List of Papers Browse by Subject Areas Author List
Abstract
Accurate classification of adrenal tumors is essential for distinguishing hormonally active tumors from non-functioning adenomas and guiding appropriate management. Current diagnostic workflows typically rely on multi-step endocrine testing and contrast-enhanced CT, which may introduce additional radiation exposure and contrast-related risks. In this study, we investigate four-class adrenal tumor classification—Cushing’s syndrome, primary aldosteronism, pheochromocytoma, and non-functioning adenoma—using non-contrast CT combined with routinely collected non-hormonal clinical variables.This task is challenged by substantial variable-level missingness in structured clinical data and pronounced class imbalance in real-world surgical cohorts. We propose a missingness-aware multimodal learning framework that explicitly encodes missing patterns and conditions cross-modal fusion on input reliability. Structured embeddings modulate imaging features via feature-wise affine transformation to enable interaction, while a reliability descriptor derived from missing patterns predicts adaptive fusion weights to dynamically integrate imaging, clinical, and interaction representations under incomplete inputs. To further address imbalance, we introduce a distribution-aware prototype regularization objective on the fused embedding space, encouraging compact intra-class representations and enlarged inter-class margins.Experiments on a retrospective cohort of 1,108 patients demonstrate that the proposed method achieves a macro-F1 of 0.822, balanced accuracy of 0.825, and AUC of 0.963 on a held-out test set, outperforming state-of-the-art unimodal and multimodal baselines. Notably, the method improves recall for minority tumor categories while maintaining stable performance on the majority class. Code is available at https://github.com/fj-sunny/AMAF-Net.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1243_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/fj-sunny/AMAF-Net
Link to the Dataset(s)
N/A
BibTex
@InProceedings{CheDeh_MissingnessAware_MICCAI2026,
author = { Chen, Dehua AND Fang, Jing AND Wu, Qihong AND Xia, Pengfei AND An, Huimin},
title = { { Missingness-Aware Multimodal Learning for Four-Class Adrenal Tumor Classification } },
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 missingness-aware multimodal learning framework for four-class adrenal tumor classification. It designs an adaptive fusion mechanism that dynamically integrates imaging and clinical data based on missing patterns, introduces distribution-aware prototype regularization to address class imbalance, and achieves state-of-the-art performance on a real-world 1,108-patient cohort.
- 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.Missingness-aware fusion mechanism: It explicitly models the missing patterns of clinical data to dynamically adjust the fusion weights of imaging/clinical modalities. Unlike simple missing value imputation or discarding strategies, this design better aligns with the characteristics of real-world clinical data. 2.Prototype regularization method: It optimizes intra-class compactness and inter-class separation in the multimodal feature space, and achieves superior classification performance for rare tumor classes compared with traditional imbalance solutions such as Focal Loss. 3.Clinically feasible design: It accomplishes four-class adrenal tumor classification using non-contrast CT combined with non-hormonal clinical data, avoiding the risks of contrast agents and complex endocrine tests, thus facilitating clinical translation and implementation. 4.Comprehensive experimental validation and reproducibility: The paper conducts sufficient comparative experiments and ablation experiments to comprehensively verify the effectiveness and robustness of the proposed method. It clearly provides detailed experimental settings, implementation details and evaluation metrics, and also offers a valid anonymous code link, ensuring good reproducibility of the experimental results.
- 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.Insufficient Effectiveness in t-SNE Visualization:The t-SNE plots in Fig. 2 fail to intuitively demonstrate the superiority of the prototype regularization loss (Lproto). Visually, the difference in feature clustering and separation performance between Lproto and Focal Loss is not distinct enough, which undermines the clear illustration of Lproto’s value in optimizing intra-class compactness and inter-class separation. 2.Lack of Framework Innovation:The overall design of the framework leans heavily towards engineering-oriented modular integration and hyperparameter tuning. It lacks breakthrough theoretical innovations or novel multimodal fusion paradigms. The originality of the core mechanisms needs to be further strengthened to enhance the academic contribution of the study.
- 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
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 paper demonstrates strong practicality and process completeness, and can effectively address real clinical deployment demands. The experiments are sufficient and the results are reliable with reproducible code. However, the overall method is still insufficient in theoretical innovation and core novelty, leading to relatively limited originality.
- 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
Review #2
- Please describe the contribution of the paper
This paper proposes a missingness-aware multimodal framework for adrenal tumor classification that combines non-contrast CT with routinely collected clinical data. The method adaptively fuses modalities based on input reliability and incorporates distribution-aware prototype regularization to address class imbalance. Experiments on 1,108 patients show improved macro-F1 and minority-class recall over strong baselines.
- 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 work targets a realistic clinical setting with missingness and class imbalance. The proposed solution carries clinical significance and potential utility. 2.The method explicitly encodes missing patterns and uses a reliability descriptor to adaptively weight modalities, which provides a more effective alternative to standard imputation or static fusion. 3.Experiments show strong empirical validation.
