Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Tongue diagnosis, a convenient and noninvasive traditional diagnostic method in Traditional Chinese Medicine (TCM), provides an important tool for early health screening. Tongue images not only reveal TCM syndrome patterns but also allow a preliminary assessment of relevant organ health. However, most existing methods face three major limitations: (i) syndrome patterns and organ states prediction are often treated as independent tasks, ignoring their coupled mechanisms; (ii) tongue region-dependent signs are insufficiently integrated with global tongue appearance, resulting in inadequate attention to salient local cues; and (iii) clinical priors and TCM knowledge are underutilized, constraining clinically grounded reasoning and interpretability. To address these limitations, we propose an MLLM-Enhanced Region-Aware Bidirectional Evidence-Based Model for tongue diagnosis. Specifically, first, an Augmented Global-Local Feature Fusion (AGLFF) module is introduced to reconcile holistic tongue context with regional cues by mutually refining global and local representations, strengthening region-sensitive evidence extraction. Second, an Uncertainty-Weighted Bidirectional Mixture-of-Experts (UWBMoE) module is designed to propagate syndrome-level and organ-level information, thereby stabilizing multi-task learning and cross-level reasoning. Moreover, a multimodal large language model (MLLM) is incorporated to enrich semantic representations and improve alignment between visual evidence and clinically meaningful concepts. Experiments demonstrate that the proposed approach outperforms state-of-the-art baselines on both syndrome patterns and organ states prediction tasks. The framework provides clinically relevant, evidence-based decision support for noninvasive screening by jointly modeling syndrome-organ associations. The source code is publicly available at https://github.com/NJU-MedAI-Lab/CycleTCM.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0903_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/NJU-MedAI-Lab/CycleTCM

Link to the Dataset(s)

tongue dataset: https://github.com/tonguedx/tonguedx

BibTex

@InProceedings{DuYiw_MLLMEnhanced_MICCAI2026,
        author = { Du, Yiwei AND Lu, Chenyang AND Fu, Baochen AND Fang, Yuqi AND Shan, Caifeng},
        title = { { MLLM-Enhanced Region-Aware Bidirectional Evidence-Based Model for Tongue 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

    This paper proposes an MLLM-enhanced region-aware bidirectional evidence-based model for tongue diagnosis. Different from previous methods that model syndrome patterns or organ status in isolation, this study fully characterizes their intrinsic correlation through unified representation learning and joint inference, and verifies the robustness of the proposed method via experiments.

  • 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 authors accurately capture region-specific tongue signs and their global contextual features via region-aware global-local evidence aggregation, generating richer fused representations. 2.Through a bidirectional evidence interaction and complementation mechanism, the authors strengthen the syndrome-organ joint modeling capability, improve cross-level reasoning performance, and enhance the consistency of joint decision-making. 3.The prior knowledge derived from MLLM adopted by the authors can improve the accuracy of semantic alignment, producing prediction results with higher interpretability and clinical significance.

  • 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.The authors should further elaborate the implementation details of the comparison methods to highlight the performance gain brought by quantitative syndrome-organ coupling modeling. 2.The GradCAM visualization results in Figure 2 should be interpreted from the perspective of Traditional Chinese Medicine (TCM) theory to verify the feasibility of the method. 3.The authors should further elaborate the mitigation strategies for the class imbalance problem in the dataset. 4.The authors state that the proposed framework provides a practical, noninvasive and scalable solution for screening-oriented clinical decision support, yet relevant validation on the model’s computational efficiency and the consistency with TCM practitioners’ diagnostic results is absent.

  • 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.

    (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?

    This paper addresses the long-standing problem of isolated modeling of syndrome patterns and organ status in existing intelligent tongue diagnosis research, proposes a technically feasible MLLM-enhanced region-aware bidirectional evidence-based framework that outperforms multiple state-of-the-art baselines on target tasks, and conducts valuable exploratory attempts to improve the clinical interpretability of artificial intelligence models aligned with Traditional Chinese Medicine (TCM) theories, demonstrating clear innovation and practical application potential. The limitations include insufficiently detailed experimental analysis, inadequate validation of clinical consistency between model outputs and TCM practitioners’ diagnoses, among other issues, which collectively determine the final assessment conclusion for the paper.

