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

Abstract

Lesion-level prediction of Gleason Grade Group (GGG) from multiparametric MRI (mpMRI) can support prostate cancer risk assessment and reduce unnec-essary invasive biopsies. Existing GGG prediction methods face two major limitations. First, they often overlook non-image information critical for GGG prediction, including age, prostate-specific antigen (PSA), and expert priors embedded in radiology reports. Second, they tend to oversimplify GGG as flat categorical labels, failing to account for its intrinsic hierarchy of primary and secondary Gleason patterns. To this end, we propose a novel Knowledge-Driven Ordinal-Aware Learning (KOAL) framework with three synergistic modules. Specifically, the Clinical-Context Modulation (CCM) module uses clinical variables (e.g., age and PSA) to dynamically modulate discriminative image representations. The Knowledge-Guided Prototype Alignment (KGPA) module leverages an LLM to extract group-specific expert knowledge from training radiology reports and clinical guidelines, producing offline semantic anchors describing grade-specific radiological findings without requiring pa-tient-specific reports at inference. Through prototype contrastive alignment, patient-specific mpMRI representations are matched with these anchors to promote pathology-aligned representation learning. The Hierarchical Ordinal-aware Constraints (HOC) module decouples primary and secondary Gleason pattern prediction and maps their probabilistic outputs to GGG via a Differen-tiable Bio-logic Mapping Layer (DBML), ensuring pathological grading con-sistency. Experiments on public PI-CAI and in-house datasets demonstrate that KOAL outperforms state-of-the-art methods. Code is available at: https://github.com/Gother-GZ/KOAL.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1084_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Gother-GZ/KOAL

Link to the Dataset(s)

N/A

BibTex

@InProceedings{GuoZhe_KOAL_MICCAI2026,
        author = { Guo, Zheng AND Cui, Jiaqi AND Xiong, Haocheng AND Han, Jize AND Liu, Bo AND Zhang, Qianwen AND Chen, Rui AND Wang, Yan},
        title = { { KOAL: Knowledge-Driven Prostate Cancer Grading with Ordinal-Aware Learning } },
        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 a multimodal framework for prostate cancer Gleason Grade Group (GGG) prediction that integrates mpMRI, clinical variables (age and PSA), and radiology reports. The framework consists of three core modules: (1) Clinical-Context Modulation (CCM) for channel-wise early fusion and clinically guided feature recalibration; (2) Knowledge-Guided Prototype Alignment (KGPA), which leverages radiology reports to construct grade-specific semantic anchors; and (3) Hierarchical Ordinal-aware Constraints (HOC), which decouple GGG prediction into primary and secondary Gleason pattern prediction via a dual-stream network. The paper is clearly written and has multiple strengths (see major strength section). It should be noted, however, there are multiple weaknesses that require clarification and additional justification to fully support the claims made.

  • 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.

    Clinically motivated multimodal integration: The integration of mpMRI with clinical variables and radiology reports is well-motivated, as MRI alone does not capture all clinical factors relevant to GGG prediction (e. g. , PSA, age). This is a meaningful step toward a more holistic grading pipeline.

    Ordinal-aware learning: The HOC module’s decomposition of GGG prediction into primary and secondary Gleason pattern streams is an interesting and principled design choice that explicitly respects the ordinal structure of Gleason grading. This is particularly clinically relevant for borderline GG3/GG4 cases where treatment decisions differ substantially.

    Comprehensive evaluation with multiple baselines: The paper compares against a wide range of baselines including multimodal methods (HyperFusion, PPAL, DSPA-MIL, XCoOp, DAFT) and ordinal regression methods (PON, CORAL, AOR-DR), providing a reasonably thorough empirical benchmark.

  • 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.

    Unrealistic inference assumption due to tumor-centered cropping: The framework crops input images around the radiologist-annotated tumor region, implicitly assuming lesion location is known. This may introduce spatial bias: since prostate cancer predominantly arises in the peripheral zone, cropping around annotated lesions may cause the model to learn zone-specific appearance cues rather than discriminative grading features, inflating apparent performance. Furthermore, a significant proportion of positive biopsies originate outside the MRI-targeted regions, meaning clinically relevant cancers detected through systematic biopsy would fall outside the cropped field of view entirely. This limits the generalizability of the framework to the broader patient population encountered in practice.

