List of Papers Browse by Subject Areas Author List
Abstract
Recent methods for pathology report generation from whole-slide images (WSIs) are capable of producing slide-level diagnostic descriptions but fail to ground fine-grained statements in localized visual evidence. Furthermore, they lack control over which diagnostic details to include and how to verify them. Inspired by emerging agentic analysis paradigms and the diagnostic workflow of pathologists, who selectively examine multiple fields of view, we propose QCAgent, an agentic framework for quality-controllable WSI report generation. The core innovations of this framework are as follows: (i) it incorporates a customized critique mechanism guided by a user-defined checklist specifying required diagnostic details and constraints; (ii) it re-identifies informative regions in the WSI based on the critique feedback and text-patch semantic retrieval, a process that iteratively enriches and reconciles the report. Experiments demonstrate that by making report requirements explicitly prompt-defined, constraint-aware, and verifiable through evidence-grounded refinement, QCAgent enables controllable generation of clinically meaningful and high-coverage pathology reports from WSI. Code is available at https://github.com/cs-w0r1d/qcagent.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1286_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/cs-w0r1d/qcagent
Link to the Dataset(s)
N/A
BibTex
@InProceedings{WanRun_QCAgent_MICCAI2026,
author = { Wang, Rundong AND Ba, Wei AND Li, Yingtai AND Zhou, Ying AND Liu, Bowen AND Wang, Baizhi AND Wang, Yuhao AND Yang, Zhidong AND Zhang, Kun AND Yan, Rui AND Zhou, Shaohua Kevin},
title = { { QCAgent: An Agentic Framework for Quality-Controllable Pathology Report Generation from Whole Slide Image } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16878},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
- Iterative QC architecture. We reformulate report generation as a multiround audit–retrieve–revise process. A structured QCAgent automatically identifies non-compliant statements and missing clinical information, enabling closed-loop refinement.
- Demand-driven evidence retrieval. We introduce a text-guided evidence retrieval mechanism. Instead of random sampling, the QCAgent converts missing fields into morphology-oriented queries to retrieve the most relevant visual patches from the WSI, grounding the report completion in targeted evidence.
- Strict anti-hallucination design. We adopt an evidence-priority scheme (supplemental visual evidence > initial draft) to reduce hallucinations. By explicitly separating verifiable from non-verifiable fields, only evidence-supported claims are retained in the final report.
- 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.
a demonstration of clinical feasibility and a novel application.
- 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.
There are too few comparison methods; it is recommended to supplement them.
- 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.
(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 presents QCAgent, an agentic and quality-controllable framework for pathology report generation from whole slide images. The work addresses important limitations of existing WSI-to-text methods, including insufficient fine-grained visual grounding, limited control over report contents, and potential hallucination in generated diagnostic descriptions. Instead of relying on a single-pass generation paradigm, the proposed method formulates report generation as an iterative audit–retrieve–revise process. It identifies missing or insufficiently supported diagnostic fields through a quality-control module, converts them into morphology-oriented retrieval queries, retrieves relevant local patches from the WSI, and revises the report based on supplemental visual evidence. This design is well aligned with the diagnostic workflow of pathologists, who iteratively examine slides at different magnifications.
The main contributions are clear. First, the paper introduces a QC-driven iterative framework that makes pathology report generation more controllable and auditable. Second, the demand-driven evidence retrieval mechanism allows the system to actively locate relevant WSI regions according to missing diagnostic fields. Third, the explicit rules such as NO-FABRICATION, EVIDENCE PRIORITY, and FIELD CLASSIFICATION provide useful constraints for reducing unsupported claims and hallucinations. Fourth, experiments on both TCGA-STAD and an in-house Chinese gastric cancer cohort demonstrate consistent improvements over the PRISM baseline, especially in clinical field coverage and semantic alignment.
To further strengthen the paper, I suggest adding one key experiment: a small-scale expert evaluation by pathologists. Since the paper emphasizes clinically meaningful, quality-controllable, and evidence-grounded report generation, automatic metrics such as BLEU, ROUGE, METEOR, BERTScore, and Field Recall may not fully capture medical correctness and clinical usability. The authors could randomly select 30–50 cases and ask 2–3 pathologists to blindly compare reports generated by PRISM and QCAgent in terms of diagnostic correctness, clinical completeness, factual consistency, evidence support, and hallucination risk. This additional evaluation would substantially improve the clinical credibility of the work without requiring a large experimental expansion.
