List of Papers Browse by Subject Areas Author List
Abstract
Multimodal imaging is essential for accurate primary bone tumor subtyping, yet its clinical utility is limited by cross-modal diagnostic conflicts and frequent missing modalities. Existing deterministic multimodal fusion methods fail to model predictive uncertainty, often yielding over-confident predictions under incomplete inputs or inter-modality disagreements. In this paper, we propose DPDEP, a novel Dirichlet Process (DP)-driven diagnostic evidence-based prototype learning framework for multimodal bone tumor subtyping. Inspired by clinical differential diagnosis, DPDEP projects multimodal features onto modality-specific prototypes to capture heterogeneous diagnostic signatures. Instead of deterministic fusion, we introduce a DP-based evidential arbitration mechanism that dynamically quantifies prototype reliability and resolves inter-modality conflicts by prioritizing the most trustworthy evidence. To handle missing modalities, we introduce a DP-guided asymmetric evidential distillation strategy that transfers robust joint-modality reasoning from teacher to student by selectively dropping the more reliable modality without hallucinating missing-modal evidence. We collected a well-curated large-scale in-house bone tumor dataset comprising 4,443 clinical cases. Experiments demonstrate that DPDEP achieves state-of-the-art performance on both three-class malignancy grading and nine-class fine-grained histotype subtyping, with strong robustness and reliability in real-world incomplete multimodal settings. Our code is available at: https://github.com/HKU-MedAI/PBT-DPDEP.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2542_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/HKU-MedAI/PBT-DPDEP
Link to the Dataset(s)
N/A
BibTex
@InProceedings{WuFen_Diagnostic_MICCAI2026,
author = { Wu, Feng AND She, Ruiling AND Zhang, Ruiyang AND Peng, Liang AND Ma, Qiang AND Zhao, Yinghua AND Yu, Lequan},
title = { { Diagnostic Evidence-Based Prototype Learning via Dirichlet Process for Multimodal Bone Tumor Subtyping } },
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
This paper proposes a novel Dirichlet Process (DP)-driven Diagnostic Evidence Prototype learning framework named DPDEP for addressing the problems of missing modalities and diagnostic conflict in multimodal primary bone tumor subtype classification. The framework contains three core components: Semantically Aligned Diagnostic Prototype Learning module, which projects multimodal features onto modality-specific prototypes to capture heterogeneous diagnostic features. DP-based Dynamic Evidence Arbitration mechanism, which introduces the DP stick-breaking prior to dynamically quantify prototype reliability and prioritizes the most credible evidence during fusion to resolve cross-modal conflicts. DP-guided Asymmetric Evidence Distillation strategy, which, to handle missing modalities in real-world scenarios, selectively drops the more reliable modality during training to robustly distill joint-modal reasoning capability from the teacher network to the student network, while avoiding the generation of hallucinated evidence. Experiments validate the method on a large-scale internal dataset containing 4,443 clinical cases, achieving state-of-the-art performance on both the malignancy three-class classification and the histological nine-class classification tasks.
- 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 clinical motivation is highly valuable and clearly defined. The paper targets the real clinical pain points in bone tumor imaging diagnosis where MRI is frequently unavailable and inter-modal feature overlap causes conflicts. The problem definition closely aligns with practical medical scenarios. The method demonstrates strong novelty and interpretability. Rather than relying on simple deterministic fusion or feature concatenation, the paper innovatively introduces the Dirichlet Process into multimodal fusion. By dynamically quantifying evidence reliability to arbitrate conflicts, it enhances model robustness while providing interpretability for diagnostic decisions. When handling missing modalities, the DP-guided “asymmetric distillation” is adopted to prevent clinically dangerous phenomena from occurring.
- 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.