  • 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 focuses on tongue image analysis for Traditional Chinese Medicine (TCM), addressing a multi-label tongue attribute recognition task involving both syndrome pattern identification and organ-state assessment. The authors identify three main limitations in existing methods: (1) syndrome and organ-state prediction are typically treated as independent tasks, without modeling their intrinsic coupling; (2) insufficient integration of region-specific tongue signs with global appearance leads to weak representation of salient local cues; and (3) limited incorporation of TCM clinical priors constrains interpretability and clinically grounded reasoning. To address these challenges, the paper proposes an MLLM-enhanced region-aware bidirectional framework. The method consists of three key components: (1) an Augmented Global-Local Feature Fusion (AGLFF) module that integrates global tongue context with region-level features; (2) an Uncertainty-Weighted Bidirectional Mixture-of-Experts (UWBMoE) module that enables bidirectional information exchange between syndrome-level and organ-level tasks for joint inference; and (3) the integration of a multimodal large language model (MLLM) to inject clinically relevant semantic priors into the visual representation. Experiments on tongue diagnosis benchmarks show that the proposed method achieves superior performance over existing state-of-the-art approaches on both syndrome pattern and organ-state prediction tasks.

  • 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.Clinically meaningful problem formulation with coupled task modeling. The paper addresses a relevant problem in TCM tongue diagnosis by jointly modeling syndrome pattern identification and organ-state assessment. Unlike prior approaches that treat these tasks independently, the proposed framework explicitly captures their intrinsic relationship, which is well aligned with clinical reasoning and improves the coherence of multi-task predictions. 2.Well-structured region-aware global–local feature fusion design (AGLFF). The proposed AGLFF module provides a systematic mechanism to integrate global tongue appearance with region-specific cues. The combination of multi-branch feature extraction, gated fusion, and cross-attention-based class token enhancement is well motivated and clearly reflected in the implementation. This design effectively strengthens the representation of salient local patterns while preserving global contextual information. 3.Principled bidirectional interaction via uncertainty-aware MoE (UWBMoE). The UWBMoE module introduces a technically sound approach to cross-task interaction. By leveraging cosine similarity to estimate uncertainty and using it to guide bidirectional feature fusion, the model dynamically balances information from syndrome and organ branches. The conditional Top-K Mixture-of-Experts further enables adaptive and selective feature refinement, providing a flexible mechanism for cross-level reasoning. 4.Clear and consistent alignment between methodological design and implementation. The proposed components (AGLFF, UWBMoE) are not only conceptually described but also concretely implemented with well-defined operations (e.g., gating, cross-attention, conditional routing). This strong correspondence between the method description and the code enhances reproducibility and makes the technical contributions more transparent and verifiable. 5.Empirical effectiveness on both tasks. The method demonstrates improved performance over prior approaches on both syndrome and organ prediction tasks, suggesting the potential benefit of joint modeling and cross-level interaction.

  • 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.Incomplete realization of the MLLM component. While the paper highlights the integration of an MLLM to provide clinically grounded semantic priors, this component is not clearly realized in the provided implementation. In particular, the code does not include an explicit pipeline for generating, encoding, or injecting MLLM-derived features into the model. Instead, the current implementation mainly operates within the visual feature space, with no concrete mechanism demonstrating how external semantic knowledge is incorporated. This gap makes it difficult to assess the actual contribution of the MLLM component, and raises concerns about whether the claimed benefits (e.g., improved interpretability and semantic alignment) are substantiated in practice. 2.Strong dependency on region-specific inputs, with unclear practicality and additional complexity. The model requires multiple region-specific inputs for the organ branch, constructed as: x_organ = torch.cat([x_heart_lung, x_spleen, x_liver, x_kidney], dim=1). This design implies that organ-level regions must be pre-extracted before being fed into the network. However, the paper does not clearly describe how these region-specific inputs are obtained. In practice, this likely requires an additional segmentation or localization pipeline to generate organ-specific regions, which introduces extra annotation requirements and preprocessing complexity. Such a dependency may limit the practicality and scalability of the method in real-world settings, particularly compared to approaches that operate directly on whole images without requiring structured region decomposition. It also raises questions about whether the reported performance gains are partially attributable to the use of additional region-level supervision.

