List of Papers Browse by Subject Areas Author List
Abstract
Dynamic contrast enhanced breast MRI (DCE-MRI) is a powerful clinical tool for breast cancer detection, providing high resolution anatomical detail together with rich temporal contrast information. However, high dimensional 4D inputs, small lesions, and heterogeneous acquisition protocols across clinical sites hinder robust automated classification of healthy, benign, and malignant cases. To address these challenges, we propose a framework that dynamically analyzes latent representations to adapt to protocol-specific characteristics. Spatial variability is mitigated by reducing confounding background uptake and compensating for misalignment caused by deformable soft tissue. Additionally, relationships in the latent space across phases are leveraged to select the most informative temporal features, improving robustness to protocol-specific temporal variability. Finally, task specific discriminative features are promoted through large scale supervised lesion segmentation pretraining, which substantially enhances downstream finetuning. Evaluated under leave-one-center-out validation on the ODELIA dataset and the held-out AMBL cohort, the proposed framework substantially outperforms finetuned radiology foundation models and prior methods, improving mean AUROC by nearly 8 points and balanced accuracy by 4 points over the strongest baseline. Additionally, our method achieved first place in the MICCAI ODELIA Breast MRI Challenge 2025, further demonstrating its effectiveness for robust breast cancer classification. We publicly release our codebase under https://github.com/MIC-DKFZ/CURIAtor.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3872_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/MIC-DKFZ/CURIAtor
https://github.com/MIC-DKFZ/MeisenMeister
Link to the Dataset(s)
ODELIA: https://arxiv.org/abs/2506.00474
AMBL: https://www.cancerimagingarchive.net/collection/advanced-mri-breast-lesions
MAMA-MIA: https://github.com/LidiaGarrucho/MAMA-MIA
Duke: https://www.cancerimagingarchive.net/collection/duke-breast-cancer-mri
ISPY1: https://www.cancerimagingarchive.net/collection/ispy1
ISPY2: https://www.cancerimagingarchive.net/collection/ispy2
EA1141: https://www.cancerimagingarchive.net/collection/ea1141
Divide-and-Conquer: https://arxiv.org/abs/2507.13830
BibTex
@InProceedings{HamBen_Feature_MICCAI2026,
author = { Hamm, Benjamin AND Kirchhoff, Yannick AND Rokuss, Maximilian AND Langenberg, Moritz AND Ulrich, Constantin AND Wald, Tassilo AND Traub, Jeremias AND Gotkowski, Karol AND Maier-Hein, Klaus},
title = { { Feature Space Guidance for Breast Cancer Classification in DCE-MRI } },
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 tackles the challenge of protocol variability and high-dimensionality in dynamic contrast-enhanced MRI (DCE-MRI) for breast cancer classification. The authors propose a three stage hierarchical framework consisting of a 3D segmentation model for spatial bounding, an adaptive temporal phase selector (CURIAtor) that uses foundation model embeddings to identify the most informative contrast phases, and an anatomically pretrained model fine-tuned for classification. Evaluated using rigorous leave-one-center-out cross-validation on the ODELIA dataset and a held-out AMBL cohort, the method significantly outperforms fixed phase selection, static maximum intensity projection (MIP) approaches, and direct finetuning of standard foundation 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 CURIAtor module introduces a solution to temporal sampling variability by dynamically selecting the most informative post-contrast phase based on feature-space dissimilarity, it bypasses the limitations and biases of static phase selection rules.
The validation pipeline is strong through utilizing leave-one-center-out cross-validation across five distinct clinical centers, alongside an entirely unseen multi-site dataset (AMBL), and provides a highly realistic assessment of out-of-distribution generalizability.
By specifically addressing the heterogeneity of acquisition protocols (varying temporal resolution, number of phases, and contrast timing) across different clinical sites, the method targets a major barrier to the deployment of automated DCE-MRI screening tools.
- 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.
While the reported performance gains are substantial (e. g. , an 8 AUROC point improvement over MIP baselines), the paper currently lacks statistical significance testing for the results presented in Table 1.Could the authors incorporate more statistical tests to validate the robustness of these multi-center improvements?
