Abstract

Lupus nephritis (LN) is a severe complication of systemic lupus erythematosus that affects pediatric patients with significantly greater severity and worse renal outcomes compared to adults. Despite the urgent clinical need, predicting pediatric LN prognosis remains unexplored in computational pathology. Furthermore, the only existing histopathology-based approach for LN relies on multiple costly staining protocols and fails to integrate complementary clinical data. To address these gaps, we propose the first multimodal computational pathology framework for three-class treatment response prediction (complete remission, partial response, and no response) in pediatric LN, utilizing only routine PAS-stained biopsies and structured clinical data. Our framework introduces two key methodological innovations. First, a Clinical-Injection Transformer (CIT) embeds clinical features as condition tokens into patch-level self-attention, facilitating implicit and bidirectional cross-modal interactions within a unified attention space. Second, we design a decoupled representation-knowledge adaptation strategy using a domain-adapted Masked Autoencoder (MAE). This strategy explicitly separates self-supervised morphological feature learning from pathological knowledge extraction. Additionally, we introduce a multi-granularity morphological type injection mechanism to bridge distilled classification knowledge with downstream prognostic predictions at both the instance and patient levels. Evaluated on a cohort of 71 pediatric LN patients with KDIGO-standardized labels, our method achieves a three-class accuracy of 90.1% and an AUC of 89.4%, demonstrating its potential as a highly accurate and cost-effective prognostic tool.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3844_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{HuaYue_ClinicalInjection_MICCAI2026,
        author = { Huang, Yuewen AND Ye, Zhitao AND Feng, Guangnan AND Zheng, Fudan AND Gao, Xia AND Lu, Yutong},
        title = { { Clinical-Injection Transformer with Domain-Adapted MAE for Lupus Nephritis Prognosis Prediction } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16887},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper targets a typical medical image analysis task with a deep learning–based multimodal framework. Although the topic falls within the scope of MICCAI, the work exhibits severe limitations in novelty, methodological design, experimental validation, and presentation clarity, rendering it unsuitable for publication in its current form.

  • 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 topic is clinically relevant and fits the general scope of MICCAI, and the authors attempt multimodal fusion and validation on a standard clinical dataset.

  • 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] lack of novelty: The proposed framework is a simple combination of existing modules (e. g. , attention, fusion, generative augmentation) without novel designs, new theoretical insights, or specialized mechanisms for medical data. [2] The technical pipeline is ad-hoc and lacks clear motivation. Key designs (e. g. , attention mechanism, feature fusion, generative module) are not well justified or tailored to the clinical task. [3] The writing is unclear, notations are inconsistent, figures are not well-explained, and some technical descriptions are ambiguous or incomplete.

  • 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.

    (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 topic but does not meet the novelty, technical depth, and validation standards required by MICCAI.

  • 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 multimodal computational pathology framework for predicting 12-month treatment response in pediatric lupus nephritis from PAS-stained renal biopsy WSIs and structured clinical data. The authors use a Clinical-Injection Transformer that injects clinical features as a condition token into patch-level self-attention, a decoupled MAE-based knowledge adaptation strategy, and patch/patient-level morphology type injection. The method reports 90.1% accuracy and 89.4% AUC outperforming all baselines suggesting potential value for cost-effective prognostic support in a rare pediatric disease setting.

  • 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.

    Pediatric lupus nephritis prognosis prediction from biopsy images appears substantially underexplored and the paper focuses on a clinically meaningful response task.

    The method uses routine PAS-stained pathology plus structured clinical variables which is more practical and potentially lower-cost.

    The Clinical-Injection Transformer is a clean, parameter-efficient fusion strategy that places clinical and image tokens in a shared self-attention space rather than relying on other heavier dual-stream fusion. The paper also reports the fusion block is lightweight with around 0.56M parameters.

    The decoupled knowledge adaptation and the multi-granularity morphology injection are well motivated and fit the pathology setting well.

