Abstract

Modern foundation models provide highly expressive visual representations, yet adapting them to high-resolution medical imaging remains challenging due to limited annotations and weak supervision. Mammography, in particular, is characterized by large images, variable multi-view studies, and predominantly breast-level labels, making end-to-end fine-tuning computationally expensive and often impractical. We propose Multiple Instance Learning on Precomputed Features (MIL-PF), a scalable framework that combines frozen foundation encoders with a lightweight MIL head for mammography classification. By precomputing the semantic representations and training only a small task-specific aggregation module (40k parameters), the method enables efficient experimentation and adaptation without retraining large backbones. The architecture explicitly models the global tissue context and the sparse local lesion signals through attention-based aggregation. MIL-PF achieves state-of-the-art classification performance at a clinical scale while substantially reducing training complexity. We release the code for full reproducibility on https://nikolajovisic.github.io/mil-pf/.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0943_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/nikolaJovisic/mil-pf-research

Link to the Dataset(s)

EMBED dataset: https://registry.opendata.aws/emory-breast-imaging-dataset-embed/ VinDr-Mammo dataset: https://physionet.org/content/vindr-mammo/1.0.0/ RSNA dataset: https://www.kaggle.com/competitions/rsna-breast-cancer-detection

BibTex

@InProceedings{JovNik_MILPF_MICCAI2026,
        author = { Jovišić, Nikola AND Škipina, Milica AND Dall’Asen, Nicola AND Ćulibrk, Dubravko},
        title = { { MIL-PF: Multiple Instance Learning on Precomputed Features for Mammography Classification } },
        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

    This paper proposes taking advantage of modern foundation models to overcome the challenges related to high resolution images and limited annotations in medical imaging. The authors claim that frozen pretrained encoders provide useful and discriminative features that generalize well for mammography, even if originally trained in other domains (eg, DINOv2). As such, these pre-computed features can be plugged into a smaller, computationally less demanding model to perform the final classification task. Their proposed framework is called Multiple Instance Learning on Precomputed Features (MIL-PF) and is designed to be lightweight and easy to adopt in clinical scenarios, since it involves training very few parameters and requires only image level labels. Evaluation of this framework is done on mammography breast-level classification with complementary lesion detection (trained in weakly supervised). Experimental validation is extensive, using three popular benchmarks in the field - EMBED, VinDr, and RSNA, and the proposed approach is compared against few but meaningful state of the art alternatives. Results suggest that the proposed framework achieves comparable or better classification results than other approaches specific of mammography image analysis.

  • 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 provides a simple yet effective approach to leverage pretrained foundation models for the downstream task of classifying multiview mammography images. This approach requires little training and only needs image level annotations. The paper provides extensive and solid evidence that the proposed MIL-PF approach often outperforms alternative state of the art approaches. The authors perform several runs for each experiment, provide sufficient ablation studies and validation to give support to their claims. Overall, it offers a simple yet effective approach to train models on downstream tasks while relying on pre-computed features obtained from frozen foundation models. The authors also show that multiple foundation models can be plugged in with similar performances, regardless of being previously trained on this type of data, thereby opening new possibilities based on generic foundation models, ie, models not necessarily trained on mammographies or even medical images).

  • 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 major weakness of the paper is its novelty. On one hand, using frozen foundation models for downstream tasks has been extensively explored in the last few years. The paper’s reference [8] is one such example, which already uses features extracted with a frozen MammoCLIP to perform the classification of mammography images. Another work that compares multiple foundation models for this specific task, including a non-medical DINOv2, is [A], also published in MICCAI 2025.Additionally, there are plenty of other works based on this type of approach for other applications. The proposed MIL framework is also not novel, as the described Perceiver-style cross attention is identical to recent approaches based on prompt learning or context optimization (see, eg [B]). While this paper provides additional evidence that this line of work is relevant and effective, it is not proposing a novel approach. Nonetheless, I conceive that it may be the first to combine the using pre-computed features from foundation models with a cross attention mechanism to aggregate patch level information for a MIL framework. The results are also not always convincing, in the sense that some of the alternative approaches achieve equally good or better results in some of the tests. This is particularly evident in the classification results for VinDr and also for the (weakly supervised) detection metrics, considering the results obtained with [8]. The latter could be indicative that modality-specific encoders can be better at capturing medically relevant features than generic ones.