The ablation study in Table 2 effectively demonstrates the value of the adaptive CURIAtor module compared to fixed temporal combinations (e. g. , Pre + Post1 + Post2). However, how does this 3-phase selection approach compare against standard sequence modeling baselines (such as recurrent networks, 3D+t convolutions, or temporal transformers) designed to natively process the entirety of the 4D sequence?
While the paper excellently evaluates center-level protocol shifts, there is no breakdown of model performance across different tumor characteristics. Can the authors discuss whether the proposed adaptive temporal sampling maintains consistent sensitivity across varying lesion sizes, morphological types (e. g. , mass vs. non-mass enhancement), or molecular subtypes?
- 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 presents an effective, clinically motivated, and well-engineered solution to handling 4D protocol variability in DCE-MRI. Its primary strength is the novel adaptive phase selection mechanism (CURIAtor), which, when combined with rigorous multi-center evaluation, clearly establishes state-of-the-art performance. Although the inclusion of statistical significance tests and a broader comparison against native 4D sequence models would strengthen the technical validation, the current results already demonstrate a clear and substantial improvement over standard clinical AI approaches, making this a strong contribution to the field.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper proposes a DCE-MRI based framework for breast cancer classification, aiming to distinguish healthy, benign, and malignant cases. The method consists of three major stages. First, a breast segmentation model is used to localize the breast region, which reduces computational burden and helps suppress irrelevant background noise outside the anatomically relevant area. Second, the method uses latent features extracted from a radiology foundation model to adaptively select the most informative post-contrast phase based on feature dissimilarity. Third, the model is pretrained on a tumor segmentation task so that the encoder can learn lesion-relevant features before being finetuned for the downstream classification task. Overall, the proposed multi-staged design is reasonable and well-motivated, and the reported results suggest that it outperforms other methods in cross center evaluation.
- 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.One major strength of this work is its clinical relevance. Robust classification on multi-center breast DCE MRI is an important and challenging task, and this study directly addresses highly practical issues, especially protocol heterogeneity across centers and the sparsity of small lesions relative to the full image volume.
2.The paper is generally clearly written and well organized. The proposed framework is conceptually coherent, with a clear internal structure: it first reduces irrelevant background through anatomically informed ROI extraction, then addresses temporal heterogeneity through adaptive phase selection, and finally improves downstream classification through large-scale lesion-aware pretraining. This multi-stage design is well motivated and easy to follow.
3.The reported experimental results are promising and support the effectiveness of the proposed approach. The model shows performance advantages over several compared baselines. In addition, the paper states that the method achieved first place in the corresponding MICCAI challenge, which further suggests the competitiveness and potential practical value of the approach.
- 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 dataset description is insufficient. The manuscript does not clearly describe the class distribution of the three categories, and in particular, it lacks detailed statistics on scanner vendors, acquisition protocols, and center-specific differences. For a paper whose main objective is cross-center generalization, these details are essential. Requiring readers to search for this information elsewhere is not helpful and weakens the clarity and completeness of the study.
2.The core rationale behind the adaptive temporal selection module has not been fully validated. The method selects the post-contrast phase that is most dissimilar to the first post-contrast phase in latent feature space. Still, the paper does not sufficiently demonstrate that this phase is truly the most clinically informative. Although Figure 2 provides three examples, it would be much more intuitive if these examples were directly matched with the actual MR images from the corresponding time points. I assume a large distance in feature space may reflect meaningful delayed washout information, but it could also arise from noise, motion artifacts, registration error, or scanner induced domain shift. In particular, Figure 2 raises a question: for the three illustrated cases, why is the final time point not the most dissimilar one? In DCE MRI, contrast enhancement typically evolves over time, so one might expect the difference to increase gradually. Additional explanation and more intuitive visualization with real DCE MRI images would strengthen the argument.