    The method improves over image-only MIL baselines, a clinical-only MLP, late fusion, and cross-attention fusion, with especially notable gains in macro-F1 for minority classes.

    The paper separately studies fusion architecture, feature extractor choice, and morphology type injection, which strengthens the technical case.

    Generally clear presentation. The framework overview, module definitions, and training setup are presented clearly enough for the overall method to be understandable.

  • 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 paper does not clearly specify how the 5-class morphology classifier used for type injection was trained relative to the downstream 5-fold CV on the 71 patient prognosis cohort. If morphology predictions for a test fold were generated by a model trained on those patients, that would inflate performance. This should be clarified explicitly.

    The 0m+3m setting includes “treatment response features”, but the exact variables are not described clearly enough to assess how directly they overlap with the 12-month endpoint definition. This matters for interpreting the clinical forecasting claim.

  • 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 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

    The manuscript appears to use pediatric clinical and histopathology data, but the text does not clearly state ethics/IRB approval. This should be made explicit in the final version.

  • 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 strengths are a strong clinical motivation, a method that is more thoughtful than a standard fusion baseline, and convincing internal results supported by focused ablations. The work also has practical appeal because it relies on routine PAS histology and structured clinical data. My main reservations are the small single-center cohort and the fold separation of the morphology classifier used for type injection and the exact content of the 3-month response-related features. These concerns reduce confidence, but they appear potentially clarifiable rather than clearly fatal from the current manuscript.

  • 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 #3

  • Please describe the contribution of the paper

    This is the first work to construct a multimodal computational pathology framework for prognostic prediction in pediatric lupus nephritis (LN). Previous prognostic studies have primarily focused on adult populations, and relevant explorations in computational pathology for the pediatric cohort remain lacking.

  • 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) Existing histopathology-based methods for LN require four costly staining protocols and do not incorporate clinical data. This work represents the first multimodal framework that performs three-class prediction using only routine PAS-stained biopsy images and structured clinical data, demonstrating strong potential for clinical translation. 2) To address the issue that fine-tuning models directly for morphological classification discards subtle texture features useful for prognosis, the authors froze the MAE-pretrained backbone for feature extraction, and injected morphological classification results via one-hot vectors and distribution vectors. Ablation studies confirmed the benefit of this design. 3) To mitigate the small dataset problem, this work projects clinical features into a condition token, which is fed into a shallow 2-layer Transformer together with image patch tokens. This reduces the number of parameters and lowers the risk of overfitting.

  • 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.

    It is not specified whether patches from patients in the test fold were excluded from the pretraining corpus for each fold in the 5-fold cross-validation. Without strict patient-level isolation, the model may have seen test data during pretraining, which could lead to inflated performance estimates. The study data were collected from only 71 patients at a single hospital, with no external multi-center independent validation set. Therefore, the model’s generalization ability across different institutions, staining batches, scanning devices, and patient populations cannot be verified. It is not clarified whether MICE imputation was performed uniformly on the entire dataset, or strictly fitted on the training set and then applied to the test set within cross-validation. The former would constitute data leakage. In the introduction, the authors mentioned mainstream multimodal fusion methods in pathology and claimed that they are prone to overfitting. However, the comparative experiments only include single-modal MIL, simple late fusion, and the cross-attention module of MCAT, without comparison against full end-to-end state-of-the-art multimodal pathology models. At least one complete SOTA multimodal pathology baseline should be added to numerically demonstrate the advantages of the proposed CIT module under small-sample conditions.

  • 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 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.

    (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 work represents the first attempt to construct a multimodal computational pathology framework for prognostic prediction in pediatric lupus nephritis. The authors design several targeted modules to improve model performance, demonstrating certain innovation. However, the experimental validation is still insufficient.

  • 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 and the Area Chair for their careful feedback. We address each concern below.

