Abstract

Accurate prediction of radiation-induced lymphopenia (RIL), a common complication of radiation therapy (RT), is clinically crucial for ensuring the safety of cancer treatment. However, accurately predicting RIL before RT is highly challenging due to the complexity of immune damage and various input data. In this study, we propose a novel multimodal learning framework named RadKAM to predict RIL severity using heterogeneous data, including CT images, dose maps, and meta-data. The proposed RadKAM leverages a “divide and conquer” strategy to learn the multimodal representation and model the dose-damage relationship for RIL prediction in an end-to-end framework. For the first time, an Attention-driven Kolmogorov-Arnold Fusion (AKaF) scheme is designed by injecting modality-adaptive attention into KAN for intra- and inter-modality interactions. Specifically, RadKAM is constructed with Multimodal Interactive AKaF (MI-AKaF) and Cross-modality Guided AKaF (CG-AKaF) to capture features related to lymphocyte-related organs, and model the dose-damage relationships by multimodal feature interactions. By leveraging the advantages of nonlinear representation, RadKAM effectively models the complex interactions of heterogeneous multimodal data, resulting in a comprehensive representation for RIL prediction. Extensive experiments validate the effectiveness of the proposed RadKAM framework, demonstrating its ability to accurately predict RIL severity through multimodal learning.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2378_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{ZhaRon_RadKAM_MICCAI2025,
        author = { Zhao, Rongchang and Wu, Zhangyue and Zhang, Jian and Zhang, Zijian and Li, Shuo},
        title = { { RadKAM: Attention-Driven Kolmogorov-Arnold Model for Automatic Radiation-Induced Lymphopenia Prediction by Multimodal Learning } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15974},
        month = {September},
        page = {519 -- 529}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes RadKAM, a novel multimodal learning framework for predicting radiation-induced lymphopenia (RIL) severity in nasopharyngeal carcinoma patients. Key contributions include:

    1. Attention-driven Kolmogorov-Arnold Fusion (AKaF): A novel fusion scheme integrating modality-adaptive attention with KAN to model intra- and inter-modality interactions.
    2. End-to-End Multimodal Learning: RadKAM avoids manual region-of-interest (ROI) delineation and leverages CT images, dose maps, and meta-data for RIL prediction.
    3. Clinical Relevance: Demonstrates improved performance (AUC: 87.32%) in fine-grained RIL severity classification (G2/G3/G4), addressing limitations of prior binary classification approaches.
  • 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. Novelty of AKaF: The integration of attention mechanisms with KAN for multimodal fusion is innovative, addressing modality heterogeneity and complex dose-damage relationships.
    2. Technical Rigor: Extensive comparisons with existing methods (e.g., MMTM, HeCNN) demonstrate RadKAM’s superiority in multimodal learning (Table 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.
    1. Limited Dataset: The study uses only 211 patients, with imbalanced classes (40 G2, 114 G3, 57 G4). Larger datasets are needed to validate generalizability.
    2. Lack of External Validation: In the absence of more data verification, the paper can consider conducting experiments on more public datasets to prove the effectiveness and generalization of the method.
    3. Incomplete Baseline Comparison: There are two aspects of baseline missing: (1) There is no comparison with the single-modality SOTA method. In order to highlight the role of multi-modality, it is possible to compare the results with the single-modality SOTA method. (2) There is no comparison with the previous results based on CNN and Transformer models.
  • 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.

  • 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
    1. Clarify how the model handles modality misalignment (e.g., spatial registration between CT and dose maps).
    2. In the module Cross-modality Guided AKaF in Fig.2, why use meta data as Q in the cross-attention paradigm? Will using image data as Q produce better results?
  • 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?

    Strengths: AKaF’s integration of attention and KAN is a notable technical advancement. The clinical focus on fine-grained RIL prediction addresses an unmet need. Weaknesses: Limited data and incomplete baseline comparisons reduce confidence in broader applicability.

  • 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 paper proposes a learning framework to predict radiation-induced lymphopenia by leveraging a “divide and conquer” strategy to model the multimodal representation of the dose–damage relationship.

  • 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 study demonstrates strong clinical relevance in the context of current radiotherapy practices. Furthermore, it includes comparative results with relevant prior studies, effectively positioning its contributions within the existing literature.

  • 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 use of an internal dataset raises the potential risk of overfitting. Additionally, a practical strategy for handling missing data is needed to enhance the robustness of the approach.

  • 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 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.

    (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?

    This study demonstrates strong clinical relevance within the context of current radiotherapy practices. It also provides comparative results with relevant prior studies, effectively situating its contributions within the existing body of literature. However, the use of an internal dataset introduces a potential risk of overfitting. Moreover, implementing a practical strategy to address missing data would further strengthen the robustness of the proposed approach.

  • 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.

    The authors appropriately address all of my concerns.



Review #3

  • Please describe the contribution of the paper

    This paper introduces RadKAM, a method for predicting severe radiation-induced lymphopenia (RIL) using multimodal clinical data, including CT images, dose maps, and metadata. Through carefully designed ablation studies, the proposed RadKAM algorithm demonstrates its superiority in predicting RIL, achieving clinically meaningful performance with AUC and F1 scores exceeding 0.8.