3.The implementation details of the foundation model baselines remain unclear. The paper compares against Merlin, VISTA3D, CT FM, and Curia as finetuned FM baselines, but it is not clear how these models were finetuned, whether any layers were frozen, and whether they used the same input setting as the proposed method. In addition, several of these models were originally developed for CT rather than breast MRI. Furthermore, Table 2 appears questionable. It reports mean values across datasets, but the reported mean AUROC of 74.89 is identical to the AMBL AUROC in Table 1, rather than the overall mean, maybe an error/typo? Please check this carefully. It would also be more informative if Table 2, like Table 1, reported center-specific results rather than only aggregated values.
- 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?
Overall, this study addresses a clinically meaningful problem, and the reported results suggest that the proposed framework is effective for cross center breast cancer classification on DCE MRI. However, the dataset description is missing and lacks important information needed to assess the validity of the cross center evaluation. Moreover, the core principle of the adaptive temporal selection module would be clearer with stronger justification and a more intuitive explanation.
- 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.
After reading the rebuttal, I think the authors have sufficiently addressed my main concerns. In particular, the explanation regarding DCE-MRI kinetics is convincing, as enhancement patterns may show wash-in, plateau, or washout, and the final phase is not necessarily the most informative or most dissimilar phase. The authors also provided additional details regarding the foundation-model baselines and acknowledged the issue in Table 2.Moreover, they indicated that they will revise the manuscript to improve the dataset description and extend the figure with corresponding MRI examples. Overall, I consider the rebuttal satisfactory and will maintain my original recommendation of Accept.
Review #3
- Please describe the contribution of the paper
The authors propose a focus on breast MRI classification in the scenario of heterogeneous protocols. That is, the authors introduce a method selecting the appropriate sequences maximizing the performances.
The method is evaluated on the recent multi-centric ODELIA dataset.
Some ablation is performed to straighten the advantages of the proposed method.
- 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 choice of the adequate MRI sequences is not obvious, especially when the protocols vary. An automated selection of the sequences may be relevant not only in breast, but in other modalities as well.
The paper is generally clear and the ablations illustrate the advantages of the proposed method.
- 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 paper seems to lack the context of the state of the art. Not much is said about the Odelia Challenge and the winners (https://odelia2025.grand-challenge.org/odelia-breast-mri-challenge-2025-winners/). This makes it difficult to position the method. With that, it seems to have limited contribution with regard to the performances increase. For instance, other works (https://arxiv.org/pdf/2510.27326) allow to achieve higher scores.
- 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
State of the art
I wonder whether the authors could discuss the ODELIA challenge and the most performant methods? And how the proposed situates in this context.
Results It appears that the most of the performances is gained by
- not using the whole volume
- using post_1 sequence Could the authors put some light on it?
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While being quite well written, and clearly presented, the manuscript lacks state of the art. Relying on the ODELIA challenge involves the comparison to the main challengers so the proposed method is easier to situate.
- 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 answers provided by the authors are structured and comprehensive. That is, the authors provide some reasonable arguments with regard to some questions raised in the review, specifically about comparison to the ODELIA challenge competitors. The adaptive selection of temporal points appears clearer now. I would expect these responses to appear in the final manuscript. Yet, I change my opinion to “accept”.
Author Feedback
We thank the reviewers for their constructive feedback. We are encouraged that they recognized the clinical relevance (R2, R3), the novelty of the adaptive CURIAtor module (R1, R3), the rigor of our multi-center evaluation (R1, R2), and the manuscript’s clarity (R1, R2, R3). The points raised are actionable and primarily concern contextualization, statistical reporting, and explanatory detail rather than the validity of the framework or the multi-center results:
@R3 Positioning We agree the ODELIA challenge context should be clearer. The cited arXiv work is a preliminary unpublished challenge write-up covering the pretraining and ROI-separation components of the present framework under static phase configuration; that configuration is precisely reflected in the Pre+P1+P2 ablation row in our Table 2, which our full framework (with adaptive CURIAtor phase selection) outperforms by 4.86 AUROC and 1.66 BA points. In that prior work, AMBL was part of the training set, and the resulting numbers are therefore not fully comparable to our six-cohort mean evaluation, in which AMBL is fully held out. Scores higher than our mean evaluation are single-fold ablation values solely on CAM, in line with our CAM fold results.