Insufficient experimental generalizability and prominent class imbalance issues. All 4,443 test cases come from a single clinical center. Although the total sample size is large, no multi-center external independent cohort validation has been conducted, nor has comparison and validation been performed on public datasets. The model’s generalization ability under different equipment, scanning protocols, and data distribution discrepancies cannot be adequately demonstrated, raising concerns about clinical practicability and transferability. Incomplete ablation study with key validations missing. The paper only conducts a simple analysis on the number of prototypes K, without performing hyperparameter sensitivity validation. This includes the dropout probability, the rationale and stability of the four balancing coefficients λ in the loss function, all of which remain unexplained, directly affecting the reproducibility of the method.
- 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 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.
(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?
The methodological novelty of introducing the DP process into multimodal evidence arbitration. The authors not only identified the clinical deficiency that existing fusion methods tend to produce overconfidence under modal conflict, but also ingeniously introduced the theoretically grounded Dirichlet Process, achieving substantial results in both interpretability and robustness. A highly challenging clinical task supported by a large-scale dataset. Primary bone tumors constitute a complex sub-field with extremely high annotation difficulty. The 4,443 finely annotated cases provided in this paper and the outstanding nine-class classification metrics (Macro-F1 35.41%) demonstrate strong potential for clinical application. Lack of multi-center validation and multi-dataset validation. All 4,443 cases come from the same hospital. Although the total sample size is large, no multi-center external independent cohort validation has been conducted, nor has comparison and validation been performed on public datasets.
- 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 addressed my concerns, I have no other questions.
Review #2
- Please describe the contribution of the paper
This paper proposes DPDEP, a multimodal bone tumor subtyping framework that combines modality-specific diagnostic prototypes, a Dirichlet-Process-based evidential arbitration module, and an asymmetric distillation strategy for missing-modality robustness. The claimed advance is to move beyond deterministic fusion by explicitly modeling prototype reliability, resolving cross-modal conflicts, and preserving joint-modality reasoning when MRI is unavailable.
- 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 targets an important and clinically relevant problem: primary bone tumor grading and histotype subtyping from incomplete CT/MRI data, which is well motivated in the introduction and abstract. 2.The method is reasonably novel in how it combines prototype learning with DP-style evidential arbitration and selective modality dropping, rather than using standard deterministic fusion alone. 3.The comparative evaluation is fairly comprehensive, covering multiple metrics including AUROC, ACC, Macro-F1, and BACC, and the manuscript additionally reports 95% confidence intervals and p-values against the best competing baseline, which strengthens confidence in the reported performance gains. 4.The construction of a relatively large in-house bone tumor dataset is itself a meaningful contribution for this task, and the scale of data collection and curation reflects substantial practical effort that deserves recognition.
- 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.Reproducibility is relatively limited. Although the paper states that code is available, the study relies on a private in-house dataset, and the manuscript does not provide sufficiently detailed information on preprocessing, input construction, and other implementation details needed for faithful reproduction. If the dataset is not publicly released, this becomes a more substantial limitation for the community. 2.The paper strongly emphasizes uncertainty, reliability, and conflict resolution, but the evaluation is almost entirely classification-centric. I did not find dedicated uncertainty/calibration metrics or failure analysis showing that the evidential scores are actually better calibrated in practice. No direct evidence found in the manuscript. 3.The missing-modality story is promising but not fully stress-tested. The dataset contains many CT-only cases and fewer paired CT-MRI cases, while the distillation module is applied only to naturally paired samples; I did not find a detailed stratified analysis by missingness pattern or severity beyond the overall benchmark tables and ablation summary. 4.The paper argues that prior methods rely on deterministic feature aggregation and provide limited interpretable diagnostic evidence, and positions DPDEP as a more evidence-aware and clinically meaningful alternative. However, the manuscript does not provide direct comparative visualization or interpretability analysis against baseline methods to substantiate this claim. While Fig. 3 offers a qualitative view of prototype assignments within the proposed model, it does not demonstrate that DPDEP yields more interpretable diagnostic evidence than prior approaches. This leaves one of the paper’ s central claimed advantages insufficiently validated.
