List of Papers Browse by Subject Areas Author List
Abstract
Placental pathological image classification plays a pivotal role in revealing pathological mechanisms underlying pregnancy-related diseases and supporting reliable clinical diagnosis. Existing whole slide image (WSI) classification methods often struggle to capture the substantial visual diversity across WSIs within a dataset, which limits their generalization ability. Additionally, current vision-language models lack effective semantic alignment between patch-level visual features and global contextual information, leading to insufficient fusion of visual and semantic cues. To address these challenges, we propose ProtoSG-Net, a Prototype-guided Semantic Graph learning network for WSI classification. Specifically, we design a non-parametric prototype generation module to model dataset-level representative visual diversity, guiding the key patch selection. Subsequently, a prototype-guided semantic calibration module is introduced to learn the relationships between selected key patches. The calibration module refines the semantic hierarchy of pathological attribute texts and further enhances the alignment between visual and textual features.
Extensive experiments are conducted on a self-constructed placental dataset and 2 public datasets, demonstrating the effectiveness and efficiency of the proposed method. Code is available at https://github.com/ECNU-MultiDimLab/ProtoSG-Net.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2839_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{GaoZix_Nonparametric_MICCAI2026,
author = { Gao, Zixuan AND Zhang, Qing AND Guo, Hang AND Wang, Yan AND Li, Qingli},
title = { { Non-parametric Prototypes Enable Semantic Graph Learning for Whole-Slide Pathology } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16891},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes ProtoSG-Net to address the insufficient modeling of dataset-level visual diversity and the lack of fine-grained vision-language alignment in WSI classification. The core idea involves capturing morphological diversity at the dataset level through a non-parametric prototype generation module to guide key patch selection. By integrating hierarchical attribute text to construct a semantic calibration graph, the framework unifies visual and semantic information within a graph learning architecture to achieve more robust WSI classification.
- 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 constructs dataset-level non-parametric prototypes through a two-stage K-means clustering across WSIs, extending visual diversity modeling from individual slides to the dataset level; this design approach demonstrates significant relevance to the WSI classification task. The paper expands coarse-grained attribute text into fine-grained hierarchical text with three intensity levels, using prototypes as a bridge to calibrate the semantic affinity matrix between patches and text. By explicitly injecting hierarchical semantic information into the graph construction process, this coarse-to-fine multimodal alignment design distinguishes itself from mainstream direct-alignment methods.
- 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 relies on K-means clustering for dataset-level prototype generation, yet it fails to discuss the sensitivity of the prototype count across datasets of varying scales or the criteria for its selection, leaving both reproducibility and generalization in doubt. The paper generates hierarchical attribute text using GPT-5.2 followed by manual review, but it does not report the evaluation metrics or criteria for text quality. Furthermore, the impact of text content on final classification performance lacks targeted ablation, leaving the relationship between text quality and model performance unclear.
- Please rate the clarity and organization of this paper
Poor
- 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.
(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 proposed integration of dataset-level non-parametric prototypes and hierarchical semantic graph learning for WSI classification demonstrates a degree of novelty, and the experimental results show consistent performance improvements over baseline methods across three datasets. However, the key design choices of the core modules lack sufficient ablation validation, the impact of attribute text quality on performance remains unanalyzed, and the evidence supporting the claims of generalization is limited.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
Although the rebuttal clarifies several implementation details, the core concerns regarding the insufficient validation of key design choices, the limited analysis of text-quality dependency, and the overall methodological significance remain only partially addressed. In particular, several important experimental results and robustness analyses were introduced only during rebuttal rather than being adequately supported in the original manuscript. Therefore, I maintain my rejection recommendation.
Review #2
- Please describe the contribution of the paper
The paper proposes ProtoSG-Net, a prototype-guided semantic graph learning framework for whole slide image (WSI) classification that aims to better capture dataset-level visual diversity and improve multimodal alignment between visual features and pathological texts. The main contribution lies in introducing a non-parametric prototype generation mechanism that aggregates intra- and inter-slide information to model representative tissue patterns across the dataset, which is then used to guide key patch selection. In addition, the method designs a prototype-guided semantic calibration and graph learning module that integrates coarse- and fine-grained attribute texts with visual features, enabling more effective modeling of inter-patch relationships beyond spatial adjacency. By combining prototype learning with semantic graph reasoning, the framework enhances visual–semantic alignment and improves classification performance across multiple WSI datasets .
- 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 use of non-parametric prototypes: The paper introduces dataset-level non-parametric prototypes that capture global visual diversity across WSIs, which is more expressive than conventional instance-level patch selection and helps improve generalization. Effective integration of multimodal information: The method combines visual features with both coarse- and fine-grained pathological attribute texts, and further models their hierarchical relationships, leading to better visual–semantic alignment than standard fusion approaches. Prototype-guided graph learning design: The proposed semantic calibration and graph learning module leverages prototypes to refine inter-patch relationships beyond simple spatial adjacency, which is a meaningful extension of graph-based WSI modeling.