  • 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 motivation to combine multimodal data (e.g., CT images, dose maps, and clinical metadata) is plausible, as the etiology of RIL is likely complex rather than attributable to a single cause.
    • Although the feature extraction modules appear relatively simple, the RadKAM module for integrating multimodal information proves to be both effective and essential for accurate prediction. In my opinion, this finding is instructive for the research community, as it highlights the critical role of information fusion in clinical outcome prediction.
  • 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.
    • This study design utilized only two datasets—a training set and a testing set. Therefore, although the RadKAM module appears effective in the model selection process, the validity of its final performance for clinical impact is difficult to confirm due to the potential risk of overfitting.
  • 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.

  • 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.

    (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?

    An inherently difficult task—requiring consideration of multimodal effects across several clinical parameters—was effectively addressed through the introduction of RadKAM, which achieved strong performance. This contributes meaningfully to both technological and clinical domains.

  • 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 thank all reviewers for their consistently positive comments and support of our paper, by appreciating “This study demonstrates strong clinical relevance within the context of current radiotherapy practices”(R1), “This contributes meaningfully to both technological and clinical domains”(R2), and “AKaF’s integration of attention and KAN is a notable technical advancement. The clinical focus on fine-grained RIL prediction addresses an unmet need”(R3). Our RadKAM is the FIRST work to combine Modal-adaptive Attention with KAN for predicting RIL, and is also the FIRST end-to-end RIL prediction model without manual Region-Of-Interest delineation. Q1: Limited internal dataset and Potential overfitting.(R1&R2&R3) A1: 1) Our dataset is collected from two hospitals (multi-center data) with 211 patients (31650 CT slices and 2773 metadata sub-items). The multi-center cross-ages dataset, which will be released to the public, provides sufficient diversity to avoid the potential overfitting. It is currently the largest dataset in the RIL prediction field that includes CT images, dose maps, and metadata simultaneously. 2) There is no publicly available dataset in the field of RIL prediction. Once released, our dataset will become the FIRST publicly available one. 3) Some optimization strategies such as 5-fold cross-validation, L2 norm, and cosine annealing have also been used to avoid overfitting risk. Q2: Missing data.(R1) A2: The primary focus of our research is not on handling missing data. It is important to note that our dataset is comprehensive, encompassing data from all relevant modalities. This work focuses on multimodal learning for RIL prediction, and contributes “divide and conquer” multimodal fusion strategy and Attention-driven Kolmogorov-Arnold Fusion (AKaF) scheme. Q3: Incomplete baseline comparison.(R3) A3: Current baseline comparison already demonstrates the novelty and advantages of our RadKAM on RIL prediction task: 1) Modal Effectiveness Study (Tab.1) and relevant study[5] clearly indicate the advantages of multimodal learning for RIL prediction task. Results show that it is difficult to achieve accurate RIL prediction using only single-modal data. Tab.1 shows that the combination of any two modalities increases AUC and F1 by an average of 6.48% and 7.04%, respectively, compared to the optimal single-modal baseline, and the combination of three modalities increases AUC and F1 by an average of 14.59% and 17.88%, respectively. 2) RadKAM is the FIRST to use CNN and Transformer for RIL prediction. No prior work has used CNN or Transformer for this task. 3) In Tab.3, multimodal fusion models such as HcCNN[1] (based on CNN) and MDLNet[12] (based on Transformer) are compared as baseline models in the comparative experiment. Q4: How the model handles modality misalignment(e.g., spatial registration between CT and dose maps)?(R3) A4: Spatial registration between CT and dose maps is conducted with the Pinnacle or Eclipse Treatment Planning System, as a preprocessing step of the proposed RadKAM in our implements. Dose maps share the same coordinate system and spatial location information with the corresponding CT images. Each dose value corresponds to the voxel at the corresponding spatial location in the CT image. Q5: Why use metadata as Q in Cross-modality Guided AKaF (CG-AKaF, Sec.2.2, Fig.2)?Will using image data as Q produce better results?(R3) A5: 1) Because metadata as Q in CG-AKaF provides balanced feature interaction and enhanced feature expression, avoiding information suppression caused by severe differences between modalities. This is the selection we obtained through experimental results. The results in Tab.2 also confirm that it significantly improved RIL prediction performance compared to baseline, with average AUC and F1 increasing by 4.92% and 9.25%, respectively. 2) Using image data as Q is not the best choice in this task because it suppresses the effective information provided by metadata.




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”.

    N/A

  • 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. 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’

    The rebuttal address the issues from reviewers from some extent, but the answers to some questions raised by reviewers are not convincing. The paper also omitted important references.



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    The paper presents a clinically relevant study with well-situated contributions and comparative results. While initial concerns were raised about overfitting risk and missing data handling, the authors have appropriately addressed these in the rebuttal. I recommend acceptance.



back to top