  • 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?

    The paper addresses a clinically relevant problem and proposes a well-structured framework for joint syndrome–organ modeling with reasonable technical design. However, several aspects currently limit the strength of the contribution. In particular, the role and practical integration of the MLLM component are not clearly demonstrated in the implementation, making its impact difficult to assess. In addition, the reliance on region-specific inputs is not sufficiently discussed, and its implications for practicality and scalability remain unclear. With further clarification on these aspects—particularly regarding the actual contribution of the MLLM component and the data requirements for region-level inputs—the work could be significantly strengthened.

  • 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 corrected a typo and stated that a subject‑disjoint split was used. They provided a concrete description with the updated code and clarified that the region‑specific inputs are obtained via simple rule‑based partitioning without extra supervision. However, whether such simple partitioned regions can be meaningfully associated with organ‑level features is not adequately explained by the authors.



Review #3

  • Please describe the contribution of the paper

    This paper proposes a unified tongue-diagnosis model that jointly predicts syndrome attributes and organ abnormality labels. The method combines three ideas: global-local feature fusion, uncertainty-weighted bidirectional interaction between syndrome and organ branches, and MLLM-derived semantic priors. The empirical results on TongueDx are reasonably strong for this dataset, and the ablation table suggests that each added module helps. The overall direction is sensible and the application is clinically interesting.

  • 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 paper targets a meaningful problem. Noninvasive screening support is a legitimate clinical motivation, and joint syndrome-organ modeling is more realistic than treating the two tasks as unrelated.
    • The method is coherent. AGLFF, UWBMoE, and the semantic-prior module fit the stated goal of combining local signs, global context, and cross-task evidence.
    • The empirical tables are better than many weak medical-AI submissions. The paper reports both syndrome and organ results, includes multiple baselines, and provides a component ablation table.
    • The bidirectional coupling between syndrome and organ branches is the most interesting part of the paper. Even if incremental, it is more specific than a generic multitask-learning story.
    • The paper makes an effort on interpretability through GradCAM and syndrome-to-organ Sankey visualization.
  • 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 novelty is moderate. Region-aware local modeling, organ-oriented tongue analysis, and multimodal semantic fusion all have close precedents in recent tongue-diagnosis literature. The main novelty is the combination, not a clearly new core idea.
    • The dataset protocol is not convincing enough for a medical-imaging venue. TongueDx has 5109 images from only 465 subjects, but the paper reports only a sample-level split count. If the split is not subject-disjoint, leakage is a serious risk and the reported gains may be overstated.
    • The clinical ground truth is underexplained. The paper does not clearly describe how the five organ-abnormality labels were obtained or validated, which makes it hard to judge whether the organ task is clinically meaningful.
    • The MLLM part is under-specified. The paper names Qwen3-VL-4B-Instruct, but does not clearly explain the prompts, inputs, outputs, freezing strategy, or why an MLLM is necessary compared with simpler semantic priors.
    • The experimental evidence is still narrow. There is no external validation, no multi-center setting, no robustness analysis across devices or demographic groups, and no statistical significance test for the claimed gains.
    • The baseline comparison is not fully persuasive. Several compared methods were designed for different tasks, and the paper does not show a strong joint-model baseline that isolates whether the gains come from the proposed bidirectional design or simply from giving the model more task structure.
    • The interpretability claims are weaker than they appear. GradCAM and Sankey plots are descriptive visualizations; they do not validate that the learned syndrome-organ associations are clinically correct.
  • 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?