- 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.
Reliance on generated textual descriptions: The attribute texts are generated (e.g., via GPT-based tools) and manually refined, which may introduce bias or inconsistency. The robustness of the method to different or noisy textual descriptions is not thoroughly analyzed. Limited clinical validation: Although the method is evaluated on multiple datasets, including a private one, the evaluation focuses on classification metrics without demonstrating clinical utility (e.g., impact on diagnosis, interpretability for pathologists). Evaluation limitations: The experiments are based on fixed data splits and lack more robust validation strategies such as cross-validation. Additionally, comparisons with more recent large-scale multimodal or foundation models are limited, making it harder to fully assess the competitiveness of the approach.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The 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?
The paper presents a technically sound and well-motivated framework for WSI classification, with a reasonable degree of novelty in how it combines non-parametric prototype learning, hierarchical text-guided semantic calibration, and graph-based patch relation modeling. The method is coherent, and the experimental results on both private and public datasets are consistently strong. Although some components are related to prior prototype-based, graph-based, and multimodal WSI methods, the overall integration is meaningful and appears effective in practice.
- 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.
N/A
Review #3
- Please describe the contribution of the paper
The paper proposes ProtoSG-Net for Whole Slide Image (WSI) classification. The core contributions are:
An attribute-aware, prototype-guided patch selection module that utilizes non-parametric dataset-level prototypes generated from WSIs and coarse-grained attribute texts. A semantic-aware, prototype-guided graph learning module that models the complex interactions between fine-grained attributes and key patches, which are then calibrated using the aforementioned prototypes to improve classification performance.
- 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 proposed integration of visual prototypes with textual attributes (both coarse- and fine-grained) to construct a semantic graph is an interesting and well-motivated approach. It attempts to explicitly model dataset-level visual diversity and align it with hierarchical pathological descriptions, which effectively addresses the limitation of isolated patch-level processing in current WSI analysis. The hierarchical progression from patch selection to semantic graph construction forms a logically sound pipeline for vision-language fusion in pathology.
- 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.
Despite the interesting design, there are critical ambiguities regarding the inference pipeline and the fairness of experimental comparisons that must be addressed:
Unclear Testing/Inference Pipeline: The testing pipeline is not clearly described in the manuscript. It is difficult to understand the exact end-to-end inference procedure when a completely new, unseen WSI is fed into the network. Potential Label Leakage: It is unclear whether (and how) textual inputs are utilized during the testing phase. If the corresponding textual prompts are generated using ground-truth class names (or descriptions derived directly from the specific diagnosis) and fed into the model for inference, this would introduce severe label leakage. The authors must clarify the exact source and generation process of textual prompts during the testing phase to rule out this possibility. Unfair Baseline Comparison regarding LLM Backbones: In the experimental comparisons, the baseline method Ref [5] utilizes an LLM (GPT-4o mini), whereas this paper reportedly employs a different/upgraded LLM (“GPT-5.2”). This discrepancy raises a critical question: Could the performance gap between the proposed method and Ref [5] be primarily attributed to the superiority of the underlying LLM rather than the architectural design of ProtoSG-Net? To ensure a fair comparison, the authors should provide experimental evidence where the LLM used in Ref [5] is replaced with the exact same LLM used in their method.
- 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 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
The methodology is intriguing, but the clinical utility of Vision-Language Models relies heavily on a strictly blinded inference process. Please explicitly detail the inference algorithm in the rebuttal and the revised manuscript. Additionally, please double-check the version name of the LLM used in your study (e.g., “GPT-5.2”), as this seems unusual.
- 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?
I recommend a Weak Accept. The paper presents a novel and methodologically sound approach to fusing visual WSI features with hierarchical textual pathological attributes via prototype-guided graph learning. The motivation is solid. However, my enthusiasm is currently tempered by critical missing details regarding the inference phase (specifically, the risk of label leakage via text prompts) and concerns about fair comparisons with baselines due to the mismatch in LLM backbones. If the authors can clarify the inference pipeline to eliminate concerns of label leakage and provide a strictly fair baseline comparison in the rebuttal, this paper would be a solid contribution to the MICCAI community.
- 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 thank all reviewers and the AC for constructive feedback. We will address the major concerns one by one.
1）@R1: ablation The sensitivity of the dataset-level prototypes have been studied in Section 3.3 and Fig. 4 (c). The optimal number of prototypes is 15 for 3 datasets, and the results fluctuate only slightly around this value, indicating low sensitivity to this hyperparameter. For reproducibility, dataset-level prototypes are obtained via standard K-means clustering, and all hyperparameters and experimental details are reported in Section 3.1, enabling straightforward replication.
