List of Papers Browse by Subject Areas Author List
Abstract
Prenatal screening for congenital heart disease (CHD) is challenged by rare pathology, heterogeneous data, uneven modality contributions, and frequent missing measurements. We propose a unified asymmetric-similarity methodology: a single Tversky prototype-comparison principle applied independently to structured clinical data and to medical imaging, rather than a fused multimodal model. For clinical data, a composite loss combining Tversky-prototype attraction, Tversky-contrastive separation, and entropy-regularised soft assignment learns embeddings capturing graded diagnostic similarity under severe class imbalance. For imaging, similarity fields classify by comparing frozen encoder features to learned prototypes via asymmetric Tversky measures, defining class membership through fiber inclusion in the similarity geometry. On the CARDIUM dataset (7.19 % CHD prevalence), our schema-aware Tversky loss raises F1 from 0.294 to 0.562 with perfect recall using only clinical data, while similarity fields improve F1 by 2-4 x over identical frozen backbones. On thyroid nodule classification (Stanford AIMI TUC), hypergraph Tversky embeddings achieve AUC 0.81 from relational TI-RADS features alone. By replacing symmetric distances with asymmetric prototype comparisons, the framework offers a principled solution for learning under class imbalance, and missing data. Code is available at https://github.com/ividja/tversky-fibers.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2375_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/ividja/tversky-fibers
Link to the Dataset(s)
Stanford AIMI Thyroid Ultrasound Cine-clip dataset: https://aimi.stanford.edu/datasets/thyroid-ultrasound-cine-clip
CARDIUM dataset: https://bcv-uniandes.github.io/CardiumPage/
BibTex
@InProceedings{MülJoh_Fibers_MICCAI2026,
author = { Müller, Johanna P. AND Baugh, Matthew M. G. AND Wright, Robert AND Day, Thomas G. AND Razavi, Reza AND Kainz, Bernhard},
title = { { Fibers of Asymmetric Similarity: A Framework for Clinical and Imaging Data } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes a framework to process structured clinical data and medical images in the field of congenital heart disease (CHD) detection. The task has several challenges, and the authors’ main goal is to especially address one of them, i. e. , its low prevalence. The authors make use of Tversky similarity for both types of data to explicitly mirror the asymmetry in the data (imbalanced positive vs negative samples) in the model design. Interestingly, the framework makes use of prototype learning instead of one-hot label-based classification, so for both tasks the authors learn class prototypes and shallow classifiers only. In the clinical data task, the authors design a comprehensive loss function, summing two Tversky-based terms and cosine similarity to ensure optimal convergence. In the image analysis task, this allows them to achieve great performance gains on pipelines containing large pre-trained backbones without having to fine-tune these models.
- 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 whole idea is interesting and seems to work well. In particular, I find what is shown in Fig. 4 particularly valid: classification performance using feature extracted by pre-trained backbones (e. g. , DINOv2b or BioMedCLIP) improves by up to 4X (F1@0.5), notably, without fine-tuning. The authors achieve this impressive results by optimizing the only learnable entities in the image processing pipeline, i. e. , class prototypes (as per the paper, CxMxD parameters, 5x256 per class) and projection layers (about which no information is given, but presumably very lightweight, too). This achievement is coupled and justified by the choice to use prototype learning instead of traditional one-hot encoded classification.
- 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 paper is not particularly clear. 1a. The first line of the abstract introduces the problems related to prenatal screening for CHD, but the end of the abstract is about thyroid nodule classification. Overall, the paper suggests that the authors dealt with CHD detection, but then thyroid nodules appear out of nowhere. In my humble opinion, a clearer way to present the work would be to introduce the actual clinical tasks and datasets at the very end, and to be more general about asymmetry and rare conditions in the introduction/abstract, but then the authors should address the lack of comparison or context for the second task (see point 1e). Maybe it would be better not to include the thyroid-related task altogether, as its results
1b. In Method, the authors talk about “clinical data” and “features”, so I assume that we are talking about tabular data (also because in Introduction the authors wrote “structured clinical data”), but this is opaque. Then, the reader is given no information about what the authors define “independent clinical parameters (e. g. , blood pressure)” and “relational or interdependent features”. Since the work gives a great importance to this division, I think the reader would benefit from a clear explanation of why this was necessary. Also, as per 2.1, we know that “temporal or repeated measurements (e. g. , gestational age)” are “embedded and pooled as h”, D-dimensional embeddings, but then what about the rest of data (i. e. , non-temporal but independent features and interdependent features)?
1c. The framework is presented as “unified”, but in fact clinical data and images are treated separately by separate models, right? This is not a weakness per se, of course, but reading through the paper I always had the feeling that the two branches would merge sooner or later, which was not the case.
