List of Papers Browse by Subject Areas Author List
Abstract
Accurate identification the primary tumor of metastatic cervical lymphadenopathy (CLA) is crucial for guiding clinical treatment, yet clinical diagnosis remains challenging due to the complexity of tracing multi-potential origins using ultrasound images and incomplete clinical information. Existing deep learning methods typically utilize the imaging semantic features from B-mode ultrasound (BUS) and color Doppler flow imaging (CDFI), or incorporate basic clinical information, neglecting the importance of patient-specific features such as tumor markers (TMs) in clinical diagnosis. To address these limitations, we propose a new multimodal imaging-features and distribution-based tumor-marker fusion network (MDFN) for five categories of CLA metastatic origins (thyroid, head and neck, respiratory, female reproductive, and digestive). First, a distribution-based TM imputation method is proposed to reconstruct missing TMs, which treats the available clinical information of each patient (such as sex, age, neck region, etc.) as a vector to construct data distributions between TMs and avoid the data bias issues. Building on these personalized TMs, we propose the first population-personalized fusion framework, which integrates semantic features related to lymph node morphology from BUS images, semantic features related to vascular distribution from CDFI images, and TM features consistent with individualized patient data, thereby simulating clinical reasoning patterns. The effectiveness of the proposed MDFN method was evaluated using extensive experimental results from 3,100 multi-origin metastatic CLA cases, achieving an area under the receiver operating characteristic (AUC) of 0.891, with corresponding accuracy, sensitivity, specificity, and F1 of 0.863, 0.604, 0.913, and 0.661, respectively, outperforming other state-of-the-art methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3619_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{LiRui_Multimodal_MICCAI2025,
author = { Li, Rui and Li, Chunyan and Lin, Xi and Xu, Jinfeng and Li, Fang and Lu, Yao},
title = { { Multimodal Fusion Network with Distribution-based Tumor-Marker Imputation for Multi-Origin Metastatic Cervical Lymphadenopathy Classification } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15960},
month = {September},
page = {453 -- 462}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose a multi-modal approach to fuse (i) B-ultrasound (BUS) and (ii) color Doppler ultrasound (CDFI) images of lymph nodes from the neck as well as (iii) tumor markers (TMs) imputed from patient clinical data. For (iii), a cycle GAN is used to predict the expression of 8 TMs from patient and tumor basic data. The cycle Gan was inspired from previous work ([4], breaking anonymity ?). The models aims to predict the type of primary cancer causing metastatic cervical lymphadenopathy (CLA). The approach is evaluated on a large dataset of 3100 lymph nodes and contains several ablation studies to reveal the contribution of each components. A comparison to one other method (MSMFN [19]) is provided.
- 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 multimodal fusion seems to value contributions from each of the 3 data sources as it is demonstrated by the feasibility studies. The evaluation of the method on a large dataset provides very encouraging performance. The use of the cycle GAN is original but several details are not provided in the paper, making it difficult to assess its value and originality when compared to previously published work ([4]).
- 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.
Several details and nomenclature issues are entailing the scientific relevance of the proposed paper.
First, concerning nomenclature and the title, I find the term “population” and “personalized” misleading as they rather refer to different modalities in the paper (image and clinical features). Multimodal fusion would better convey the core ideas of the approach.
Second, while the cycle GAN constitute the main originality of the approach, it was already published [4]. This paper seem to consist of a new application of the latter. However, many details are missing to really demonstrate the relevance of the latter in the context of CLA classification. In particular, the motivation, details and performance of the cycleGAN in imputing (predicting) the 8 TMs from five patient/tumor parameters are not provided. The motivation of predicting TMs from clinical parameters is not provided, and it would be needed to have an experiment where the clinical parameters are directly used for the prediction of CLA primaries. Even more importantly, the performance of the prediction of TMs from clinical parameters is not evaluated and not even discussed, making it very difficult to see the quality of this key component. The training of the cycleGAN is not explained in terms of data splitting. The ratio of missing data (clinical or TM) is not mentioned. The cycleGAN is compared to other approaches in Table 2 only on the final CLA classification performance but not in terms of prediction of TMs.