- 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.
(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 is a borderline positive submission. The paper addresses an important multimodal clinical classification problem and proposes a technically interesting framework that combines diagnostic prototype learning, evidential arbitration, and missing-modality distillation in a coherent way. The empirical study is fairly solid, with a relatively large in-house dataset, multiple baselines, several evaluation metrics, confidence intervals, p-values, and supportive ablation results. However, confidence is moderated by limited reproducibility due to the private dataset and insufficient implementation detail, the lack of direct evaluation for calibration/uncertainty claims, and the absence of comparative interpretability analysis to support a key part of the paper’ s motivation and novelty positioning. Overall, I find the work promising and potentially acceptable, but some of its central claims would benefit from stronger empirical support.
- 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 questions have been adequately addressed in the rebuttal. The authors stated that they will release source code and data which addresses reproducibility concerns. I think this is a well evaluated approach which shows good results and would be interesting to the community.
Review #3
- Please describe the contribution of the paper
1 The paper proposes a multimodal framework, DPDEP, for primary bone tumor subtyping under missing-modality settings, covering both malignancy grading and fine-grained histotype classification.
2 Its main technical contribution is a DP-based evidential fusion mechanism that performs prototype-level reliability estimation and cross-modal evidence arbitration, instead of relying on standard deterministic multimodal fusion.
3 The paper also introduces an asymmetric evidential distillation strategy to transfer joint CT–MRI reasoning to incomplete-input settings, with the goal of improving robustness when one modality is missing.
- 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 studies a clinically relevant problem in a realistic setting.
2 The proposed formulation is technically interesting. A key strength is the attempt to combine modality-specific prototype learning with an evidential fusion mechanism, rather than relying on standard deterministic multimodal fusion. In particular, the paper frames multimodal fusion as prototype-level evidence arbitration, which is an interesting perspective and gives the method a clearer conceptual identity than a conventional fusion network.
3 The paper includes a practically motivated design for missing-modality training. The asymmetric evidential distillation component is a reasonable and application-relevant design choice, as it explicitly aims to transfer joint CT–MRI reasoning to incomplete-input settings. This is a meaningful attempt to preserve multimodal benefits when one modality is absent at inference time.
- 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.It seems that the role of the DP is not sufficiently justified. It does not convincingly show why DP is necessary beyond acting as a more complex sparsity-inducing prototype weighting mechanism. In particular, it remains unclear whether the gains come from the DP formulation itself or, from adaptive weighting over prototype evidence, since simpler non-DP alternatives are not compared.
2.The main evaluation setting in Section 3.2 is not clearly specified. Table 2 reports the headline comparison, but it is unclear whether these results are obtained on naturally incomplete cases, synthetically degraded inputs, paired CT-MRI samples, CT-only samples, or some mixture of these settings. Since robustness to missing modalities is a central claim of the paper, this ambiguity makes it difficult to interpret what the reported gains actually reflect.
3.The main uncertainty/reliability claim is not directly evaluated. The paper only reports standard classification metrics such as AUROC, ACC, Macro-F1, and BACC. It appears that no results on reliability are provided. 4 The experimental comparison is not sufficiently up-to-date to support the “SOTA” claim. While the included baselines are relevant, the paper does not compare recent work on incomplete/missing-modality multimodal medical diagnosis. In particular, recent MICCAI 2025 methods such as Robust Incomplete-Modality Alignment for Ophthalmic Disease Grading and Diagnosis via Labeled Optimal Transport and MM-DINOv2 address closely related issues of diagnosis/classification under incomplete or missing medical imaging modalities, yet are not discussed or compared. The paper would benefit from clarifying why such recent methods were excluded.
5.The method appears largely generic rather than specifically designed for primary bone tumors. Although the clinical application is PBT subtyping, the methodological components are mostly generic missing-modality multimodal learning modules. The manuscript does not clearly show what part of the method is truly tailored to bone tumor diagnosis, rather than transferable to a broader multimodal classification setting.
