List of Papers Browse by Subject Areas Author List
Abstract
There is an increasing demand for medical image classifiers that deliver both high diagnostic accuracy and clinically actionable interpretability. Existing post-hoc explanation methods, such as saliency maps and Grad-CAM, can highlight discriminative regions but often provide limited insight into the underlying decision logic, which can reduce clinical confidence. Prototype learning offers a promising case-based alternative; however, conventional prototype-based models are often limited by single-scale feature representations and sensitivity to class imbalance, which can degrade accuracy and yield less stable explanations.
To address these limitations, we propose MDPNet, a Multi-Scale Deformable Prototype Network that improves classification performance while providing prototype-grounded interpretability. MDPNet integrates a multi-scale feature extraction module to capture both fine-grained morphological patterns and global contextual semantics, and introduces a deformable prototype learning mechanism that adaptively aligns prototypes with spatial variability in clinical images. In addition, we adopt a logit-adjusted loss to mitigate the impact of class imbalance during training. Extensive experiments show that MDPNet achieves 93.28±1.54% accuracy on a private premature ventricular contraction (PVC) dataset, 97.90±0.37% accuracy on the public OCT dataset, and 87.65±2.54% accuracy on the public CT dataset, consistently outperforming strong CNN/Transformer baselines and recent prototype-learning approaches.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4123_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YanQin_MDPNet_MICCAI2026,
author = { Yan, Qinglan AND Gu, Fei AND Li, Wenlin AND Li, Chaozhuo AND Sun, Shikun AND Wu, Jing AND Ma, Changsheng},
title = { { MDPNet: A Multi-Scale Deformable Prototype Network for Accurate and Interpretable Medical Image Classification } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16882},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors introduce an interpretable framework for medical image and time-series classification that consists of multi-scale feature extraction and a deformable prototype learning strategy.
- 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 method is evaluated on three different datasets representing diverse data modalities (ECG, OCT, and CT) in terms of classification performance. -An ablation study is provided (again focused on classification performance). -Some qualitative exploration of learned prototypes is included for the ECG case (Figure 3)
- 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.
My main concern is that the authors claim interpretability as their primary contribution; however, this aspect is only indirectly evaluated through classification performance. The study would be significantly stronger if it included a quantitative evaluation of the robustness and quality of the explanations. Even more compelling would be an assessment involving clinicians, who are the intended end users of the model.
Minor: -There is repetition between the last paragraph of the introduction and the listed key contributions.
- The manuscript claims the framework is designed for image classification, but Figure 1 illustrates a time-series signal. It may be more accurate to state that the framework is generalizable to data with spatial or temporal relationships.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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?
I find the idea of multiscale deformable prototypes very intresting I am just worried that the hypothesis around the improved interpretibility os not really directly eveluated.
- 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.
Authors promised to add quantitative evaluation for model interpretibility and anlysis of failure cases.
Review #2
- Please describe the contribution of the paper
The paper proposes MDPNet, a prototype-based framework for medical image classification that combines multi-scale feature extraction with a deformable prototype learning mechanism. The authors argue that this framework is higher performance compared to the baselines while being more interpretable.
- 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 addresses an important problem of combining performance and interpretability in medical imaging and presents a conceptually coherent method that integrates multi-scale features with deformable prototype learning.
- 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 has several notable weaknesses. First, the overall novelty is limited, as the method primarily combines existing components such as multi-scale feature extraction, deformable prototypes, and logit adjustment without introducing a fundamentally new idea. Second, the interpretability claims are not convincingly validated; the evaluation relies mainly on qualitative visualizations without quantitative metrics or user studies to demonstrate clinical usefulness. Third, the empirical improvements over strong baselines are relatively modest, especially on public datasets, which weakens the impact of the contribution. In addition, the choice of benchmarks is not particularly strong, with one dataset being saturated (OCT) and another being small and potentially unreliable (COVID-CT), raising concerns about generalizability. Finally, the paper lacks deeper analysis, including limited discussion of failure cases, when the method is most effective, and the computational overhead introduced by the proposed components.