    [A] Germani, E. , Selin-Türk, I. , Zeineddine, F. , Mourad, C. , & Albarqouni, S. (2025, September). Bias and generalizability of foundation models across datasets in breast mammography. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 24-34). Cham: Springer Nature Switzerland. [B] Chen, H. , & Martel, A. L. (2025, September). Breast Cancer Detection from Multi-view Screening Mammograms with Visual Prompt Tuning. In Deep Breast Workshop on AI and Imaging for Diagnostic and Treatment Challenges in Breast Care (pp. 91-100). Cham: Springer Nature Switzerland.

  • 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.

    (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 MIL-PF approach proposed in this paper provides a simple and effective way of exploiting existing foundation models for the classification of breast cancer. Despite its merit, the core contributions of the paper are in line with other previous works, limiting its novelty, as explained in the major weaknesses part of the review. Below is a list of additional comments, questions, and suggestions.

    • In the Introduction, the authors claim that training of CLIP-like models is challenging due to the lack of enough textual data. However, there are multiple approach that counter this argument, including eg MammoCLIP, MedCLIP, MedSigLIP.
    • In the Experimental Setup, the authors state that “Lesions, defined as tissue changes that could be related to breast cancer, represent our R_i”. What do they mean? R_i is mentioned a couple of pages before referring to the sparse regions of interest in an image. It is not clear, in the context of the paragraph, what is implied in this sentence.
    • The authors address a binary problem where BI-RADS 1 is the negative class, and BI-RADS 4, 5 and 6 are all associated with the positive class. What do they do for image with BI-RADS 2 and 3? Are they discarded?
    • And what about the state of the art approaches to which they compare against? Are they all trained on same labels? For instance, [8] originally tackles the binary problem of classifying the presence of a lesion: BI-RADS 1 means no lesion, 2-5 means there are suspicious lesions.
    • What is the original embedding size (possibly different for each encoder)? Why are they reduced to 16 then 8 before the fusion? Did the authors try other sizes?
    • “MIL-PF being the first in the literature to discover generalization properties of the general encoders to the mammography domain” – The work in [A] seems to suggest similar properties, although DINOv2 is not the best performing encoder.
    • It is also arguable that MedSigLIP may be considered a generic foundation model, especially because its training data included some types of x-ray images and CT’s, which are similar to mammograms.
    • The results do not show the standard deviation or the p-values assessing the statistical significance of the improvements compared to the alternative approaches.
    • It does not mention on which dataset the Explainability Analysis (detection capability analysis) was performed. Can you clarify?
    • The ablation study comparing the performance with different combinations of attention paths (global or local attn only VS both), in Table 4 (right), shows very similar performances between the single attention and both. It is not clear if there is any benefit to the dual stream – p-value could help.

    Minor comments

    • “Mutliple runs” – typo
    • Table 4 is the first table to be referenced in the text.
  • 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.

    Overall, this paper presents a valuable contribution to the specific application of breast cancer and mammography image analysis. They show that a simpler and more accessible framework achieves state of the art results, requiring training less parameters than the alternatives. Despite its limit methodological novelty, application-wise it is a good contribution.



Review #2

  • Please describe the contribution of the paper

    This paper demonstrates that general-purpose vision foundation models, such as DINOv2 and MedSigLIP, exhibit outstanding generalization performance in the mammography domain even without task-specific fine-tuning. To address the input resolution constraints of these base models, we propose the Multiple Instance Learning on Precomputed Features (MIL-PF) framework, which extracts features through a dual-stream approach: a global stream for general tissue context and a local tile stream for capturing fine-grained lesion signals. Also, the paper incorporated a Perceiver-style cross-attention mechanism as the aggregation module to effectively focus on sparse, task-relevant local signals among a vast number of unannotated tiles. The proposed architecture achieves state-of-the-art (SOTA) performance on large-scale clinical benchmarks, such as the EMBED dataset, while requiring only a lightweight task-specific module.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The paper demonstrates that general-purpose vision foundation models (e.g., DINOv2, MedSigLIP) can achieve state-of-the-art (SOTA) performance in mammography classification without the need for task-specific fine-tuning of the backbones. This is achieved through global-local dual-stream architecture that effectively balances overall tissue context with fine-grained lesion signals
    • While maintaining a highly efficient and lightweight structure (~40k parameters and 2M FLOPS per forward pass), the framework provides interpretable attention maps. This visual feedback demonstrates strong potential for xAI.
    • Overall, the manuscript is well-written and follows a clear, logical structure that effectively communicates the technical framework and clinical motivations. The method section is theoretically sound, and the experimental setup is robust, characterized by large-scale evaluations and a rigorous validation process.
  • 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 empirical findings regarding the generalization of foundation models to mammography are valuable, the proposed architecture lacks significant technical novelty. The framework primarily relies on a combination of existing components—specifically, a frozen backbone with a standard Perceiver-style cross-attention mechanism and a late-fusion strategy
    • Despite the competitive classification performance, the model exhibits notably low detection accuracy (mAP) for small lesions, though the authors correctly identify this as a limitation tied to the large input tile sizes and mark it for future work
    • Given the 11% variance in Spec@Sens=0.9, the authors should provide p-values and confidence intervals to establish statistical significance over baselines. Reporting the mean and standard deviation across all 36 runs, rather than only the best-performing validation model, is essential to confirm the framework’s true clinical stability and sensitivity.
  • 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