    Over-engineered architecture with insufficient justification: The framework comprises three complex modules with considerable design overhead. In particular, the necessity of KGPA, constructing grade-specific semantic anchors from radiology reports via BioBERT, is not clearly justified. It is unclear whether BioBERT is pretrained on an in-house dataset or directly transferred, and no ablation clearly isolates the contribution of this component in a convincing way.

    Missing implementation and dataset details: Several critical details are absent: (a) train/validation/test split information is not reported; (b) it is unclear whether the model is trained on in-house data only or also on PI-CAI; (c) AUC values appear to be averaged without reporting standard deviation; (d) radiology report content is underspecified. It is unclear whether reports include patient history or treatment group information, which could introduce confounding.

    Inconsistent and underspecified baseline comparisons: Several baseline methods are not designed for prostate MRI and the adaptation procedure is not described. For example, DSPA-MIL is a pathology method, HyperFusion is designed for simple classification without tabular input (how tabular data was incorporated is unexplained), and it is ambiguous whether inputs are tabular or text for some baselines. Only Ref [17] (DAFT) appears to be a direct methodological match. Unsubstantiated claims: The claim that HOC “encourages the model to learn discriminative and fine-grained visual cues related to dominant and subordinate pathological patterns, rather than simply memorizing global labels” lacks empirical or theoretical evidence.

    CCM clinical grounding is unclear: The paper claims that CCM incorporates clinician reasoning, but it is not demonstrated how clinical variables (age and PSA) actually influence the final GGG prediction quantitatively. A feature importance or sensitivity analysis is needed. Additionally, it is stated that doctors consider tumor aggressiveness, yet this does not appear to be reflected in the GGG prediction target.

    Missing comparison with relevant concurrent work: e. g. , ProstNFound (image + text multimodal prostate model) is a relevant and closely related work that should be discussed or compared against.

    Q: Unexplained performance gap in ablation. The large performance boost in the ablation study on the in-house dataset (component O) is not explained or discussed.

  • 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.

  • 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 meaningful problem and the ordinal-aware Gleason pattern decomposition in HOC is an interesting design choice. However, several factors led to the current score.

    The most significant concern is the unrealistic inference assumption: the framework relies on radiologist-annotated tumor regions as input, which are unlikely to be available in the deployment scenarios the paper targets, and this is not acknowledged as a limitation. Additionally, the architecture lacks sufficient justification across its three modules, particularly KGPA, and critical implementation details, including train/validation/test splits and BioBERT setup, are missing, making results difficult to assess reliably. Several baseline comparisons are also inconsistent, as methods not designed for prostate MRI grading are included without describing their adaptation.

    It is understandable that page limits may have affected the level of detail provided, but the work requires revision in experimental transparency, architectural justification, and honest discussion of its core assumptions before it can be considered for acceptance.

  • 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 addressed the major concerns raised and promised that the discussion will be reflected in the camera-ready version.



Review #2

  • Please describe the contribution of the paper

    The method aims to combine multimodal learning, knowledge priors and ordinal structure modelling into a unified framework for Gleason Grade Group (GGG) prediction from mpMRI. The method is evaluated on the public PI-CAI dataset (n=643) and an in-house dataset (n=1002).

  • 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 addresses the clinically meaningful problem of non-invasive GGG prediction
    • the idea of treating the classification as an ordinary problem instead of a 5-class problem
    • interesting idea of using prior knowledge in the form of an anchor
  • 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 experimental section is very weak and does not support the claim.
    • There is no information on how the training/testing/validation are performed
    • It is not clear what the numbers shown represent. Is it mean values? no std or confidence interval is given
    • No information on the in-house dataset is provided
    • How did the authors use PI-CAI given it does not have medical report associated?
    • No information of GGG results is provided -No inormation of GGG distribution of both cohorts is provided
    • No statistical significance test is provided -Per-class metrics are not provided
    • Important methodological details are missing; the authors did not provide any example of the prompt used, of the structured information exctracted of the LLM-based extrtaction results
    • HyperFusion, DAFT, XCoOp, and DSPA-MIL were all designed for different tasks. No implementation detail is given for adapting them to prostate mpMRI and GGG prediction
    • A trivial yet essential baseline, such as late fusion of clinical variables with image features, is missing
    • The choice of the methods used as comparison is not justified. - - Of the nine compared methods, only one (PON) was originally developed for prostate cancer grading on mpMRI, The prostate mpMRI literature offers a substantial body of directly comparable methods that the authors inexplicably omit.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • 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.

