single point, discarding intra-bag heterogeneity, while GMM models the instance distribution via multiple components. This is important in medical imaging, where variability is intrinsic. We observe consistently lower variance across runs, which may indicate improved robustness despite similar average performance. Weak clinical motivation R1: Yes, cervical length is a standard metric in clinical practice, but it is a rudimentary, subjective, and has limited predictive power [AJOG 2015, PMCID: PMC4675685]. Our goal is to assess whether full-image information can improve risk prediction by capturing clinically relevant imaging patterns beyond a single measure (e.g., cervical funneling and shape variations). Our model implicitly learns cervix characteristics through segmentation, resulting in a more comprehensive and consistent PTB risk assessment. Segmentation motivation unclear R1,R2: The segmentation branch is not a primary contribution; rather, it serves as an auxiliary task that encourages the model to learn anatomically meaningful features. This acts as an inductive bias, guiding the feature representation used for pooling and classification. The reported Dice scores are included for completeness, but the paper’s main focus is the pooling mechanism. We will revise the manuscript to better clarify the role of the segmentation branch as an auxiliary inductive bias. Dataset & evaluation R2,R4: The dataset size is limited, but the fetal application is unique. To ensure a rigorous evaluation, we perform 5-fold cross-validation with patient-level splits to prevent data leakage, repeated across 3 random seeds (15 runs), and report aggregated results. Given the cohort size, an independent held-out test set is not available, and cross-validation is standard practice in this setting. Regarding cohort composition, all patients were prospectively identified as high risk, reflecting the real-world clinical use of transvaginal ultrasound, where screening is not performed in the general population. This also makes the prediction task more challenging due to reduced class separability. We agree that evaluating on a broader population, including low-risk cases, would be valuable, and consider this an important direction for future work. We will clarify patient-level splitting in the revision. How GMM captures feature dependencies R2: GMM pooling represents a bag as a mixture of Gaussian components, where each component models feature relationships through its covariance structure. This allows dependencies between features to be captured. In contrast, density-based pooling treats each feature dimension independently, ignoring such relationships. Ablation analysis R2: While additional ablations could add detail, we emphasize that the paper already includes the most critical analyses (e.g., temperature and probes), which capture the core behavior of the method. Further ablations will be included in future work. Typo R1: Fig3: “K” → “P” thanks. We will revise the manuscript to clarify these points and incorporate the suggested improvements.
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 reviews, including two weak accepts and one weak reject. The authors are invited to submit a rebuttal. The paper proposes a GMM-based pooling strategy for MIL, which reviewers considered an interesting attempt to address a limitation of conventional pooling methods: collapsing instance embeddings into a single point estimate. By modeling the distribution of instance embeddings, the method has potential value for capturing richer within-bag information. However, the reviewers raised several important concerns, including the lack of clear improvement over simpler baselines, insufficient clinical motivation for the preterm birth application, small dataset size, missing ablation studies, unclear role of the segmentation branch, and limited evidence of statistical significance. Overall, the method is potentially useful, but its practical advantage and empirical support require further clarification; therefore, the authors are invited to submit a rebuttal before a final decision is made.
- 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 rebuttal, this paper receives three accept recommendations and one rejection. The AC acknowledges that the proposed GMM pooling strategy addresses an important limitation of conventional MIL pooling by modeling within-bag instance distributions rather than collapsing them into a single point estimate. Although the empirical gains on the preterm birth task are modest and the segmentation branch remains somewhat tangential, the rebuttal reasonably clarifies the methodological motivation, the role of segmentation as an auxiliary inductive bias, the patient-level cross-validation protocol, and how GMM pooling captures feature dependencies through mixture covariance. The majority of reviewers view the approach as technically interesting, generally well motivated, and potentially applicable beyond the presented tasks. Therefore, despite remaining concerns about limited clinical comparison, small dataset size, and marginal performance improvement over simpler baselines, the AC believes that the paper makes a useful methodological contribution to MIL and recommends acceptance.
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.
This paper received two accept and one reject after rebuttal. I recommend rejection of this paper. The main reason is that this paper formulate a MIL problem, without comparing to sota instance aggretators in the MIL field. The MIL techniques, especially the instance feature aggregators, have developed very fast in recent years. These techniques are mostly used for pathology image analysis, which is also a task used for evaluation in this paper. There are too many aggregators that are more powerful than the simple baselines used in this paper.
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.
GMM Pooling sounds technically interesting regardless of its limitations.