1d. No reference is given for the hypergraph encoder, and specifics about this component are not provided in the Implementation section (unlike everything else, which is explained in detail). This hinders reproducibility and actual comprehension of the paper. In general, I feel the clinical data part is particularly underdescribed: while one can picture what foetal US images look like, it is hard to understand what those “26 maternal clinical variables” might be.
2.Comparison with the state of the art is particularly weak, which is understandable because the CARDIUM dataset was release at the end of 2025.There is little to no context, however, about Vega et al. [22], that appears out of nowhere in Table 1, and no context at all for Yamashita et al. [24]. Also, results from the vision pipeline on the thyroid nodule dataset are not reported at all. I understand there is limited space, but taking everything into account, I again suggest that maybe it would be better to remove this second task from the paper and explain the methods more in detail. Moreover, I am not an expert, but are thyroid nodules that rare? Yamashita et al. report (citing Haugen et al. 2016) that the prevalence is 19%–68% (at least, it was in 2016).
3.While the authors show the loss component ablation study, the actual role of similarity field-based approach is unknown. For example, it would be particularly beneficial if the authors included performance achieved by a traditional pipeline trained via one-hot based classification instead of prototypes, e. g. , (frozen BB + trainable classifier) pipeline with the same number of trainable parameters as the pipelines the authors present.
- 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
1.In Introduction, the authors talk about a “pathology of interest” that is never presented (is it CHD itself? what referred populations are we talking about? any source for this?).
2.I think Fig. 1 would be clearer if “L_contrastive” changed to “L_cosine”.
- 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?
I believe the work is technically sound, but the presentation currently limits its impact. Several unconventional design choices, from Tversky loss to similarity field, appear to contribute to strong performance, yet their individual roles are not clearly explained. However, in my opinion the paper can be accepted only if the authors show the actual contribution of every one of these componets and better explain the methods.
- 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 clarified a few points in the rebuttal, but the manuscript needs to be heavily improved upon acceptance, by including the points addressed by the reviewers and partially addressed by the authors. Above all, clarity in terms of: clinical aim, clinical framework/state of the art, variables used, and comparison.
Review #2
- Please describe the contribution of the paper
The paper proposes a unified framework for handling rare and imbalanced medical classification problems using asymmetric Tversky similarity in latent space. Instead of relying on standard symmetric distances such as cosine similarity or Euclidean distance, the authors learn class prototypes and compare samples to these prototypes using an asymmetric Tversky formulation that can explicitly prioritize recall over precision. This is particularly relevant for screening tasks such as congenital heart disease, where missing a positive case is more costly than producing a false positive.
- 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 paper proposes a genuinely original formulation based on asymmetric Tversky similarity in latent representation space. 2.A major strength is the unified nature of the framework. 3.The proposed “similarity field” formulation for image classification is particularly interesting. Instead of fine-tuning a large pretrained backbone, the method keeps the encoder frozen and learns only a small projection layer and class prototypes. This makes the approach lightweight, computationally efficient, and potentially easier to deploy in settings with limited data. 4.The statistical analysis is another strength.
- 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 paper is quite abstract and mathematically heavy, and at times the presentation obscures what is actually new. 2.Although the clinical-data results are strong in recall, the precision remains quite low.
- 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?
The paper presents a novel and well-motivated methodological contribution, and the gains on the CARDIUM dataset are meaningful, especially for the clinically important minority class. Although the presentation is somewhat abstract and the comparison with other imbalance-aware approaches is incomplete, the asymmetric Tversky formulation appears sufficiently original and effective to place the paper 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.
Accept
- [Post rebuttal] Please justify your final decision from above.
I thank the authors for their rebuttal. My two main concerns were the abstract and mathematically heavy presentation, and the low precision of the clinical results. Both have been addressed satisfactorily. The additional implementation details promised for the camera-ready (clinical Tversky formula, prototype initialization, patient-level split protocol, hyperparameter selection) address the reproducibility gaps.
Review #3
- Please describe the contribution of the paper
This paper proposes a prototype-based framework built around asymmetric Tversky similarity for two settings relevant to medical AI: structured clinical/tabular data and ultrasound image classification. For clinical data, it introduces schema-aware and hypergraph encoders trained with a composite objective that combines Tversky-prototype attraction, Tversky-contrastive separation, and cosine regularization. For images, it keeps pretrained visual backbones frozen and learns a small projection layer plus class prototypes, interpreting prediction as graded membership in class-specific similarity fibers, with a multiple-instance learning extension for patient-level prediction from multiple images. Empirically, the strongest reported clinical result on CARDIUM improves F1 from 0.294 to 0.562 with recall 1.0, the TUC structured-feature model reaches AUC 0.81, and the image-side similarity-field head improves minority-class CHD performance over frozen-backbone classifiers, with the best MIL result at F1 0.519.