The justification for a 6:2:2 cross validation (CV) is not provided. Usually CVs contain equally sized folds. It therefore seems to be a single training/validation/test split. Details on how 95% confidence intervals (Table 1) are built are missing.
In the dataset, the observation is a cervical lymph node. Therefore, there is a risk of overfitting if the data is not split by patients in the case of patients having more than one lymph node, which is not mentioned. Having normal lymph nodes in the dataset would also be beneficial to better represent the clinical population and use case.
- 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.
- 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/2025/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 — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Overall, missing information, evaluation and previous publication concerning the main original component, the cycle-GAN, strongly hinders the contribution of the work.
- 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 clarified several important aspects in the rebuttal, and are partially explaining how they will modify the paper to include these clarifications. It is crucial that they include most clarifications in the camera ready version.
Review #2
- Please describe the contribution of the paper
The authors propose a method to predict the primary cancer site in patients with metastatic cervical lymphadenopathy by a multi-modal neural network which encodes and then fuses features from B-mode ultrasound imaging, color doppler flow imaging, and tumor markers.
- 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.
Thorough ablation studies are used to show that all three modes contribute to the performance of the model, as well as other modules such as the channel attention module used to process the doppler input, and the cycleGAN used to impute missing values in the tumor markers input.
The paper is working toward solving a difficult but useful problem in computer-aided diagnosis.
Comparisons are made to the state-of-the-art.
- 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 benefits of the imputation method are less clear than other benefits. Fig 2b shows that the RoC curves are not very different. While the new method does dominate, one wonders if it is a difference of any clinical significance. On the other hand, the average F1 score is indeed substantially better (0.661 vs 0.588), so perhaps a breakdown by class would help clarify this.
Clinical framing could be improved — the authors note that 90% of CLA would fall into the five categories of primary sites they are using, but in practice, what would need to be done to accommodate patients in the other 10%?
- 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.
- 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/2025/en/REVIEWER-GUIDELINES.html
In question 9 I have said that the submission does not provide sufficient information for reproducibility. Ultimately, the best way to ensure this would be to release the code. If code cannot be released, some additional details may be needed as to the precise implementation of the layers from the tensor fusion (which is described in the supplementary) through to the final prediction output.
Similarly, a description of the implementation of the CAM module would be needed.
Additionally, a description of the training in more detail may be required. Was any pretraining of the image encoders used? etc
Finally the paper also needs some editing for consistency of terminology, and correctness. A few examples:
-
“thyroid” is repeated twice in section 3.4.
-
should use consistent terminology when referring to the different classes (in some places respiratory is used, in the introduction “breath” is used, in some places “female reproductive system” is used, but in other places “female tumors”, “gynecological tumors”, etc)
-
- 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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While I do think the paper is overall a good contribution and highly relevant for MICCAI, I think some of the shortcomings in terms of clarity need to be addressed for acceptance. Furthermore, I think the authors should consider releasing the code, or providing more implementation details in the supplementary, if not the main paper.
- 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.
I think the authors responded well to the reviews.
Review #3
- Please describe the contribution of the paper
They studied the tumor auto-labelling process for ultrasound-assisted imaging for the case of missing label by using available patient data. They proposed for various regions of body for metastatic sources.
- 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.
they proposed Personalized tumor marker imputation Population-based imaging features and personalized TM fusion Large-scale multi-origin metastatic classification.
They studied a old but power imaging methodology, ultrasound imaging. Trying to empower standing-still technologies deserves promotion, greeting.
Their proposal not only for one region but also 5 different metatatic sources of body.
They can fill the gap of missing data labeling issue in medical imaging repositories.
Adding Ablation study deserves rewarding.
- 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.
- summary is soo long. Plz shorten it.