1.Cross-Validation Protocol (Reviewer #2, Reviewer #3, Meta-Reviewer #2)

Morphology Classifier (Reviewer #2). The type labels injected downstream (Sec. 2.4) are the discrete predictions of the fine-tuned classifier described in Sec. 2.2.For each CV fold, this classifier is retrained on the training and validation splits; its predictions on the held-out test patients are used as the type labels, ensuring no information transfer across folds.

MICE Imputation (Reviewer #3). MICE is applied exclusively to pre-outcome clinical features (0-month and 3-month), strictly excluding outcome labels. Crucially, to prevent feature leakage, imputation is performed independently within each cross-validation fold. The imputer is fitted solely on the training split and transforms the unseen validation/test splits, ensuring no prospective or cross-sample leakage occurs.

2.Clinical Feature Specification (Reviewer #2)

The 59 features (3 demographics, 22 baseline, 17 three-month, 17 deltas) span urinalysis, blood routine, renal biochemistry, immunological markers (C3, C4, ANA, anti-dsDNA), anthropometrics, and disease activity (SLEDAI-2K). The 3-month features and deltas capture early treatment response and are temporally distinct from the 12-month outcome. A more detailed specification will be provided in the camera-ready version.

3.Methodological Novelty (Reviewer #1)

CIT introduces a novel cross-modal paradigm. Existing multimodal pathology methods either use separate cross-attention modules (MCAT) that overfit on small cohorts, or channel-wise modulation (FiLM, DAFT) that applies the same transformation to all patches without spatial selectivity. CIT places clinical features as a condition token within the same self-attention sequence as image patches, so every patch can attend to clinical context and the clinical token can attend to relevant patches—all within a single module. This unified attention space is a structurally different approach to multimodal interaction.

The decoupled MAE addresses a limitation of standard fine-tuning: fine-tuning on an auxiliary diagnostic task narrows representations and discards prognostic features (yielding a +6.0% Acc improvement over standard fine-tuning, Table 2b). Our decoupling—frozen MAE for self-supervised features, discrete labels for morphological knowledge—preserves both.

Multi-granularity type injection exploits the hierarchical nature of pathological assessment. Patch-level one-hot encoding constrains which instances the model attends to, while patient-level type distribution provides global lesion context to the clinical token (+2.3% Acc, +5.5% M-F1, Table 2c).

4.Ethics Statement (Reviewer #2, Meta-Reviewer #2)

We appreciate this important reminder. This study was approved by the Institutional Review Board of Guangzhou Women and Children’s Medical Center. We will include an explicit ethics statement in the camera-ready version.

5.Single-Center Limitation (Reviewer #3)

We agree with this assessment and have acknowledged it in Sec. 4.The 71-patient cohort reflects the inherent rarity of pediatric LN (childhood-onset SLE incidence: 0.3–0.9 per 100,000 child-years); even the largest multicenter biopsy-confirmed pediatric LN studies report ~300 patients [Chan et al., KI Reports, 2023]. We are actively pursuing multi-center collaboration to extend this work.




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.

    I recommend a Provisional Accept because the paper addresses a significant, underexplored clinical need in pediatric lupus nephritis prognosis. It introduces a parameter-efficient Clinical-Injection Transformer (CIT) containing approximately 0.56M parameters, which effectively mitigates the risk of overfitting on small datasets. The framework demonstrates strong quantitative performance by achieving 90.1% accuracy and 89.4% AUC, showing clear improvements over single-modality baselines. While Reviewer 1 raised concerns regarding a lack of novelty and an ad-hoc pipeline , the authors’ use of decoupled MAE knowledge adaptation is well-motivated for preserving prognostically relevant features in this specific pathology context. Reviewers 2 and 3 questioned the evaluation protocol and the limitations of a small, single-center dataset ; however, this cohort size is consistent with the field given the rarity of childhood-onset SLE and the difficulty of acquiring digitized longitudinal biopsy data. For the camera-ready version, the authors must explicitly detail their cross-validation protocol to ensure no data leakage occurred during MICE imputation and morphology classifier training , and they must include a clear statement confirming ethics and IRB approval for the clinical data.



back to top