    While maintaining frozen encoders is more efficient , the paper would be more compelling if it explored the trade-off between performance gains and efficiency when applying Parameter-Efficient Fine-Tuning such as LoRA. Furthermore, it would be beneficial to include experiments with additional vision encoders known for strong generalization, such as I-JEPA.

  • 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?

    Although the work does not propose a fundamentally new architectural paradigm, it successfully demonstrates that state-of-the-art performance can be achieved by leveraging pretrained vision foundation models coupled with a Perceiver-based aggregation head. This study is a solid contribution that shows a possibility of using frozen encoders for high-resolution medical imaging tasks.

  • 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.

    My main concern in the initial review was the lack of architectural novelty, as the proposed models are built upon existing architectures. After reading the authors’ rebuttal and the other reviews, I believe the empirical finding — that strong performance can be achieved with pretrained backbones under a parameter-efficient setting — is itself a meaningful contribution relevant to the MICCAI community. Given the practical value of this finding, I update my recommendation to accept.



Review #3

  • Please describe the contribution of the paper

    The paper proposed an MIL pipeline for mammography classification. The pipeline extract features from foundation model and train a lightweight MIL head on top of the extracted features.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper tackles relevant and important clinical problems. The writing is clear and easy to follow. The pipeline is computationally efficient.

  • 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 methodological novelty is very limited, as the approach mainly combines existing components, including frozen foundation encoders, standard MIL formulations. The claimed contribution “We formalize a class of mammography-motivated MIL problems and propose an architecture designed to address them.” is also very weak. Similar problem formulations have already been explored in prior work across various medical imaging domains, such as cardiology and retinopathy screening, and these prior literature are not explicitly discussed in the current manuscript.

  • 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 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.

    (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?

    While the paper presents a practical and efficient framework for mammogram classification, the overall contribution is primarily an integration of existing techniques with limited methodological novelty.

  • 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.

    I have read the authors rebuttal as well as reviewers from other reviewers. Overall i am ok with the paper being accepted, but i think the novelty should not be over-claimed and related works should be more carefully discussed.