- Please share your dataset
- Fig1 and its caption is not enough to express what is desired to tell. Especially colorful dots in CDFI image.
- 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.
- 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/2025/en/REVIEWER-GUIDELINES.html
Share your dataset even your code
- 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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Enough and strong references. Supplementary material is nice.
- 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 sincerely appreciate the reviewers’ constructive suggestions to help improve our paper’s quality, and will explain their concerns point by point below. 1) This paper seems to present a new application of CycleGAN. (R1) Authors: Our study highlights two key contributions: First, the originality of our method lies in being the first to treat clinically discrete parameters as a vector to construct data distributions and propose a new distribution-based imputation method via generative AI, such as CycleGAN, to establish a mapping from one distribution to another, rather than traditional value-based interpolation methods. Second, another original aspect of our study is application of this method to the critical clinical challenge of multi-origin metastatic cervical lymphadenopathy classification. Due to space limitations, the experimental validations of both contributions are presented together for the application performance. 2) Please share your dataset or implementation details. (R1, R2, R3) Authors: We will share the code implementation in the revision. The dataset can be obtained with request. 3) The benefits of the imputation method are less clear than other benefits. (R2) Authors: This study proposes a new data imputation framework for multi-origin metastatic CLA classification by mapping vector-based distributions, different from traditional value-based interpolation methods. The framework uses CycleGAN as a generative tool, achieving an AUC improvement of 0.02-0.033 over value-based imputation methods. Future work will explore advanced generative AI methods (e.g., diffusion models) to further enhance performance. Due to space limitations, we focus on the average performance of five primary tumor sites, where the average F1 improvement is of at least 7.3%, highlighting our method’s effectiveness. 4) The observation is based on a single cervical LN, not containing normal LNs. (R1) Authors: Only 44 of 3,100 cases (1.4%) involved two nodes with different biopsies, and were all allocated to the training, while test set included only one node, minimizing overfitting. Clinically, the prediction of metastasis source is conducted only for confirmed malignant CLA nodes, thus excluding normal LNs. 5) How to deal with the remaining 10% of CLA patients? (R2) Authors: The clinical value of this study aims to help clinicians prioritize primary tumor sites for follow-up examinations, assist pathologists in selecting optimal biomarkers, and reduce clinical resource utilization while streamlining diagnostic workflows. In the future, we plan to employ few-shot learning on the remaining 10% of patients. 6) Concerning nomenclature and the title. (R1) Authors: Our contributions are as follows: We propose a distribution-based imputation method and apply it to a challenging clinical task. Following your advice, we will modify the title to: “Multimodal Fusion Network with Distribution-based Tumor-Marker Imputation for Multi-Origin Metastatic Cervical Lymphadenopathy Classification”. 7) Regarding the writing details: missing data, data division, 95% CIs, summary, terminology consistency and the caption of Fig1. (R1, R2, R3) Authors: Some descriptions in the previous manuscript were ambiguous. We will revise them as follows: First, our large dataset has high missing ratios for ten TMs, which are 54.5%, 3.6%, 57.6%, 39.7%, 36.9%, 84.9%, 60.1%, 36.1%, 53.3%, and 85.9%, respectively. Therefore, data imputation is a challenging and crucial research problem for clinical applications. Next, we split the dataset into training and test with an 8:2 ratio and performed 3-fold cross-validation on the training. Then 95% CIs were derived by stratified bootstrapping with 1000 iterations. Finally, we will condense the summary, thoroughly check terminology for consistency and update the caption: “BUS images were analyzed to extract anatomical semantic features, while CDFI images were analyzed to extract vascular distribution features indicated by colorful dots.”
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
The authors should clarify the points raised by the reviewers in their rebuttal.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
N/A
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
Though the authors have attempted to address the major score driving critiques, several concerns still remain/or are inadequately addressed. For eg: Benefits of the imputation method not clear, unclear clinical framing, cross-validation justification