    (1) Strong Reject — must be rejected due to major flaws

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The KOAL framework is conceptually interesting and the design is clinically well-motivated; however, the experimental validation is too weak to substantiate the central claim of state-of-the-art GGG prediction.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    Almost none of my comments were addressed. Also, as experiments are a very important part of evaluating the paper, statistics shouldn’t be omitted for the sake of the page limit.



Review #3

  • Please describe the contribution of the paper

    This paper proposes KOAL, a knowledge-driven ordinal-aware learning framework for non-invasive Gleason Grade Group (GGG) prediction from mpMRI. The method integrates heterogeneous data sources, including mpMRI, clinical variables (e.g., PSA, age), and radiology reports. It consists of three main components: (1) Clinical-Context Modulation (CCM) to incorporate clinical variables into image representation learning, (2) Knowledge-Guided Prototype Alignment (KGPA) leveraging LLM-extracted semantic anchors from radiology reports, and (3) Hierarchical Ordinal-aware Constraints (HOC) that explicitly model the compositional structure of GGG via primary/secondary Gleason patterns. The method is evaluated on both the PI-CAI dataset and an in-house dataset, showing improved performance over existing 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.
    • The paper addresses a clinically important problem, namely non-invasive prostate cancer grading, which has direct implications for reducing unnecessary biopsies.
    • The integration of heterogeneous modalities (imaging, clinical variables, and radiology reports) is well-motivated and aligns with real clinical decision-making processes.
    • The use of ordinal-aware modeling that reflects the hierarchical structure of Gleason grading is a meaningful improvement over flat classification approaches.
    • The idea of leveraging LLM-derived knowledge from radiology reports to construct semantic prototypes is interesting and adds a knowledge-driven perspective to representation learning.
    • The method demonstrates strong performance improvements on both public and in-house datasets (Table 1, page 7), suggesting good generalizability.
  • 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 framework combines multiple components (CCM, KGPA, HOC), each with its own design complexity, making it difficult to clearly attribute performance gains to individual modules beyond the ablation study.
    • The use of LLM-generated knowledge introduces an additional layer of dependency, but the robustness and reproducibility of this process (e.g., prompt design, variability of outputs) are not sufficiently discussed.
    • The construction and quality of radiology report-derived semantic anchors are not quantitatively evaluated, leaving uncertainty about their contribution.
    • The method requires access to radiology reports and clinical variables, which may not always be available or standardized across institutions.
    • While results are strong, comparison with more recent large-scale multimodal or foundation models is limited.
  • 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?

    The paper presents a well-motivated and comprehensive framework that integrates multimodal information for prostate cancer grading. The approach is promising and supported by strong experimental results. However, the complexity of the framework and the reliance on LLM-derived knowledge would benefit from further analysis and clarification.

  • 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?

    The paper addresses an important clinical problem and proposes a comprehensive multimodal framework with meaningful methodological contributions, particularly in ordinal-aware modeling and knowledge integration. The experimental results are strong and consistent across datasets. Despite some concerns regarding complexity and reliance on LLM-based components, the overall contribution is solid and above 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.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    After carefully considering the authors’ rebuttal, I recommend Accept, although my final opinion is closer to a weak or marginal accept rather than a strong accept.

    The rebuttal clarifies several important points that were ambiguous in the original submission. In particular, the authors explain that KOAL is intended as a lesion-level Gleason Grade Group grading model rather than a full detection-and-grading pipeline, and that tumor-centered ROI cropping is used after suspicious-lesion localization. They also clarify that multi-lesion cases are handled at the lesion level, that KGPA anchors are constructed offline and are not dependent on patient reports at test time, and that BioBERT is publicly pretrained. In addition, the rebuttal provides useful clarification regarding the role of PSA and age in the Clinical-Context Modulation module, the motivation of HOC for primary/secondary Gleason pattern decomposition, and the patient-level 80/20 train/test split.