Author Feedback

We thank the reviewers for the valuable feedback. They note that MIL-PF is simple yet effective [R1], tackles relevant clinical tasks [R3] and demonstrates valuable and extensively evaluated performance [R1, R2]. All found the method highly efficient and clearly presented. [R1, R2, R3] Novelty: We agree that MIL-PF relies on existing parts and, in the paper, we never claim any of the fundamentals (frozen encoders, Perceiver, dual-stream, etc.) as novel. Rather, we demonstrate the mammography benefits of our particular component arrangement (end of Sec. 3.3). Known parts and ~30x less trainable parameters (single encoder forward pass in training) are preferred when constrained clinically or in hardware, and, since we achieve state-of-the-art results this way, we find MIL-PF valuable for MICCAI; especially since no significantly similar mammography setting is present in the literature. Finally, we will adapt the final wording to emphasize these specifics in the light of reviewers’ comments. [R3] Novelty: R3 raises novelty versus cardiac/retinal, but doesn’t point to specific work. Closest analogs we found are Shad et. al (Nature 2026, arXiv:2312.00357) for frozen foundation models, Huang et al. (ISBI 2025, arXiv:2403.06024) for MIL, and MVCINN (Luo et. al., AAAI 2023) for multi-view/attention. Crucially, none combine all, and none target MIL-PF’s regime: ~5k×5k images, variable number of views, sparse ROIs, bag-labels without text reports. [R1] Novelty: We consider [A] (Germani et al.), [B] (Chen & Martel) and [8] valuable, but none comprehensively covers MIL-PF’s goal. [A] focuses on the foundation model biases, not classification, and not important domain architectural specifics like MIL-PF, and, even in this context, the DINOv2’s value is not recognized (as R1 notices), which is an important MIL-PF’s insight, thanks to the large clinical benchmark (EMBED). We respectfully disagree that [B] is identical to Perceiver-style attention used in our paper. In [B], the prompts are injected in the backbone’s internals; MIL-PF’s latent sits entirely outside the encoder, querying over precomputed embeddings - precisely enabling efficiency. We acknowledge similarities with [8], but demonstrate superior classification performance, precomputing importance, simpler architecture and general encoders’ potential over MammoCLIP. [R1] Results not convincing: On the largest (most diverse) dataset, MIL-PF improves AUC by 4% and Spec. by 20% over the strongest baseline, remaining best or almost the best on the other classification tasks, while massively decreasing trainable parameters over the baselines. [R1, R2] p-values: would strengthen the claims. However, MIL-PF’s EMBED gains are large and reproduce well across 2 encoders, 3 datasets, and 3 metrics, while the 36 runs’ val/test near-perfect alignment indicates stability. [R1] Dual-stream ablation: While p-values would be useful, we consider the 36 runs’ increase of >1% in AUC and especially increase of up to 10% in Spec. significant enough for our dual-stream claims. [R1] CLIP: No scaled text-image mammography corpus exists; MammoCLIP relies on templated/synthetic captions as a workaround; requires sacrifice of arbitrary input resolution and native multi-view support, important specifics MIL-PF handles by design. MedCLIP and MedSigLip are not trained on mammography. [R2] LoRA/JEPA: We consider these very promising, but the limitation of MIL is the encoder backprop, as the signal is sparse. They would, moreover, reintroduce expensive encoder forward passes per epoch; precomputing avoids this. [R1] Embedding size: is encoder-specific, as R1 expects. 16/8 ablation is missing due to space limit, but we see no gains increasing them; decreasing degrades the results. [R1] Will clarify: lesion regions are sparse and of clinical interest (R_i phrasing); baselines are trained on the exact splits as MIL-PF (Sec 3.3); XAI is shown on VinDr; BI-RADS 2,3 cases are discarded; formatting issues will be fixed.




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 reject. The work addresses a practically relevant problem in mammography classification and presents a simple, efficient, and clearly organized framework based on frozen foundation-model features and lightweight MIL aggregation; the positive reviews highlight its empirical strength, including evaluation on multiple mammography benchmarks and evidence that precomputed features from general or medical foundation models can achieve competitive, and in some cases state-of-the-art, performance. However, the main concern is limited methodological novelty, as the framework largely builds on established ideas such as frozen feature extraction, MIL aggregation, and Perceiver-style cross-attention, making its distinction from prior work insufficiently clear. The paper would benefit from stronger positioning against related studies, including MammoCLIP-based feature extraction, foundation-model comparisons in mammography, and visual-prompt or Perceiver-style aggregation methods. Several experimental and methodological details also require clarification, including label definitions, dataset handling, embedding-size choices, detection-analysis settings, and the statistical significance of the reported improvements. Overall, the paper is practically meaningful and empirically promising, but its contribution relative to existing work remains somewhat unclear; 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, including two reviewers who maintained weak acceptance and one reviewer who changed from rejection to acceptance. The AC acknowledges that the paper addresses a clinically important problem in mammography classification and proposes a simple, efficient, and practically attractive MIL framework built on precomputed frozen foundation-model features. The reviewers consistently recognize the paper’s strengths in clear presentation, computational efficiency, use of only image-level labels, and extensive evaluation across multiple mammography benchmarks.



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 presents MIL-PF, a simple and efficient framework using frozen foundation-model features with a lightweight MIL head for mammography classification. Reviewers agree that the methodological novelty is limited, as the work combines existing components, but the rebuttal clarified the positioning and practical contribution. The empirical results are strong, especially on large-scale mammography benchmarks, and the framework is computationally attractive with very few trainable parameters. Overall, the paper makes a valuable application-oriented contribution and is recommended for acceptance, provided the final version avoids overclaiming novelty and better situates the work relative to prior methods.



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 have addressed the concerns regarding novelty, ablation study, and p-values, and discussed related works. All reviewers are convinced to accept this manuscript after the rebuttal.



back to top