    The paper has a reasonable motivation, a technically coherent design, and promising results on TongueDx. However, for acceptance, I would want stronger evidence that the contribution is more than an incremental assembly of known components, and much stronger validation of the data protocol and clinical labels. The subject-split ambiguity alone is a major concern. If the authors can clarify the split, explain the organ-label provenance, and justify the MLLM module more rigorously, the paper would become much more competitive.

  • 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 author has addressed most of my concerns. So I recommend accepting this paper.



Author Feedback

We thank all reviewers for feedback: “higher interpretability and clinical significance”(R1), “empirical effectiveness”(R2), and “more realistic joint syndrome-organ modeling”(R4). We address concerns below and will release source code.

1.Dataset splitting (Meta&R4Q2) We sincerely apologize for the typo. The number of subjects should be 4650, not 465.Correct description is “TongueDx contains 5109 tongue images from 4650 subjects”. The dataset is split at subject level, not image level. All images from same subject are assigned to only one of the train/validation/test set to avoid data leakage.

2.Ground truth and computational efficiency (R4Q3, R1Q4) Five organ-abnormality labels are diagnosed by expert TCM physicians. Our model achieves 86.05% syndrome accuracy and 80.09% organ accuracy, aligning with TCM diagnoses (ground truth), while requiring only 5.1ms per sample for inference.

3.Details of MLLM (R4Q4, R2Q1) Our system prompt defines MLLM as a TCM tongue-diagnosis expert for analyzing tongue attributes and organ abnormalities. The whole tongue image and TCM semantic syndrome descriptions are fed into frozen MLLM (Qwen3-VL-4B-Instruct) together with system prompt. We extract last hidden-state features, project them with a trainable adapter (MLP), and fuse them into syndrome and organ branches. Unlike text-only semantic methods such as BERT, MLLM jointly encodes tongue image and textual priors, yielding fused features that better align visual signs with TCM knowledge. Code has been updated.

4.Interpretation (R4Q7, R1Q2) GradCAM: In Fig 2, highlighted regions generally align with TCM tongue theory. For example, for syndromes, TonguePale and TipSideRed correspond to pale areas and reddish tongue tip/edges, respectively, aligning with[3]. For organs, tongue tip, middle, and root correspond to Heart/Lung, Spleen, and Kidney, aligning with[9]. Sankey: For example, highest-weight TonguePale-Liver connection aligns with studies linking abnormal tongue color to liver diseases, e.g. fatty liver[1]. Second-highest-weight ToothMark-Spleen connection agrees with studies linking tooth-marked tongue to Spleen deficiency[2]. [1]DOI: 10.1007/s11655-020-3275-4 [2]DOI: 10.3389/fphys.2022.847267

5.Novelty clarification (R4Q1) We agree that region-aware strategies and multimodal fusion have been explored. Rather than simply extracting regional features or performing multimodal fusion, our model refines global context and local regional cues, and uses MLLM to align visual evidence with TCM priors, which existing works overlook. Moreover, following TCM holistic reasoning, this is the first work modeling syndrome patterns and organ states as coupled factors through bidirectional joint inference rather than independent tasks.

6.Comparative and ablation experiments (R4Q6, R1Q1) Since bidirectional syndrome-organ modeling remains unexplored, we use recent SOTA methods from the most closely related tasks as baselines. For fair comparison, we adopt their models and replace final classification heads for prediction. Ablation study validates the effectiveness of bidirectional structure: introducing bidirectional structure improves all metrics for both tasks (see 1st and 3rd rows of Table 3).

7.Clarification of region-specific inputs (R2Q2) Our region-specific inputs are generated from the whole tongue image with a simple rule-based partition: using image center as reference and partition the tongue into upper/lower, left/right strip regions and a central region as organ-related cropped regions. This requires no extra supervision, segmentation model, or manual annotation.

8.External validation(R4Q5) TongueDx is the only public dataset supporting both syndrome and organ tasks. Note that it includes images from multiple devices and diverse age groups, providing certain diversity. Future work involves multi-center validation to evaluate generalization.