- 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 does not provide sufficient information for reproducibility.
- 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?
I think the paper tackles too many concepts at once, making it not suitable for MICCAI format in terms of required space to cover concepts and results. As a consequence of this, all major contributions of the paper appear lacking a detailedd justification as well as discussion.
- 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.
The authors carefully and correctly addressed my comments.
Review #3
- Please describe the contribution of the paper
This paper addresses the need for both accuracy and interpretability in medical image classification by proposing MDPNet, an interpretable classification framework that integrates a Multi-Scale Feature Extraction Module (MSFEM), a Deformable Prototype Learning Module (DPLM), and a logit-adjusted loss. Specifically, MSFEM incorporates a multi-scale spatial pyramid attention mechanism into the DenseNet bottleneck layers, while DPLM adopts the deformable prototype-matching strategy of Deformable ProtoPNet. The method is evaluated using five-fold cross-validation on a private premature ventricular contraction (PVC) dataset, as well as the NEH OCT dataset and COVID-CT-Dataset.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Clear motivation. The paper identifies two key limitations of existing prototype-learning methods in medical image classification, namely spatial rigidity and single-scale representation. 2.Relatively comprehensive experimental design. The method is evaluated on three datasets from different imaging modalities, and the ablation studies provide reasonable support for the effectiveness of the individual components. 3.The discussion of failure cases is informative. Section 3.4 provides an honest analysis of two representative failure modes: prototype–feature mismatch under ambiguous morphology, and confusion caused by subtle inter-class differences. This discussion is valuable for future research.
- 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 novelty is incremental and primarily lies in combining existing methods for a medical imaging task. MSFEM is directly derived from the multi-scale spatial pyramid attention mechanism in [22] (Yu et al., 2024), with the main adaptation being its insertion into DenseNet bottleneck layers rather than the introduction of a new multi-scale mechanism. DPLM directly adopts the deformable prototype strategy of Deformable ProtoPNet [7] (Donnelly et al., CVPR 2022), and the logit adjustment is taken directly from Menon et al. [15]. The authors should further clarify why this combination is non-trivial and what benefits it provides beyond a straightforward aggregation of existing components. 2.The evaluation of interpretability is too limited. I encourage the authors to include prototype visualizations on the OCT and CT datasets. 3.The writing and figure presentation could be further improved. Figure 3 is difficult to follow, and I suggest adding more detailed annotations, legends, or textual explanations.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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.The authors should clarify what the fundamental innovation is beyond assembling existing techniques. 2.The authors should explain why DenseNet was chosen as the backbone for MSFEM, instead of alternatives such as ResNet or ViT. 3.It is also worth discussing whether using 100 prototypes per class for PVC, or 50 per class for OCT/CT, is excessive. Have the authors analyzed the impact of the number of prototypes on both classification performance and interpretability?
- 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 addresses a clinically meaningful problem, and the proposed method consistently outperforms the baselines across three datasets. The ablation studies are reasonably designed, and the discussion of failure cases is candid and useful. However, the central weakness of the paper lies in its limited technical novelty: all three major components are directly adopted from, or only slightly adapted from, existing methods, and their combination does not yet demonstrate sufficiently deep technical insight.
- 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.
