List of Papers Browse by Subject Areas Author List
Abstract
Multimodal ultrasound, combining B-mode and color Doppler, is vital for the early screening of skin lesions. However, its automated clinical application is severely hindered by modality quality disparities, incidence-driven data imbalance, and model over-confidence. To this end, we propose ProSyn-Net, a unified multimodal diagnostic framework comprising three synergistic modules. First, a dynamic cross-attention fusion module adaptively aligns morphological and hemodynamic features while suppressing modality noise. The fused representations are then processed by a non-linear prototypical mapping module, which effectively disentangles complex, non-convex minority samples from majority clusters in a high-dimensional space. Finally, an uncertainty-aware manifold ensemble leverages the geometric diversity of these prototypes to yield reliable predictions and extract predictive uncertainty. Extensive experiments demonstrate that ProSyn-Net achieves state-of-the-art performance on the primary in-house imbalanced 6-class dataset (78.60% Macro F1 and 96.66% AUC), while showing promising transferability on a complementary independent public binary dataset (87.80% accuracy). We further analyze uncertainty-guided selective deferral for ambiguous cases. Code is available at https://github.com/EdualcLin/ProSyn-Net.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4942_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/EdualcLin/ProSyn-Net
Link to the Dataset(s)
Public dermatologic ultrasound dataset: https://www.kaggle.com/datasets/alfageme/dermatologic-ultrasound-images
BibTex
@InProceedings{LinJic_ProSynNet_MICCAI2026,
author = { Lin, Jicheng AND Dai, Xiangning AND Wang, Beidi AND Feng, Juncai AND Liu, Haotian AND Yu, Chenke AND Liu, Ruimeng AND Qin, Ziwei AND Zhao, Yujing AND Luo, Ye},
title = { { ProSyn-Net: A Reliable Prototypical Synergy Network for Imbalanced Dermatologic Multimodal Ultrasound Diagnosis } },
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
The paper proposes ProSyn-Net, a multimodal dermatologic ultrasound framework that combines dynamic cross-attention fusion for B-mode and Doppler images, a non-linear prototypical mapping module for long-tailed classification, and an uncertainty-aware manifold ensemble with selective deferral. The method is evaluated on a 917-pair in-house 6-class dataset with an independent hold-out test set and on an independent public binary 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.
This is a well-focused submission that addresses three real deployment problems at once: multimodal quality disparity, incidence-driven class imbalance, and over-confident predictions in a clinical setting. The method is technically coherent. Dynamic cross-attention is used to adaptively suppress low-quality modality evidence, the non-linear prototypical mapping is a meaningful response to minority-class manifold distortion, and the uncertainty-aware ensemble gives the reliability story a clear operational interpretation rather than treating uncertainty as an afterthought. The experimental section is stronger than average: the internal study uses an explicit hold-out test set after an 85/15 split, the baselines are described with attention to fairness, the paper reports minority-class gains (for example on SK), and the submission includes ablations, a public external evaluation, runtime trade-offs, and a clinically relevant deferral analysis.
- 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 main limitation is that the external validation does not match the primary task. The internal benchmark is a 6-class problem, whereas the external public cohort is binary, so the paper should be more careful when framing this as broad cross-dataset generalization. A second limitation is deployment cost: the strongest setting uses a 5-fold ensemble with 50 MC-dropout passes, and the paper reports about 1.25 s per sample for the full ensemble. That is acceptable for research, but less lightweight than the narrative sometimes suggests. Third, the deferral story is promising but should be contextualized more carefully: the autonomous accuracy exceeds 91% at a 50% deferral rate, which is potentially useful but only if such a referral burden is acceptable in the intended workflow. Finally, the fairness claim is mostly strong but not perfectly uniform, because some baselines are reconfigured with the authors’ preferred ConvNeXt-Tiny plus focal-loss recipe while others retain their original designs.
- 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
For rebuttal, I would focus more on calibrating claims than on adding many new experiments. Please state explicitly that the public external validation is binary rather than 6-class, explain how that should and should not be interpreted, and contextualize the clinical meaning of the 50% deferral result. It would also strengthen the paper to summarize the performance-latency trade-off between single-pass inference and the full ensemble, and to briefly justify why the chosen baseline adaptations are fair.
- 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?