6.Sensitivity analysis on λ_kd, λ_dp-dist, λ_div, and λ_dp is missing. The paper only analyzes the effect of the prototype number K.
- 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.
(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?
Please refer to the major weaknesses.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
I thank the authors for their rebuttal and additional analyses. However, I remain unconvinced that my main concerns have been adequately addressed. The rebuttal largely relies on post-hoc clarifications and newly added analyses, which suggests that key claims in the original submission were not sufficiently supported by the manuscript itself.
My central concern about the necessity of the DP formulation remains. The rebuttal states that DP induces sparse evidence allocation and conflict-aware fusion, but it still does not compare against simpler non-DP sparse or adaptive prototype-weighting alternatives. The DPL/DPEF/DPAED ablation is useful, but it does not isolate whether the gain comes from DP itself or from a more generic prototype weighting / evidential fusion mechanism. Thus, the main methodological claim remains under-justified.
Regarding reliability, the newly added analyses are relevant, but they do not fully resolve my concern. Since reliability-aware arbitration is a central claim, the lack of direct reliability evaluation in the submitted manuscript remains a serious weakness. Comparing DP-derived uncertainty with softmax confidence is also insufficient to establish that the proposed reliability mechanism is intrinsically better.
I find the response on recent baselines not directly addressing my concern. My criticism was not merely that RIMA and MM-DINOv2 must be included as off-the-shelf baselines, but that the manuscript’s SOTA claim requires comparison against, or at least careful discussion of, the most recent 2025 methods for incomplete/missing-modality multimodal medical diagnosis. The rebuttal shifts the issue to domain incompatibility, arguing that these methods were designed for different imaging settings. However, this does not justify the absence of an up-to-date baseline discussion, nor does it establish that the original baseline set is sufficient. Moreover, if the authors adapted these methods using frozen 3D features, the rebuttal should provide enough detail to judge whether those adaptations are faithful and competitive.
Finally, I remain unconvinced that the method is specifically tailored to primary bone tumor diagnosis. The rebuttal mostly restates that CT/MRI evidence can be complementary or conflicting in PBT diagnosis, but this is a general multimodal imaging issue. It does not identify a clearly PBT-specific modeling component or show that the method exploits tumor-specific clinical or imaging knowledge.
Overall, the rebuttal adds clarifications and new analyses, but it does not resolve my core concerns about methodological necessity, reliability validation, baseline completeness, and disease-specific contribution. I therefore have to stick to my previous negative rating.
Author Feedback
We appreciate the favorable comments on the clinical importance of PBT diagnosis(R1,R2,R3), our novel DP-based evidential design(R1,R2,R3), and the comprehensive evaluation(R2). We address the main concerns below with additional analyses and clarifications.
Q1: Reproducibility/Generalizability(R1,R2) For reproducibility, we will release synthetic/demo raw data and runnable code for preprocessing, input construction, feature extraction, mask generation, training, and evaluation, with splits/settings/hyperparameters in configs. Class imbalance is addressed by class-weighted CE and Macro-F1/BACC. Generalizability is supported by the 2010–2025 cohort with multiple scanners/protocols and frozen 3D foundation features. We also add a preliminary 188-case external-center malignancy validation, where DPDEP remains best under cross-center shift, supporting transferability while broader multi-center validation remains future work.
Q2: Hyperparameter sensitivity(R1,R3) We add sensitivity analyses for p_drop, λ_kd, λ_dp-dist, λ_div, and λ_dp. Performance remains stable around the reported settings. Lower p_ drop weakens missing-modality robustness, while overly high p_drop hurts complete-modality reasoning. Too-small loss weights underuse distillation/regularization, whereas too-large weights over-constrain the student/prototype allocation. The values in Sec.3.1 are best or near-best across metrics.