N/A
Author Feedback
We sincerely thank the reviewers for their constructive comments. We have revised the manuscript to better clarify the interpretability, novelty, dataset-specific results, efficiency, backbone choice, and prototype settings. Common Q1: Interpretability validation. We agree that interpretability should be supported by both visual and quantitative evidence. In the revision, we added a prototype-removal analysis, where top-ranked matched prototypes are removed or masked during inference to evaluate their contribution to prediction. The performance drop shows that the prototypes are not only visual explanations but also decision-relevant evidence. Moreover, we expanded the clinical discussion. For PVC classification, subtype-specific cues may appear across different ECG leads and local waveform segments. Our prototype matching helps trace predictions back to representative diagnostic patterns, improving case-level interpretability over black-box classifiers. Common Q2: Incremental novelty. Our contribution is not a simple combination of existing modules, but a clinically motivated evidence-matching framework. Medical diagnostic cues are often subtle, distributed, and non-contiguous. In PVC classification, discriminative evidence may appear across multiple ECG leads and waveform segments. Conventional prototype methods usually rely on fixed or local matching, which may be insufficient for such evidence. We therefore integrate multi-scale representation with deformable prototype matching to capture flexible evidence correspondence while preserving prototype-level interpretability. We revised the Introduction and Method sections to clarify this motivation. R1 Q1: Repetition and inconsistent task description. We have checked the manuscript and revised repeated or inconsistent descriptions in the Abstract, Introduction, and Method sections. R3 Q1: Limited baseline gains and dataset representativeness. We agree that gains on OCT and CT are smaller than on PVC due to differences in task difficulty and diagnostic patterns. PVC is a five-class fine-grained task with subtle inter-class differences and distributed cues across ECG leads, where deformable prototype matching is particularly useful. OCT and CT contain more visually salient or localized patterns, leaving less room for improvement over strong baselines. Still, our method achieves competitive results on these datasets while providing additional prototype-level interpretability. We clarified this in the revision. R3 Q2: Failure cases and efficiency. We strengthened the failure-case and efficiency analysis. Two failure types were discussed: weak/incomplete discriminative cues and highly similar classes such as LVOT vs. RVOT. These cases show the remaining difficulty in fine-grained inter-class discrimination. For efficiency, we added parameters, FLOPs, and inference time. The extra cost mainly comes from multi-scale feature extraction and deformable prototype matching. We clarified that our method aims to balance performance, interpretability, and computational cost. R3 Q3: Too many concepts. We reorganized the method by clarifying the roles of multi-scale representation, deformable prototype matching, prototype-based interpretability, and logit adjustment, while removing redundant terminology. R4 Q1: Writing and Fig. 3.We revised the writing and improved Fig. 3 by enlarging key regions, adding clearer labels, and rewriting the caption to better explain the visualization pipeline. R4 Q2: DenseNet choice. DenseNet was selected because it provided better feature reuse and stability than ResNet, and was more suitable than ViT for limited medical data. R4 Q3: Prototype number. We added prototype number sensitivity analysis. Too few prototypes may miss diagnostic variations, while too many introduce redundancy. The final setting balances performance, interpretability, and efficiency. More prototypes were used for PVC due to its more challenging five-class classification task.
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 a clinically meaningful problem and presents a conceptually coherent framework. However, the limited novelty and insufficient interpretability validation prevent a stronger recommendation. I recommend acceptance with the expectation that the authors will substantially improve the interpretability evaluation (e.g., quantitative faithfulness metrics, expert assessment) and clarify the novelty contribution in the final version. If the authors cannot address these concerns in rebuttal, the paper may need to be rejected.
- 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 successfully address the concerns raised by reviewers.
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 a clinically meaningful problem and presents a coherent prototype-based framework for medical image classification. The reviewers raised valid concerns regarding incremental novelty, limited interpretability validation, modest gains on some public datasets, and insufficient analysis of failure cases and efficiency. The rebuttal responds to these concerns in a generally reasonable manner by clarifying the motivation, promising additional interpretability analysis, failure-case discussion, efficiency reporting, and prototype/backbone analysis. While some concerns, especially regarding novelty and clinical interpretability, are not completely eliminated, the reviewer consensus after rebuttal is positive. I therefore recommend acceptance. The authors should clearly incorporate the relevant clarifications, interpretability discussion, failure-case analysis, efficiency details, and prototype/backbone analysis in the final version.
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 rebuttal has adequately addressed the main concerns from all reviewers and is sufficient to support acceptance. Therefore, I recommend Accept.
