List of Papers Browse by Subject Areas Author List
Abstract
Anomaly detection in medical images is challenging due to limited annotations and the domain gap. Existing reconstruction-based methods often rely on frozen pre-trained encoders, restricting adaptation to domain-specific patterns and degrading localization accuracy. Meanwhile, prototype-based learning offers interpretable representations but commonly suffers from prototype collapse, where a few prototypes dominate training and reduce diversity. To address these issues, we propose DNP-ConFormer, a unified framework that integrates a trainable encoder with prototype-guided reconstruction and a Diversity-Aware Alignment Loss. A momentum encoder enables stable domain-adaptive representation learning, while a lightweight Prototype Extractor discovers informative normal prototypes and injects them into the decoder via attention to guide reconstruction. The proposed alignment objective further encourages balanced feature-to-prototype assignments, effectively mitigating prototype collapse. Extensive experiments on multiple medical imaging benchmarks demonstrate improved representation quality and anomaly localization compared with prior methods. Visualization and prototype assignment analyses further validate the effectiveness and interpretability of our approach. The source code is available at https://github.com/liluhu0/DNP-ConFormer.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0431_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/liluhu0/DNP-ConFormer
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiLuh_Diverse_MICCAI2026,
author = { Li, Luhu AND Liu, Bin AND Lin, Bowen AND Shen, Zihan AND Wang, Chengwei AND Fu, Shujun},
title = { { Diverse Normal Prototypes-Guided Contrastive Reconstruction for Medical Anomaly Detection } },
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 DNP-ConFormer, a framework for medical image anomaly detection that combines a trainable encoder with prototype-guided reconstruction. It introduces a diversity-aware alignment loss to prevent prototype collapse and improve feature representation. Experiments show improved anomaly localization and interpretability over existing methods.
- 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 is generally well-motivated, with clear experimental validation and solid presentation. 2.The paper includes sufficient ablation experiments to evaluate the contribution of each component.
- 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.Using teacher-student feature discrepancy at the patch level to generate anomaly maps is not well justified. The paper lacks explanation on why this strategy reliably reflects anomaly regions. 2.The number of prototypes 𝑀 is a critical hyperparameter, but no analysis is provided. Its impact on performance should be systematically evaluated. 3.It is unclear how the prototypes are updated during training. The authors should clarify whether they are learnable and how they are optimized. 4.It is unclear whether all methods in Tables 1 and 2 use the same backbone architecture, raising concerns about the fairness of comparisons.
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper lacks sufficient methodological clarity, particularly regarding the rationale behind the anomaly scoring strategy and the learning mechanism of prototypes. Key design choices are not well justified, and important factors such as the sensitivity to the number of prototypes are not analyzed. In addition, the experimental evaluation raises concerns about fairness and completeness, making it difficult to fully validate the claimed improvements. Overall, these issues weaken the technical soundness and impact of the work.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The manuscript proposes an improved method for knowledge distillation-based anomaly detection. The student model extracts normal prototypes from the encoder features to guide the transformer decoder to reconstruct the features of the teacher model. Both the student and teacher encoder are initialized with DINOv2 pretrained weights and the teacher encoder is adapted to the medical domain using an exponential moving average of the student encoder.
- 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.
- A simple change in the loss formula improves prototype diversity in comparison to INP-Former, which is visually demonstrated in Fig. 1.- The findings regarding the fine-tuning strategy of the foundation model is a valuable insight beyond the anomaly detection community of MICCAI.
- The manuscript contains a thorough ablation of the proposed changes.
- 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.
- Comparison is limited since the state of the art methods only contain knowledge-distillation based methods, but no image-reconstruction based methods such as AnoDDPM [1] or AutoDDPM [2] or embedding based methods such as PatchCore [3] (which although devised for industrial vision applications reach high performances on medical imaging tasks).
- Missing knowledge distillation based baseline Skip-ST [4].
- The manuscript lacks a few details in the methods: it does not tell how the cosine similarities are aggregated over the layers, it does not explain S in Eq. 4., and it does not provide any details on the bottleneck aggregation module.
- Is there a reason for using DINOv2 instead of DINOv3?
[1] Wyatt et al., “Anoddpm: Anomaly detection with denoising diffusion probabilistic models using simplex noise”, CVPR (2022). [2] Bercea et al., “Mask, stitch, and re-sample: Enhancing robustness and generalizability in anomaly detection through automatic diffusion models”, IMLH (2023). [3] Roth et al., “Towards total recall in industrial anomaly detection”, CVPR (2022). [4] Liu et al., “Anomaly detection for medical images using teacher–student model with skip connections and multiscale anomaly consistency”, IEEE IMS (2024).
- 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
- There is no appendix (the authors referred to it at the end of implementation details).
- Is MIR a typo and should be MRI in the “Brain MIR” dataset?