2）@R1@R2: attribute text quality All attribute texts were reviewed by professional pathologists to check their quality without any modifications, so very few bias or inconsistency is introduced. Coarse-grained texts describe high-level pathological attributes (e.g., the overall differentiation and maturity of cells within the tissue) without defining“correct”values, and fine-grained texts simply expand these attributes into three intensity levels. We additionally conduct robustness experiments by adding Gaussian noise (mean=0, std=0.01, 0.1, 0.5) to textual features. The results show that model performance remains stable (e.g., Acc and F1 vary within 2% on all datasets), indicating insensitivity to minor variations in text content. For instance, on TCGA-NSCLC, acc stays at 96.65% across different noise levels.
3）@R3: inference pipeline and label leakage clarification During testing, the dataset-level prototypes are not regenerated. Instead, the prototypes generated from the training set slides are directly used for patch selection, ensuring a fully end-to-end inference for a completely unseen case without any access to its label. Regarding textual inputs, all attribute texts describe visual properties of the tissue (e.g., the overall differentiation and maturity of cells within the tissue) and are initially coarse-grained, then expanded into fine-grained intensity levels (high, moderate, low), all of which are used as input. These texts contain no information derived from ground-truth class labels or patient diagnosis, eliminating any possibility of label leakage.
4）@R2: clinical validation The clinical relevance of the model is supported by attention visualization in Fig. 3 and Section 3.2, where the patches with the highest attention scores correspond to pathologically meaningful regions. These visualizations highlight that the model focuses on lesion areas, demonstrating interpretability and alignment with clinical expectations.
5）@R2: evaluation We conduct 5-fold cross-validation on Camelyon+ comparing our model with TransMIL (previous second-best). Our model outperforms it by 0.42% in Acc, 1.93% in F1, 0.11% in AUROC, and 3.54% in PRAUC. We also conduct zero-shot classification on two large-scale multimodal foundation models: PLIP (Acc: 56.74%, AUROC: 46.26%) and CONCH (Acc: 59.01%, AUROC: 57.41%) on Camelyon+. These additional results suggest that our method remains competitive under cross-validation and compares favorably with recent foundation models.
6）@R3: fairness check for LLM backbone EmmPD’s metrics remain lower than ours across all three datasets with GPT-5.2-generated text (acc on TCGA-NSCLC: 94.74%; Placenta: 75.27%; Camelyon+: 87.59%). Besides, the LLM used in our study is indeed GPT-5.2, which was officially released in December 2025 and the latest available version at the time of our submission.
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.
The paper is potentially acceptable. The methodological direction is relevant and moderately novel for MICCAI, especially in combining dataset-level prototypes with hierarchical semantic graph learning for WSI classification. The reported performance appears promising across multiple datasets. However, the current concerns are substantial and directly affect validity: possible label leakage during inference, unclear use and generation of text prompts at test time, unfair baseline comparison due to different LLM backbones, insufficient ablation of prototype/text/graph components, and weak reproducibility details.
- 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.
Despite relying on some previously established components, the paper’s integration strategy provides a distinct, valuable, and proven contribution to multimodal and graph-based WSI classification. The robust experimental results firmly support its utility, making it a strong addition to the conference/journal.
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 ProtoSG-Net, a prototype-guided semantic graph learning framework for whole slide image (WSI) classification that aims to better capture dataset-level visual diversity and improve multimodal alignment between visual features and pathological texts. Overall, the paper is interesting, technically sound, and addresses a relevant problem in computational pathology.
After carefully considering the manuscript, the reviews, and the rebuttal, I find that the major concerns raised during the review process have been adequately addressed. In particular, the authors provided satisfactory clarifications regarding the use of GPT-5.2, the fairness of the corresponding comparisons, the quality of the generated attribute texts, and the inference procedure. These responses significantly improve confidence in both the methodology and the reported results. Therefore, I support acceptance of this paper at MICCAI.
For the final version, I encourage the authors to incorporate the key clarifications provided in the rebuttal into the manuscript. In addition, Figure 2 could be further improved for readability. Specifically, the arrows and symbols in the green-highlighted region would benefit from more explicit explanations. The current depiction of the matrix multiplication operation is also somewhat confusing, particularly where two inputs are first merged by arrows before entering the matrix multiplication symbol. A clearer visual representation of the data flow and operations would substantially improve the accessibility of the method.
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 sufficiently clarifies the inference pipeline, training-set prototype usage, and absence of label-derived textual prompts, which alleviates the most serious concern regarding potential label leakage. Although some concerns remain regarding text-quality dependence, robustness of the generated semantic attributes, and the reliance on additional rebuttal-only analyses, the prototype-guided semantic graph formulation is reasonably motivated and the post-rebuttal evidence places the paper marginally above the acceptance threshold.