- 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.While the overall framework is well integrated, most components build on existing ideas. Feature-wise modulation follows prior work such as FiLM (Perez et al., 2018), and multimodal interaction with tabular data has been explored in DAFT (Wolf et al., 2022) and HyperFusion (Duenias et al., 2025). Handling missing modalities via adaptive mechanisms is also related to DrFuse (Yao et al., 2024). As such, the contribution lies more in the combination of these techniques rather than a fundamentally new formulation. 2.The use of prototype-based objectives is inspired by prior work such as ProCo (Du et al., 2024), but the paper provides limited intuition for why this specific formulation is particularly suitable for the multimodal fusion setting. The improvement over alternatives (e.g., focal loss) is shown empirically but lacks deeper theoretical or geometric justification. 3.The study is based on a single-center retrospective cohort, which may introduce selection bias and limit generalizability. There is no external validation or cross-institutional testing.
- 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.
(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?
My score is based on the fact that the paper presents a well-motivated clinical problem, a coherent missingness-aware multimodal framework, and strong empirical validation with meaningful gains. However, the methodological novelty is somewhat incremental, and external validation is lacking. The overall quality, clarity, and clinical relevance place it near the acceptance 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 #3
- Please describe the contribution of the paper
The paper proposes a missingness-aware multimodal learning framework for four-class adrenal tumor classification that explicitly models incomplete clinical data and class imbalance. Its key contribution is an adaptive fusion mechanism that encodes missingness patterns to guide cross-modal interaction and dynamically weight imaging and clinical features based on input reliability, making the model robust to real-world incomplete data. In addition, it introduces a distribution-aware prototype regularization in the fused feature space to improve class separability, particularly for minority classes.
- 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.
-
Novel missingness-aware fusion mechanism: The paper explicitly models missing clinical data and uses it to guide cross-modal interaction and adaptive weighting. This is interesting because, unlike prior work that relies on imputation or assumes complete inputs, the model conditions fusion on input reliability, making it more robust to real-world clinical settings.
-
Effective handling of class imbalance via prototype regularization: The proposed distribution-aware prototype loss improves intra-class compactness and inter-class separation in the fused embedding space. This goes beyond standard reweighting (e.g., focal loss) and leads to better minority-class performance, which is critical in medical datasets.
-
Clinically relevant evaluation: The work demonstrates effectiveness on a realistic multimodal setup (non-contrast CT + incomplete clinical data) with substantial missingness and imbalance, supported by comprehensive experiments, ablations, and analysis, including improved recall for underrepresented tumor types.
-
- 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.
-
Limited methodological novelty relative to prior multimodal fusion work: While the paper emphasizes missingness-aware fusion, many components build directly on existing ideas, e.g., FiLM-style modulation (Perez et al., 2018), DAFT (Wolf et al., 2022), and missing-modality handling in DrFuse (Yao et al., 2024). The adaptive weighting via a learned reliability signal is incremental over these works, and the paper does not clearly articulate what fundamentally distinguishes it beyond combining these ideas.
-
Evaluation lacks external validation and generalizability analysis: All experiments are conducted on a single retrospective cohort from one institution. There is no multi-center validation or testing on external datasets, which is critical for clinical deployment. Given known domain shifts in medical imaging, this limits confidence in real-world applicability.
-
Handling of missing data is relatively simple and potentially suboptimal: Missing values are still imputed using basic statistical methods, and missingness is only encoded as a binary indicator. More advanced approaches (e.g., generative modeling, uncertainty-aware imputation, or masked modeling) are not explored, making the approach somewhat limited compared to state-of-the-art missing-data techniques.
-
Clinical utility and comparison to standard workflows are not fully established: Although the paper motivates reducing reliance on hormonal testing and contrast imaging, it does not provide direct comparison with clinical baselines or decision protocols, nor does it evaluate how the model would integrate into practice (e.g., decision curves, cost-benefit analysis, or physician-in-the-loop scenarios). This weakens the claim of practical clinical impact.
-
- 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?
NA
- 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
NA
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a clinically relevant problem and presents a well-motivated multimodal framework that explicitly considers missing clinical data and class imbalance, which are common in real-world settings. The proposed missingness-aware adaptive fusion and prototype-based regularization are reasonable and empirically effective, and the experimental evaluation is fairly comprehensive, including strong baselines, ablations, and analysis of model behavior. However, the overall novelty is somewhat limited, as key components build on existing approaches such as FiLM/DAFT-style modulation and prior missing-modality fusion methods like DrFuse, with the main contribution being a combination and extension of these ideas rather than a fundamentally new formulation. Additionally, the lack of external validation and limited discussion of clinical deployment reduce confidence in generalizability and real-world impact.
- 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
N/A
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.
This paper introduces a missingness-aware multimodal learning framework for four-class adrenal tumor classification, integrating adaptive cross-modal fusion and prototype regularization to address incomplete clinical data and long-tailed class distribution. There are strong clinical motivation, comprehensive ablation studies, reliable performance on a large single-center cohort, and accessible anonymous code as key strengths of this work.