- Fig. 1d): Probably output should be f_D^l instead of f_D^{l-1}
- Text in Fig. 1 and 3a and 3b is too small and only readable after zooming in a lot. The font size in Fig. 2 is also borderline.
- The order of methods in Tab. 1 and Tab. 2 are not the same.
- 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 evaluation is limited only to comparisons with knowledge-distillation based anomaly detection methods, the manuscript has interesting take-homes that are also of interest for people in the community that work with vision foundation models.
- 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 thank the authors for the additional details in the method section. Even though the evaluation is limited to knowledge distillation based methods due to space constraints, the comparison covers the most relevant methods in that line of work and allows to asses the improvements over previous methods.
Review #3
- Please describe the contribution of the paper
This paper addresses the domain gap and prototype collapse issues in unsupervised medical image anomaly detection by proposing a unified framework named DNP-ConFormer. The main contributions are:
1.Introducing a student-teacher reconstruction architecture equipped with a momentum-updated encoder (the M2+ framework) to replace the traditional frozen pre-trained encoder. This design aims to achieve better domain adaptation to medical images while preserving pre-trained knowledge.
2.Proposing a Diversity-Aware Alignment Loss from a long-tail learning perspective. This loss encourages balanced feature assignments to diverse normal prototypes, effectively mitigating the prototype collapse commonly observed in low-contrast medical images.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Clear Motivation: The authors provide a profound analysis of why existing reconstruction methods struggle with medical images. The proposed momentum update and diversity-aware alignment loss are logically coherent solutions to the issues of pre-trained feature mismatch and the loss of representational diversity caused by coherence objectives.
2.Convincing Architectural Evolution: The method derivations and ablation studies (e.g., the progression from M0 to M1, M2, and finally M2+ in Fig. 3) clearly demonstrate why direct fine-tuning is infeasible and justify the necessity of the momentum encoder.
- 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.Insufficient Verification of the Momentum Encoder (M2+) Applicability: Replacing the frozen pre-trained encoder with a momentum-updated encoder is theoretically sound. However, this strategy is currently only integrated and evaluated within the INP-Former architecture. If the authors claim this is an effective solution for medical image domain shifts, it is highly recommended to adapt and validate this momentum strategy on other baseline reconstruction-based methods (e.g., RD4AD, ReContrast, Dinomaly). This would provide stronger evidence of its value and generalizability.
2.Lack of Discussion and Comparison with the Core Baseline (M2): The motivation and network architecture of the proposed M2 framework are similar to the frozen-encoder+trainable-encoder of ReContrast. However, the authors did not to provide any comparison or explanation of this relationship in the main text. Please clarify: Is M2 same as the arch used in ReContrast, or are there specific architectural innovations?
3.Given the rich modalities and high variability of medical images, evaluating on only three public datasets is somewhat limited. It is recommended that the authors perform additional experiments on a more comprehensive medical anomaly detection benchmark, such as the BMAD benchmark (which includes sub-datasets from 5 different modalities), to solidify their empirical conclusions.
4.In Fig. 3(b), which shows the performance trends over training iterations, the authors provide the curves for OCT2017 and ISIC2018 under the M2+ framework, but omit the M2+ results for APTOS2019.Is this because the M2+ framework performs poorly (e.g., fails to converge or drops in performance) on the APTOS dataset?
- 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.
(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 overall quality of the paper is above the general level. But lack enough results on more benchmarks, ablation studies, and justification comparing to core baseline.
- 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 all reviewers (R) and the Meta-Reviewer (MR) for their constructive feedback. As rebuttals are limited to 4,000 characters and text-only, we provide concise clarifications below.
=== Response to R1 and MR ===
1.Rationale of the anomaly map -Principle: Follows prior works [6,8,9,10,14,18], which assume a network trained only on normal data poorly rebuilds anomalous regions, thus giving larger rebuild errors there. We measure this error by cosine distance, then resize and apply Gaussian smoothing to get the final anomaly map. -Empirical: Localization quality is shown by heatmaps (Figs. 4–5) and pixel-level metrics (Table 2). On our in-house dataset, we further compare predicted maps with masks (Fig. 5), where high-response areas overlap well with real lesions.
2-3.Number of prototypes and updating -Following INP-Former [14], we fix prototype number M=6 across all datasets. In practice, smaller M under-represents normal diversity, while larger M mainly increases computation with marginal gains. -As in Sec. 2 & Fig. 2(c), prototypes P = E(F_Q, {t_m}) are dynamically generated via cross-attention with encoder features and optimized end-to-end by L_total (Eq. 6).
4.Backbone fairness We faithfully follow each baseline’s original backbone:
- ViT-Small/14 (35.13M): Ours, INP-Former, Dinomaly
- WideResNet50 (136.28M): RD4AD, ReContrast
- ResNet50: EDC (34.64M), EA2D (49.06M) Our improvements are clear against baselines with the same backbone (INP-Former, Dinomaly) or comparable parameters (EDC, EA2D). Against larger-backbone baselines, we still outperform despite using fewer parameters. We will clarify the backbone of all methods in the final draft.