I am positive on this submission because it offers a coherent methodological contribution tied to real clinical failure modes, and the empirical evidence is broadly convincing. The paper does not only optimize a closed-set score; it also shows minority-class improvements, examines uncertainty estimation, and discusses selective deferral in a way that is relevant for deployment. My score is not higher because the external validation is not on the same 6-class task and the strongest inference setting is somewhat heavy. Even with those caveats, I view this as clearly above threshold.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The main contributions of the paper present the ProSyn-Net framework: It constructs a unified multimodal diagnostic architecture, using a dynamic cross-attention fusion mechanism to adaptively align morphological and hemodynamic features, effectively suppressing the noise interference from low-quality modalities. A nonlinear prototype mapping module is designed: It proposes to transform the Euclidean distance distribution into a discriminative subspace through nonlinear mapping, successfully separating the non-convex and irregular feature manifolds of minority classes (such as SK) from the majority classes, directly addressing the extreme imbalance problem of data driven by incidence rates. A manifold integration and safety rejection mechanism for perceiving uncertainty is introduced: By leveraging the geometric prototype topological diversity brought about by different data partitions in cross-validation, prediction uncertainty indicators are extracted. By handing over high-uncertainty samples to human experts for processing, the autonomous diagnostic accuracy of the system is increased to over 91%.
- 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.Strong clinical motivation: The problem definition is accurate and precisely targets the actual pain points in real clinical scenarios such as the absence of Doppler blood flow signals, the scarcity of rare lesion samples, and the “overconfidence” of AI models. 2.Solid experimental data: Rigorous comparative tests were conducted on the real clinical 6-class dataset (917 pairs, with a maximum imbalance ratio of 4.7:1) and an independent public binary dataset, proving its excellent generalization ability. 3.Significant performance improvement: In extremely rare classes (such as Seborrheic Keratosis, SK), the Macro F1 reached 80.79%, which was a significant leap compared to the baseline model CMX (62.14%), effectively demonstrating the effectiveness of the nonlinear mapping module. 4.Detailed ablation experiments: Detailed module ablation analysis and t-SNE visualization were provided, clearly showing the intuitive differences in feature decoupling between the linear Voronoi boundary and the nonlinear prototype manifold.
- 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.Limitations of the dataset: Although an open dataset was included for generalization testing, the core 6-class imbalanced dataset originated from a single institution (in-house, specialized hospital). The lack of multi-center (Multi-center) external independent validation has weakened the robustness persuasiveness of the model under different ultrasound devices and acquisition protocols to a certain extent. 2.The generation logic of fusion weights is relatively weak: The dynamic weight generator merely uses feature concatenation followed by MLP and Softmax to calculate the weights. Compared to the highly complex prototype mapping in the subsequent stage, this design of the fusion depth is slightly simple and may not be able to fully capture the fine-grained spatial dependencies across modalities. 3.Computational Overhead: The model inference stage combines 5-fold ensemble and 50 times of Monte Carlo random forward propagation (M=50 MC Dropout). This means that a single sample requires 250 forward propagations, and the total time consumption is approximately 1.25 seconds. For edge ultrasound devices, this delay is too high, limiting their possibility of real-time deployment. 4.Lack of visual interpretability: For dermatologists, in addition to the “uncertainty score”, if the model can provide visual attributions for the lesion area through methods such as CAM (Class Activation Map), the clinical acceptance will be higher.
- 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 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.
(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?
1.The use of cross-attn for fusion in B-ultrasound + Doppler color ultrasound is already a highly saturated routine operation. The dynamic weight generation merely outputs two global scalar weights through a simple MLP, which cannot achieve pixel-level or channel-level fine noise suppression. The claim of “suppressing low-quality modalities” lacks persuasiveness. 2.The greatest advantage of the classic prototype network lies in its interpretability - it directly obtains classification probabilities through strict Euclidean distance. In the paper, it is sent to a two-layer MLP with batchnorm and dropout for nonlinear mapping. Essentially, it is no different from the MLP’s decision-making for classification. There is a suspicion of packaging concepts. 3.Model integration based on 5-fold cross-validation (deep ensembles) and Monte Carlo random sampling (mc dropout) are very classic and ready-to-use techniques in the field of Bayesian deep learning. However, sacrificing inference time to obtain an uncertainty indicator for devices with limited computing resources on the clinical edge ultrasound equipment is not of practical value for such integration. 4.Although the authors did not provide open-source code, the description of the experiments is still relatively detailed. The comparison experiments and ablation experiments are also quite complete. The t-sne method is used to visually compare the non-linear mapping and pure distance in the paper to explain the reason for the improvement of the F1 index from 72.19% to 78.6%. 5.The framework diagram drawn by the authors is very professional. It decomposes the complex network structure into three subsystems using dotted lines and stage1-3 labels, and draws a three-dimensional scatter diagram of “discriminative manifold”.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The paper introduces ProSyn-Net, a framework designed for reliable skin lesion detection using multimodal ultrasound data (morphological and hemodynamic features). The work addresses three critical barriers to clinical deployment: uncertainty estimation, high class imbalance, and multimodal data integration. The authors propose a dynamic cross-modal attention fusion to suppress noise from low-quality modalities, a non-linear prototypical mapping for imbalance handling, and an uncertainty-aware manifold ensemble to estimate predictive uncertainty.