@R1 Statistical testing We performed bootstrap 95% confidence intervals for all methods; summarized here ours against the strongest baseline (MIP): AUROC: 75.36 [72.25, 78.29] vs. 67.38 [64.63, 69.95]; BA: 65.78 [63.16, 66.68] vs. 60.42 [58.46, 62.42]. Both improvements are statistically significant. Fold-level CIs for all methods will be added to Table 1. @R1 Subgroup analysis Unfortunately the cohorts lack harmonized subgroup metadata: ODELIA is labeled at the breast/exam level without lesion sizes, morphology, or molecular subtype annotations, and AMBL provides richer pathology data only for a subset. We will note this limitation.
@R2 Dataset details We will include a concise summary of dataset characteristics and class distributions in the manuscript.
@R2 CURIAtor rationale and validation DCE-MRI kinetics may show wash-in, plateau, or washout, so the final phase is not necessarily the most dissimilar from P1, accounting for the cases in Fig. 2.CURIAtor uses P1 as a clinically motivated anchor and selects the phase most complementary in latent space, not merely the most temporally distant. Table 2’s static ablations show the adaptive choice outperforms fixed last- and mid-phase variants, indicating the selected phase is genuinely more informative. To reduce the risk that dissimilarity reflects noise, motion, registration, or scanner shifts, distances are computed after breast cropping and weighted by pooled ROI-token overlap. We will extend Fig. 2 with the corresponding MRI slices.
@R2 FM baselines & Table 2 All non-CURIA baselines were fine-tuned end-to-end with input resolution and CURIAtor-selected inputs matched to our method; We used AdamW with a cosine schedule (100 epochs, batch 32, 50 steps/epoch). For CURIA: Frozen 2D embeddings, slice-wise mean aggregation and linear probe. Results indicate that CT-only FMs perform as well as CT+MRI FM. R2 is correct that the Table 2 value duplicates the AMBL column from Table 1; we will correct it and add further implementation details.
@R1 Native 4D models Native 4D models constitute a research direction of their own, beyond the rebuttal window and out of scope here, where we target improvements within the prevailing fixed-input paradigm. Post-contrast phases each relate to pre-contrast and have bidirectional dependencies with one another; variable phase counts and timings require interpolation, padding, or protocol-specific resampling, each a temporal prior that confounds model vs. harmonization gains; challenges compounded by small cohorts and high VRAM costs. The field reflects this: the ODELIA challenge had no temporal entries, and in MAMA-MIA the only temporal architecture used a fixed 3-phase input. We will more explicitly position 4D modeling as future 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 addresses an important clinical problem in breast DCE-MRI classification under heterogeneous acquisition protocols. Reviewers generally found the proposed framework clear and clinically motivated, especially the adaptive phase selection strategy and the strong multi-center validation design.
The main concerns are related to positioning and evidence rather than fundamental flaws. The paper should better situate itself against the ODELIA challenge winners and other recent state-of-the-art methods. The rationale behind the adaptive temporal selection module also needs clearer validation, including whether feature-space dissimilarity truly identifies clinically informative phases rather than noise or protocol effects. Reviewers also requested more dataset details, clearer baseline implementation details, and statistical testing or subgroup analysis.
- 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 an important and clinically relevant problem, and reviewers generally agreed that the framework is well designed and supported by strong multi-center evaluation. The main concerns were related to positioning, statistical reporting, and clarification of the adaptive phase selection strategy rather than fundamental methodological issues. The rebuttal addressed these points clearly and satisfactorily, and the reviewers who initially raised concerns updated or maintained their final recommendation as accept
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.
Congratulations!
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 authors delivered structured and comprehensive responses addressing reviewer concerns, particularly regarding comparisons with ODELIA challenge methods and adaptive temporal point selection.