=== Response to R2 and MR ===
1-2.Additional baselines We compared the most directly related SOTA methods, e.g., EDC [9] (IEEE TMI 2024), EA2D [18] (IEEE TMI 2025) within the page limit. Since rebuttal guidelines preclude new experiments, we will discuss the methods you mentioned in related-work section. As for PatchCore’s performance, one may refer to Recontrast [8] (Table 5), which reports generally moderate results.
3.Methodological details -Multi-layer aggregation: following [10], we adopt group-to-group supervision, summing features from layers 2–5 and 6–9 of ViT-Small into two groups. Cosine similarity is computed between the encoder and decoder feature groups, and then taking the average. -In Eq. 4, S is cosine distance. -The bottleneck module is a two-layer MLP.
4.DINOv2 vs DINOv3 (i) When this work was conducted, DINOv3 and its evaluation ecosystem were not yet mature; (ii) for fair comparison, ViT-based competitors (INP-Former, Dinomaly) all use DINOv2; (iii) the framework is backbone-agnostic and can be directly migrated to DINOv3. The above and other issues (typo ‘MIR’, Fig. 1d output, figure font size, table order, appendix ref) will be corrected in camera-ready.
=== Response to R4 and MR ===
1.Applicability of M2+ We agree with the suggestion to apply this to other baselines. M2+ is method-agnostic by design: it only modifies the teacher-update rule (Eq. 2) and is decoupled from the rest of the framework. Systematic studies on other baselines will be pursued in future work.
2.Relation to ReContrast/M2 M2 shares the asymmetric twin-path philosophy of ReContrast [8], but is presented only as an intermediate ablation in Sec. 2.1/Table 3a, not as our contribution. Our contribution is M2+ and tackling prototype collapse from a long-tail perspective.
3.Datasets Our datasets follow the protocol of EDC [9] and EA2D [18], and our private brain MRI dataset, i.e., BrainMRI. We note that rebuttal guidelines preclude new experiments; extension to the BMAD benchmark is planned as future work.
4.APTOS curve in Fig. 3b Fig. 3b is intended to illustrate the limitation of M2: it cannot keep improving with iterations on some datasets (OCT2017, ISIC2018), which M2+ resolves. On APTOS2019, M2 already improves monotonically, while M2+ further achieves higher AUC.
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 proposes DNP-ConFormer for medical anomaly detection, using prototype-guided reconstruction, a momentum-updated encoder, and a diversity-aware alignment loss to address domain gap and prototype collapse. Reviewers generally recognized the clear motivation, solid presentation, useful ablation studies, and the potentially valuable insight of adapting DINOv2 to medical images with EMA. However, several recurring concerns should be addressed in rebuttal. Reviewers requested clearer justification of the anomaly scoring strategy based on teacher–student patch-level discrepancy, more details on how prototypes are learned or updated, and analysis of the critical prototype number hyperparameter. Multiple reviewers also raised concerns about limited comparisons, including missing reconstruction-, embedding-, and KD-based baselines, and insufficient discussion of the relationship to core baselines such as ReContrast/M2.Further clarification is needed on methodological details and comparison fairness. A focused rebuttal addressing these and other reviewer-raised points would help reassess the contribution.
- 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.
DNP-ConFormer addresses domain gap and prototype collapse in medical anomaly detection via a momentum-updated EMA teacher and a diversity-aware alignment loss. Reviewers found the motivation clear, the presentation solid, the ablations useful, and the DINOv2-EMA adaptation insight valuable beyond anomaly detection; R2 upgraded to Accept post-rebuttal. The rebuttal reasonably clarified the in-scope concerns, including comparison fairness, prototype generation, anomaly-scoring rationale, method details, and the relationship to ReContrast. Remaining limitations, including M2+ being validated only within INP-Former, evaluation on three datasets, and lack of sensitivity analysis for M, are real but non-fatal. The authors are encouraged to address these in the camera-ready where possible, particularly by more carefully scoping claims about M2+ generality and incorporating the rebuttal clarifications. Recommend 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.
Accept
- Please justify your recommendation.
The rebuttal addressed several of the main methodological concerns and improved the clarity of the paper, especially regarding the anomaly scoring rationale, the prototype mechanism, and implementation details. Although some concerns remain about comparison completeness and methodological justification, the overall post-rebuttal assessment is positive. I therefore support acceptance.
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 paper received Weak Reject (no post-rebuttal submitted), Accept, and Weak Accept (no post-rebuttal submitted).
The authors’ rebuttal addressed shared methodological concerns, such as anomaly-map rationale, prototype mechanism, multi-layer details, backbone fairness, and ReContrast positioning.
For the camera-ready, the authors should incorporate the rebuttal clarifications and add a related-work discussion of the missing reconstruction- and embedding-based baselines.