I also suggest that the authors clarify the definition and computation of Field Recall, including the selected clinical fields and matching criteria. In addition, the ablation result where “w/o Patho-R1” achieves a higher FieldRecall than the full model should be better explained, as it appears inconsistent with the current textual interpretation. Overall, the paper addresses an important problem, proposes a reasonable and clinically motivated framework, and provides promising experimental results. These suggestions are intended to further improve the presentation and clinical validation, and do not affect my recommendation 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.
N/A
Review #2
- Please describe the contribution of the paper
The paper presents an interesting direction by proposing QCAgent, an agentic framework that iteratively improves the quality of pathology reports through a QC-round refinement process. The framework leverages feedback-guided CONCH retrieval and Patho-R1 evidence to iteratively refine reports until predefined quality control (QC) criteria are satisfied.
- 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.Agentic formulation for report generation. The paper reformulates WSI report generation as an iterative quality-control loop (audit–retrieve–revise), enabling on-demand backtracking to refine missing or uncertain findings. This allows more controllable and evidence-grounded report generation instead of relying on a single-pass prediction.
2.Performance improvement over PRISM. The experiments show consistent improvements over PRISM (used as the base model for making WSI Report Generation in QCAgent) across several WSI report generation metrics.
3.Clinically intuitive design. The iterative refinement process resembles how pathologists review slides by revisiting regions at different fields of view when information is missing, making the framework conceptually aligned with real diagnostic workflows.
- 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.Limited novelty of the iterative feedback mechanism: Although using LLM feedback to guide patch selection in subsequent iterations is relatively new for report generation, similar ideas have already been explored in computational pathology agents. PathAgent [1], for example, applies a self-reflection mechanism to identify missing information and uses this feedback to guide patch retrieval in the next iteration.
[1] Chen, J., Cai, L., Wang, Z., Huang, Y., Jiang, S., Huang, S., … & Zhang, Y. (2025). Pathagent: Toward interpretable analysis of whole-slide pathology images via large language model-based agentic reasoning. arXiv preprint arXiv:2511.17052. 2.Comparison limited to a single baseline: The experiments mainly compare against PRISM, which is also used as the base model for report generation in QCAgent. The paper does not include comparisons with other recent models such as SlideChat (CVPR 2025) [1] and WSI-LLaVA (ICCV 2025) [2], making it difficult to assess the relative advantage of the proposed framework.
[1] Chen, Y., Wang, G., Ji, Y., Li, Y., Ye, J., Li, T., … & He, J. (2025). Slidechat: A large vision-language assistant for whole-slide pathology image understanding. In Proceedings of the Computer Vision and Pattern Recognition Conference (pp. 5134-5143).
[2] Liang, Y., Lyu, X., Chen, W., Ding, M., Zhang, J., He, X., … & Shen, L. (2025). Wsi-llava: A multimodal large language model for whole slide image. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 22718-22727).
3.Dependence on large proprietary models: QCAgent relies on Qwen3-VL-30B-A3B-Thinking, which may be computationally expensive and difficult to deploy in low-resource environments.
- 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?
The paper proposes an agentic formulation for WSI report generation, where reports are iteratively refined through a quality control loop consisting of audit, retrieve, and revise steps. Using feedback to guide evidence retrieval is intuitive and clinically meaningful, as it reflects how pathologists re examine slides to validate findings.
The overall novelty is somewhat limited, since similar iterative self reflection mechanisms have been explored in prior work such as PathAgent. However, the paper still provides a solid and well motivated extension in the context of report generation. The experimental validation could be more comprehensive, as comparisons are primarily limited to PRISM and do not include other recent approaches such as SlideChat or WSI-LLaVA. The reliance on large proprietary models such as Qwen3-VL-30B-A3B-Thinking also raises concerns about computational cost and practical deployment.
Overall, despite these limitations, the paper addresses an important problem and presents a reasonable methodological contribution that meets the standard for MICCAI. Given the practical importance of WSI report generation and the soundness of the approach, I lean toward a weak accept.
- Reviewer confidence
Very confident (4)
- [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.
Given the unique motivation of the work, I am inclined to accept the paper, as most existing agentic frameworks for WSI analysis primarily focus on WSI-VQA, whereas this work instead emphasizes improving the quality of WSI report generation. However, I would still recommend adding more baselines, along with additional analyses and ablation studies (including experiments with different backbones such as CONCH, Patho-R1, and Qwen3-VL-30B-A3B-Thinking), to further strengthen the paper.
Author Feedback
We thank the reviewers and AC for their positive comments and for recognizing QCAgent’s clinical motivation (R1), pathologist-like iterative workflow (R2), improvements over PRISM (R1/R2)，and the transparency of the audit-retrieve-revise design(R1/R2).