    I still find the paper clinically meaningful and methodologically interesting. The problem of non-invasive prostate cancer grading is important, and the proposed combination of mpMRI, clinical variables, radiology report-derived knowledge, and ordinal-aware learning is well aligned with clinical reasoning. In particular, the hierarchical ordinal-aware modeling of primary and secondary Gleason patterns remains a strong aspect of the paper, because it reflects the compositional and ordinal structure of Gleason grading more explicitly than a flat multi-class classifier.

    However, several concerns remain only partially addressed. First, the reliance on tumor-centered ROI cropping remains an important limitation for clinical translation. Although the rebuttal clarifies that the method is lesion-level rather than a detection pipeline, the original claims should be carefully framed so that the method is not interpreted as a complete clinical screening or diagnostic pipeline. Second, many important experimental details, including GGG distributions, confidence intervals, per-class metrics, confusion matrices, and baseline adaptation details, were missing from the submitted manuscript and are described in the rebuttal as items to be added. Third, the reproducibility of the LLM-derived semantic anchor construction remains somewhat limited in the current submission, although the authors indicate that prompt and BioBERT details will be included. Finally, the contribution of PSA and age through CCM is supported by ablation results, but a more direct feature-importance or sensitivity analysis would still be helpful to substantiate the clinical interpretation.

    Overall, I believe the rebuttal resolves enough ambiguity to support acceptance, but the paper remains borderline. I recommend acceptance mainly because the clinical motivation, ordinal-aware formulation, and multimodal knowledge-driven design are promising and above the acceptance threshold. At the same time, I encourage the Area Chairs to consider the remaining limitations regarding clinical deployment assumptions, experimental transparency, baseline comparability, and reproducibility when making the final decision.



Author Feedback

We thank all reviewers for their constructive comments. Q1: Clinical setting, ROI cropping, and multi-lesion handling(R1) A1: (1) KOAL is a lesion-level GGG grading model, not a detection+grading pipeline. Tumor-centered ROI cropping is used to study GGG prediction after suspicious-lesion localization, consistent with the clinical workflow where radiologists first identify/delineate ROIs. (2) Detection false positives and systematic-biopsy-positive lesions outside MRI-targeted ROIs are upstream detection/sampling issues. We will clarify this scope. (3) For multi-lesion cases, each lesion is treated as one sample; the input is always one lesion ROI. Q2: KGPA, LLM and reports(R1,R2,R3) A2: (1) KGPA is needed because GGG labels provide only category supervision, while radiology reports contain radiologist priors. Class-level semantic anchors provide pathology-consistent references for ordinal GGG representation. Table 2 shows KGPA improves QWK on both datasets. (2) KGPA builds GGG-level population anchors offline and uses no patient reports at test time. Anchors are generated from pre-treatment reports in the In-house training set and clinical guidelines, contain no test-case information, and are also used for report-free PI-CAI, avoiding label leakage. Template-constrained LLM outputs serve as initial priors; anchors are calibrated during training. (3) Reports only provide pre-treatment MRI phenotypes, without pathology, treatment, or prognosis. BioBERT is publicly pretrained. We will add the prompt and BioBERT details. Q3: PSA/Age and CCM(R1) A3: (1) PSA and age are clinical variables commonly used in prostate cancer risk assessment; Ref. [20] in our paper supports the clinical relevance of PSA. We do not use them as independent GGG determinants, but as patient-specific context to modulate mpMRI features via ACMA. The adaptive gate G balances image and clinical-modulated features. (2) Table 2 shows CCM’s value: +10.27% QWK on PI-CAI and +12.82% on In-House. Feature-importance/sensitivity analysis will be considered as future interpretability work. Q4: HOC and ordinal learning(R1) A4: (1) HOC predicts primary/secondary Gleason patterns and maps them to GGG via fixed DBML. It is constrained by pattern-level, grade-level ordinal, and distance-aware ordinal losses, avoiding ordinary 5-class classification. (2) Ablation shows standard ordinal loss improves over M+C+K, and HOC further improves results, supporting primary/secondary decoupling. (3) The larger In-House gain of O is because this heterogeneous cohort is more prone to distant-rank errors, which ordinal loss penalizes. We will clarify this. Q5: Experimental details and reporting(R1,R2) A5: (1) Both datasets use a patient-level 80%/20% train/test split, ensuring no patient overlap. GGG distributions were omitted due to space and will be added. (2) PI-CAI contains GGG1 labels, e.g., 10007_1000007; thus the GGG1–5 setting is valid. (3) Results are reported from one fixed split. AUC is macro-AUC. We also computed 95% CIs, per-GGG confusion matrices, and class-wise metrics; these were omitted due to space and will be included in the final version. Q6: Baselines, related work, and reproducibility(R1,R2,R3) A6: (1) Baselines cover KOAL’s four design axes: prostate mpMRI grading, ordinal learning, multimodal fusion, and prototype/prompt learning. Non-native prostate mpMRI GGG methods used the same split, mpMRI input, GGG target, and metrics; methods supporting tabular variables used PSA/Age, otherwise mpMRI only. Details will be added. (2) We tested simple late fusion with image features+PSA/Age; its QWK 73.23/67.04 is lower than CCM 75.78/68.89, showing the gain comes from clinical-context modulation rather than merely adding variables. (3) We will cite and discuss ProstNFound. Given limited mpMRI-based ISUP/GGG grading work, we compared representative methods related to KOAL. Foundation-model comparison will be noted as a limitation, and code will be released after acceptance.