- 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 prototype based approach is lightweight and potentially more interpretable than large end to end classifiers.
Parameter efficient design: frozen encoders + learned prototypes is clinically realistic.
Evaluation on two public datasets, includes ablations, compares multiple backbones and adds a statistical ranking analysis. The ablation study is strong and well shown.
F1 improvement from 0.294 to 0.562 on CARDIUM is substantial.
- 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 baselines are narrow with no comparison with focal loss, SMOTE, cost sensitive learning or class balanced sampling.
Hard to attribute gains to Tversky specifically vs any imbalance aware method.
The paper is framed as a unified multimodal framework, but experiments are actually clinical only and image only not a joint fused model.
a) Some reproducibility details are unclear:
b) Exact clinical Tversky formula
c) How the embeddings are constrained
d) How the hyperparameters are chosen
e) Whether all splits are strictly patient level
MIL random sampling of N=4 images per patient introduces stochasticity that isn’t systematically analyzed.
- 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
I would encourage the authors to compare the proposed approach against stronger standard imbalanced-learning baselines, provide a sensitivity analysis for α, β, and η, and explain the very low variance reported for the schema-aware + Tversky-full configuration. It would also be useful to report the computational overhead of the Tversky similarity computation relative to standard cosine or Euclidean prototype comparisons, and to clarify the exact implementation choices required for reproduction: the clinical Tversky formula, embedding constraints/normalization, hyperparameter selection, prototype initialization, and the patient-level split protocol. Finally, the limitations discussion should be strengthened, especially with respect to prospective validation, subgroup fairness, site/scanner shift, calibration, and the referral burden associated with operating at very high recall.
- 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?
I found the paper interesting and potentially impactful because the prototype-based asymmetric-similarity formulation is lightweight, clinically plausible, and produces meaningful gains on some imbalanced tasks. However, I currently view the submission as marginally below the acceptance threshold because the experimental comparisons are not broad enough to demonstrate that Tversky similarity itself is the key factor, the claimed unification across modalities is not supported by an actual fused multimodal experiment, and several implementation details important for reproducibility remain underspecified. Overall, I see promise here, but the current evidence is not yet strong enough for acceptance.
- 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
We thank R1-R3+AC for: “genuinely original” (R2), “lightweight, clinically plausible, potentially highly impactful” (AC, R3), with “substantial” gains and a “strong” ablation (R3). All refs/vals are in the manuscript or released code.
1.Abstract/introduction (AC, R1, R2). We reframe both around asymmetric similarity for rare and imbalanced clinical learning, on two tasks: CHD detection from heterogeneous data (CARDIUM, 7.19% prevalence) and thyroid stratification from relational TI-RADS features (TUC, AUC 0.81 from structured features alone, matching CNN SOTA [24]). “Pathology of interest” denotes CHD in referred populations [4]. TUC shows the hypergraph encoder transfers to relational features, not that nodules are rare. Fig. 1 will be relabelled L_cosine (R1.2). The “unified” framing (R1.1c, R3) is methodological (one asymmetric-similarity prototype principle), not a fused multimodal model; we will rephrase.
2.Tversky loss and clinical data (R1.1b/d, R3.a-d). “Clinical data” are structured tabular features (numeric, ordinal, categorical). The schema-aware encoder handles independent parameters (e.g., maternal age, BP, BMI); valid ranges and orderings carry signal. The hypergraph encoder handles relational features (e.g., TI-RADS composition + echogenicity + margins) via a transformer over typed tokens with a patient CLS token. Tversky (alpha, beta) are fixed; learnable are the projector, prototypes, and (image side) convex weights across similarity fields (Prop. 5 of [11]). The clinical Tversky reuses Eq. 4 with sigmoid-mapped h, p in (0,1)^D and beta > alpha: missing target-class features penalised more than extra embedding features, recall-prioritised for rule-out screening. This will be added to the Implementation paragraph.
3.Role of the similarity field, isolated from Tversky (R1.3, R3). Fig. 4 is exactly this experiment: grey bars are a frozen backbone + linear head at identical learnable-parameter budget, the baseline R1.3 requests. Across five backbones, frozen + linear collapses on CHD with F1 0.106-0.238; replacing only the head with our prototype + Tversky similarity field yields F1 0.347-0.440 (single instance) and 0.519 (MIL). Tab. 2 then isolates asymmetry: symmetric Tversky (eta = 0.5) gives F1 0.422 +/- 0.20; asymmetric (eta = eta*) gives 0.562 +/- 0.008.This 25-fold variance reduction is itself the answer to R3’s “explain the low variance” question: asymmetry stabilises optimisation under severe imbalance.