- 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.
Major Strengths:
- The approach is highly innovative and the “ProSyn” concept is both well-explained and easy to understand. That can be easily translate to other task and further enhance the handling of imbalanced dataset issues inherent to medical application.
- The manuscript includes a clear ablation study to demonstrate the impact of each component and provides statistical analysis to verify the significance of the performance improvements.
- The experiments are well-detailed, featuring comparisons against numerous baselines and utilizing rejection plots to validate the effectiveness of the chosen uncertainty measure U(x)
- 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.
Major Weaknesses: The study mentions a second dataset with similar characteristics but does not utilize it for external validation (i.e., training on the larger in-house dataset and directly validating on the publicly available one). Given the importance of generalizability in clinical contexts, this is a missed opportunity to further demonstrate the robustness and clinical readiness of this very well-designed model.
- 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 the rejection plot is useful, the analysis would be significantly strengthened by including metrics such as Area Under the Risk Curve (AURC) or Excess AURC. These metrics would provide a more direct and standardized measure of the quality of the uncertainty-based ranking.
- 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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This is an exceptional and timely paper that addresses core challenges in medical AI, including uncertainty estimation and class imbalance. The proposed ProSyn-Net is innovative. The framework is supported by comprehensive experiments, a clear ablation study, and rigorous statistical analysis. The technical quality and potential for clinical impact make this a standout contribution.
- 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
Author Feedback
We sincerely thank the reviewers and the meta-reviewer for their careful evaluation and constructive comments. We appreciate the recognition of the clinical relevance and overall methodological design of our work.
External validation scope. We will further clarify that the public dataset used in this work is an independent binary dermatologic ultrasound cohort, while our primary in-house task is 6-class imbalanced classification. Accordingly, we view this experiment as complementary evidence of transferability on an independent public cohort, rather than as a strictly matched external 6-class validation. Given the current scarcity of publicly available multimodal dermatologic ultrasound datasets closely aligned with our task setting, we used this cohort as complementary external evidence and will make this scope and limitation more explicit in the camera-ready version.
Deferral interpretation. We will also clarify that the uncertainty-guided deferral analysis is intended to illustrate a tunable, risk-aware human-AI collaboration setting, not to advocate a fixed clinical referral rate.
Prototype rationale. In addition, the prototype-based component is intended primarily to provide a structured class-distance representation for imbalanced, limited-sample learning, especially for irregular minority-class manifolds, rather than to make a standalone interpretability claim. Empirically, its value is reflected in the improvement over the pure-distance variant and in the stronger cross-fold ensemble gain observed in our ablation study.
Baseline fairness and efficiency. For fairness, we will clarify that the adaptations applied to architecturally flexible baselines were intended to place them in a more consistent and task-appropriate setting under a unified backbone and training recipe, while highly specialized methods retained their original designs. We also acknowledge that the full 5-fold ensemble with MC Dropout is a higher-performance but computationally heavier inference configuration, and we will make the corresponding runtime-performance trade-off more explicit in the final version.
We greatly appreciate these suggestions and will reflect them in the camera-ready revision to further improve clarity and calibrated presentation.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
After reviewing this manuscript alongside the expert evaluations, I note that the reviews are mixed but overall lean positive, with strong support highlighting the paper’s coherent methodological design, clinical relevance, and solid experimental validation, including ablations and uncertainty analysis. The main concerns—such as the mismatch between internal and external validation tasks, computational overhead, and the need for clearer justification of certain components—are valid but appear addressable through clarification and minor revision. Given the strength of the contribution and its relevance to reliable medical AI, I recommend a Provisional Accept, contingent on the authors adequately addressing these points in the rebuttal and final revision.