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 carefully read the study and the reviewers’ input, and I’m confident that the reviewers have identified some of the strainghts and weaknesses of this study. Notebly, the approach is predicting Grade Group from crop ROI within MRI. The approach works as a proof of concept, but cropping the ground truth (without radiologists False positives) makes this approach work better that it would perform in clinical setting. This preprocessing is ensuring better performance that it would be shown clinically. Is not clear who are the authors dealing with patients with multiple lesions which need to be identified in the MRI and report and matched. I think is a great idea to decouple the primary and secondary grade, but as the reviewers mentioned, I would like to see how did the PSA and Age end up, playing in this predicton. Another reviewer pointed to PICAI data and the fact that it doesn’t have reports. nor cancers with Grade Group 1,

  • 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.

    Congratulations!



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 KOAL, a multimodal framework for prostate cancer Gleason Grade Group prediction integrating mpMRI, clinical variables, and radiology-report-derived semantic priors. Reviewers generally agreed that the problem is clinically important and that the proposed ordinal-aware formulation is well motivated. In particular, the decomposition of GGG prediction into primary and secondary Gleason pattern prediction was viewed as a meaningful design choice that better reflects the ordinal and compositional structure of Gleason grading than standard flat classification approaches.

    The rebuttal successfully clarified several important ambiguities in the original submission. Most notably, the authors clarified that KOAL is intended as a lesion-level grading framework rather than a complete detection-and-grading pipeline, which better contextualizes the use of tumor-centered ROI cropping. The rebuttal also clarified the construction and use of KGPA semantic anchors, including that anchors are generated offline from training reports and guidelines and are not dependent on patient reports at inference time, helping address concerns regarding information leakage and deployment feasibility. Additional clarification was also provided regarding the patient-level train/test split, baseline adaptation strategy, the role of PSA and age in CCM, and the rationale for the ordinal-aware HOC formulation.

    At the same time, several weaknesses remain. Important experimental and reproducibility details, including GGG distributions, confidence intervals, per-class metrics, confusion matrices, prompt details, and baseline adaptation specifics, were missing from the submitted manuscript and are primarily addressed in the rebuttal rather than the paper itself. The reliance on lesion-centered ROI cropping also remains an important limitation for broader clinical deployment, even if the rebuttal appropriately narrows the scope of the method. In addition, the reproducibility of the LLM-derived semantic anchor construction and the interpretability of the clinical-context modulation mechanism could be better supported.

    Overall, however, the rebuttal resolves a substantial portion of the original concerns and clarifies several misunderstandings regarding the scope and methodology of the work. While the paper remains somewhat limited by incomplete reporting and deployment assumptions, the multimodal formulation, ordinal-aware modeling strategy, and integration of report-derived semantic priors constitute a meaningful contribution that is above the acceptance threshold.



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.

    Although the experimental validation still needs improvement, reviewers agreed that the paper is conceptually interesting.



back to top