Q3: Interpretability comparison(R2) DPDEP provides built-in evidence decomposition via activated CT/MRI prototypes and modality reliability/conflict before fusion, unlike deterministic fusion baselines. We further add a model-agnostic evidence retrieval analysis: each method retrieves top-k nearest training cases from its final embedding, and we compare macro-averaged same-label C@1/C@5(top-k purity) and mAP. DPDEP scores best, showing more diagnosis-consistent evidence.
Q4: Reliability evaluation(R2,R3) DPDEP estimates reliability from marginal prototype evidence and down-weights lower-reliability evidence under CT/MRI conflict. We added uncertainty-based failure detection and calibration analysis. For DPDEP, uncertainty is derived from DP-based modality reliability/arbitration scores; for baselines, we use softmax confidence. DPDEP achieves higher error-detection AUROC(0.87 vs. 0.83) and lower ECE(0.032 vs. 0.044), showing better failure identification and calibration than deterministic fusion baselines.
Q5: Missing-modality setting(R2,R3) The main evaluation(Tab.2–3) follows real clinical missingness, where 69.6% cases are MRI-missing. To further stress-test robustness, we randomly masked MRI in 10%, 25%, and 40% of naturally paired test cases and computed the relative drop from the paired full-modality setting. DPDEP remains best at all rates and shows the smallest drop, indicating robustness beyond the observed missing pattern.
Q6: DP necessity(R3) DP turns prototype matching into sparse evidence estimation and reliability-aware CT/MRI arbitration. In ablation, DPL already serves as an adaptive weighted prototype baseline; adding DPEF brings clear gains, with further improvement after DPAED. Thus DP contributes beyond generic prototype weighting by inducing sparse evidence allocation and conflict-aware evidential fusion. R1/R2 also recognized this DP-specific design as innovative/reasonably novel.
Q7: Recent baselines(R3) RIMA targets 2D fundus/OCT ophthalmic diagnosis and MM-DINOv2 focuses on foundation-model adaptation with modality masking, so neither is directly comparable to our 3D CT/MRI PBT setting without adaptation. We adapted both using the same frozen 3D features, DPDEP still outperforms them across metrics.
Q8: PBT specificity of DPDEP(R3) DPDEP addresses a PBT-relevant problem(R2) and is well motivated by bone tumor imaging diagnosis(R1). DPEF arbitrates complementary/conflicting CT/MRI evidence used in PBT diagnosis, while DPAED targets real MRI-missing PBT cases without hallucinating missing-modal evidence.
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.
A further rebuttal is required for the reviewers to reach a consensus.
- 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 clinical challenge in multimodal primary bone tumor diagnosis and presents a novel evidence arbitration framework based on Dirichlet Process modeling. The work is supported by a large-scale clinical dataset and comprehensive experiments demonstrating strong empirical performance. While one reviewer remains unconvinced regarding the necessity of the DP formulation and the completeness of baseline comparisons, the majority of reviewers found these concerns sufficiently addressed in the rebuttal and agreed that the method provides a meaningful contribution. Given the strong experimental evidence and clinical relevance, I 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.
Reject
- Please justify your recommendation.
The rebuttal fails to substantiate the necessity of the proposed DP design, provide validation of reliability claims, or ensure sufficiently SOTA baselines, leaving the core contributions insufficiently supported.
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 represents a strong and clinically meaningful contribution with clear potential impact. The work addresses an important real-world problem in multimodal bone tumor diagnosis, particularly under missing-modality and conflicting cross-modal evidence settings, and introduces a technically interesting Dirichlet Process-based evidence arbitration framework that balances predictive performance, interpretability, and clinical safety. The experimental validation is comprehensive, involving a relatively large clinical cohort with strong performance across a challenging nine-class classification task, supported by ablation studies and statistical analysis. While some limitations remain regarding external validation, calibration evaluation, and implementation details, the reviewers agreed these are addressable issues that do not diminish the core novelty and practical significance of the work.