9.Imbalance(R1Q3) We use the inverse positive sample ratio as class weight in weighted BCE loss.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • 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 rebuttal rather than early accept, deviating from R1 (5/Accept) and R2 (4/Weak Accept) and aligning more closely with R4 (3/Weak Reject). The deviation is justified by a paper-grounded concern that R1 and R2 did not test, but which would materially overturn the headline empirical claim if confirmed. Per MICCAI guidelines, mixed scores (5, 4, 3) do not satisfy the “all ≥4” rule for early accept, and one of the negative concerns is decisive enough to require a rebuttal cycle for the following reasons.

    1.Subject-disjoint splitting is not stated and is almost certainly violated. The paper reports a 3,371/843/895 sample-level train/val/test split (Sec. 3.1) on the TongueDx dataset, which the authors describe as 5,109 images from 465 subjects — i.e., ~11 images per subject on average. Sec. 3.1 is silent on whether the split is subject-disjoint. Under any random sample-level split with this image-to-subject ratio, the same patient almost certainly appears in train, val, and test. If true, all reported numbers are inflated. This is the central concern raised by R4, and reading the paper directly confirms it as a verifiable gap rather than a subjective worry. 2.The reported gains are modest enough to disappear under correct splitting. The proposed method beats the second-best baseline by only +1.60 F1 on syndrome and +0.47 F1 on organ tasks (Table 1) — well within the range that subject leakage could account for. Until the splitting protocol is clarified, the empirical contribution is unverifiable.

  • 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.

    Two reviewers support acceptance based on the novelty and clinical relevance of the joint modeling design, while R4 raises concerns about data-protocol transparency. I agree with the majority view that the methodological contribution justifies acceptance, but the following should be clarified in the camera-ready: (i) confirm whether the 3,371/843/895 train/val/test split on TongueDx (5,109 images / 465 subjects, Sec. 3.1) is subject-disjoint, and if not, re-run under subject-disjoint splitting; (ii) document the MLLM (Qwen3-VL-4B-Instruct) usage, including prompts, freezing strategy, language, input format, and how H_MLLM is injected, with an ablation isolating the MLLM contribution from the adapter capacity; (iii) provide annotation provenance for the five organ-abnormality labels, describe how syndrome-local and organ-local image branches are constructed, and report statistical significance tests on the gains given the small subject pool.



Meta-review #2

  • 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.

    This paper proposes an MLLM-enhanced region-aware bidirectional evidence-based model for tongue diagnosis. Different from prior work that models syndrome patterns and organ status in isolation, the study jointly learns their intrinsic correlation through unified representation learning and joint inference. The topic is clinically important and the bidirectional evidence-based design is well motivated.

    During the rebuttal phase, the authors satisfactorily clarified the critical subject-disjoint splitting protocol, resolving a central concern that threatened the validity of the main results. Reviewer #4 and Reviewer #1 now explicitly recommend acceptance, and Reviewer #2 raises only limited, addressable concerns regarding the explanation of how partitioned regions are associated with organ-level features.

    Given that the major experimental validity issue has been resolved and the majority of reviewers support acceptance, I recommend accept. For the camera-ready version, the authors must: (1) reinstate confidence intervals and statistical tests; (2) substantially improve the clarity of the methodology, especially the association between region-based partitioning and organ-level feature semantics; (3) add or properly discuss appropriate multimodal or multi-region baselines; and (4) ensure that all standard experimental details and data splitting procedures are explicitly reported.



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.

    The rebuttal resolved the main concerns sufficiently. In particular, the authors clarified the subject-level split, correcting what was previously the most serious issue, and explained the provenance of the organ labels as expert TCM annotations. They also provided a concrete description of how the MLLM is used in the pipeline and clarified that the region-specific inputs are obtained via a simple rule-based partition rather than requiring additional supervision or segmentation models. Some concerns remain, especially regarding moderate novelty, limited external validation, and the clinical meaning of the rule-based organ partitions. However, these do not outweigh the paper’s strengths.



back to top