4.Imbalance baselines (R1.2, R3). Vega [22] is CARDIUM’s baseline; Yamashita [24] is TUC CNN SOTA. On focal/SMOTE/cost-sensitive: focal reweights per-sample BCE; SMOTE oversamples in input space; both are sample- or output-space fixes for symmetric losses. Asymmetric Tversky acts earlier, in embedding-prototype geometry, with alpha != beta reweighting feature mismatches. Tab. 2’s symmetric row is the ablation against which such reweighting competes; it is dominated at p < 0.05 (Fig. 5). Our Contrastive/Hybrid rows further use cosine with class-prevalence weighting, equivalent to cost-sensitive learning.
5.Reproducibility (R3). Splits are strictly patient-level: CARDIUM 3-fold by design (folds by patient ID); TUC 2-fold stratified on 191 annotations. Prototypes are Xavier-initialised; LR warm-up 5 epochs (MIL) or 3.Clinical Tversky alpha = 0.3, beta = 0.7, m = 0.3; image side alpha = 0.2, beta = 0.8 (Sec. 3). Hyperparameters via grid search on the train fold. MIL N = 4 stochasticity is captured in n = 5 seed variance (+/- 0.073, MedViT-HF), smaller than the gap to the fine-tuned baseline (0.519 vs 0.345). Tversky adds one elementwise op over cosine, ~5% wall-clock on A100. 6.Precision and limitations (R2, R3). At 7.19% prevalence with recall 1.0, precision 0.391 yields a referral burden below current obstetric practice (~30% detection [4]); high recall is the screening-appropriate setting. Limitations: prospective validation, scanner shift, calibration, and referral burden at high recall.
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.
All reviewers acknowledge that the article tackles a relevant problem and that the proposed prototype-based asymmetric-similarity formulation is clinically plausible, interesting, and potentially highly impactful. However, all reviewers also raised concerns about several methodological choices that should be addressed by the authors in the rebuttal. Please note that the rebuttal should focus on clarification rather than adding new experiments or results, particularly by providing more details and insights regarding the Tversky loss, the similarity field, reproducibility and experimental details, and how the abstract and introduction could be revised to improve clarity. Currently, all reviewers found them a bit difficult to understand.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
I generally agree with the concerns raised by all reviewers regarding the clarity of the manuscript and the tendency to overstate some of the claims. In particular, the framing of the work as a “unified multimodal” approach appears insufficiently supported by the experimental validation presented in the paper.
Another major concern is the lack of comparison against standard imbalance-aware baselines. Since the proposed contribution is closely related to handling class imbalance, the experimental evaluation should include well-established methods such as focal loss, SMOTE-based approaches, or other commonly used reweighting and resampling strategies. Without these baselines, it is difficult to assess whether the proposed method provides a meaningful improvement beyond existing and widely adopted techniques.
Additionally, the paper lacks a sufficient sensitivity analysis of the proposed hyper-parameters. Such an analysis is important to demonstrate that the proposed loss formulation is genuinely responsible for the reported gains and that the method is robust to parameter choices. At present, it remains unclear whether the improvements are attributable to the core methodological contribution or to favorable hyper-parameter selection.
Overall, while the problem addressed by the paper is relevant, the current version suffers from insufficient clarity, limited experimental validation, missing baseline comparisons, which collectively weaken the strength of the contribution.
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 paper addresses the challenge of identifying rare conditions under severe class imbalances, where missing a positive case carries a high clinical cost. To resolve this, the authors propose a parameter-efficient prototype framework built on asymmetric Tversky similarity fields that optimizes local class representations with frozen foundation models. Its key strengths include a highly parameter-efficient footprint, substantial empirical gains on fetal ultrasound, and an objective mathematical target prioritizing clinical recall. The rebuttal corrected the misleading unified multimodal text narrative, and clarified the optimization mechanics. The technical adjustments addressed most of the reviewer concerns and support recommendation.
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 work proposes a unified framework for handling rare and imbalanced medical classification problems using asymmetric Tversky similarity in latent space. Specifically, the authors interestingly use prototype learning for this rather than relying on standard distance metrics. The authors demonstrate their framework on two public datasets with detailed ablations about their components.
All the reviewers have explicitly agreed that this is a really genuine innovative approach for screening tasks such as screening tasks such as congenital heart disease. They have appreciated the design of the framework, the ablations, and the experimental evaluation. Some concerns regarding the clarity of writing, reducing the heavy mathematical descriptive parts, limited baselines on CARDIUM, and minor issues with figures were raised. The authors have convincingly addressed all the issues, however, the final camera-ready version needs to be revised in terms of writing and presentation. I request the authors to make these changes genuinely in writing and if required by consulting their peers to go through the presentation and writing quality, as this will be an important contribution to the community. Please also consider final suggestions from the other reviewers post rebuttal to incorporate them into the camera-ready version.