Comparison methods (R1/R2/AC): QCAgent should be viewed as a model-agnostic QC layer on top of a WSI-to-report generator, not a single fixed generator. The clean way to isolate its contribution is a paired comparison on the same backbone: the same PRISM draft is compared with the QC-refined report. Tab.1 measures the added value, with Field Recall improving from 0.3208 to 0.6335 on TCGA-STAD and from 0.1908 to 0.4863 on the in-house cohort. SlideChat and WSI-LLaVA primarily target VQA/assistant-style interaction, not full report generation; direct conversion would mix task adaptation with method quality. We appreciate this suggestion and will make the QC-layer framing clearer.
Novelty compared with PathAgent (R2/AC): We appreciate the reference to PathAgent. Both use iterative WSI analysis but differ in task and control. PathAgent is a training-free VQA/WSI-analysis agent that decides whether evidence suffices to answer a specific question. QCAgent targets complete pathology report generation, covering fields such as tumor type, grade, invasion, LVI/PNI, margins, lymph nodes, and staging. In Sec. 2, quality-controllability is controlled by criteria K and budget T, specifying schema, required fields, anti-hallucination rules, evidence priority, and template. Each round outputs structured QC results: missing fields M_t, morphology queries Q_t, and a revised draft. It also separates image_related from admin_required fields and enforces NO-FABRICATION/EVIDENCE PRIORITY. Thus, the novelty is not a generic loop, but a QC-constrained, auditable audit-retrieve-revise formulation for report generation.
Computational cost and deployability (R2/AC): We agree this should be clearer. QCAgent is a lightweight QC layer, not a monolithic system. It can be combined with any WSI-to-report model that provides an initial draft, and the QC agent can be replaced by a stronger or lighter LLM depending on budget. In our implementation, Qwen3-VL-30B-A3B-Thinking is only one plug-in QC agent, and its MoE activates only a subset of parameters. Cost is bounded by T=3 QC rounds and K=3 retrieved patches per query. CONCH retrieval is FAISS Top-K over pre-extracted embeddings, so extra cost mainly comes from limited patch descriptions and revisions for missing/uncertain fields. K and T are quality-cost controls; reducing them gives a direct low-resource setting.
Metrics and Field Recall (R1/AC): NLG metrics alone are insufficient for clinical reports. Field Recall measures whether ground-truth clinical fields are covered, including tumor type, grade/differentiation, invasion depth, LVI, PNI, margins, lymph-node status, and TNM/stage-related information. It is a coverage metric, while BERTScore captures semantic alignment. Together with recall-centric ablations and qualitative evidence traces, these metrics reduce dependence on n-gram overlap alone. We will fully incorporate these suggestions and evaluate additional report generation baselines in our planned journal extension.
w/o Patho-R1 Field Recall observation (R1): The higher Field Recall of w/o Patho-R1 does not imply better reliability. Without visual evidence, the LLM can fill field names from priors, e.g., “LVI: not identified”, increasing coverage without grounding. The full model is constrained by retrieved evidence and anti-hallucination rules, so unsupported items are marked undetermined rather than fabricated. This yields slightly lower Field Recall but higher evidence-supported coverage, reflected by higher ROUGE-L Recall (0.5139 vs. 0.4495).
We hope these clarifications address the main concerns and thank the reviewers for helping us sharpen the framing of the work.
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 two reviewers acknowledge the clinical motivation of the formulation, but raise concerns that warrant clarification in a rebuttal. The authors should focus on the following points:
1) Insufficient experimental comparison (raised by both reviewers). The evaluation is largely limited to PRISM, which is also used as the base model inside QCAgent, making it hard to assess the standalone advantage of the proposed framework. 2) Novelty relative to prior agentic pathology work (R2). The iterative feedback-and-retrieval mechanism is conceptually close to PathAgent (Chen et al., 2025), which already employs a self-reflection mechanism to identify missing information and guide patch retrieval. The authors should clearly articulate the technical differences between QCAgent and PathAgent (and other agentic pathology systems), and quantify the contribution that is unique to the QC formulation. 3) Computational cost and deployability (R2). QCAgent depends on Qwen3-VL-30B-A3B-Thinking and multiple heavy components (CONCH, Patho-R1). The authors should discuss inference cost, the number of QC rounds in practice, and feasibility of deployment in realistic clinical / low-resource settings. 4) Evaluation relies almost entirely on traditional NLG metrics (BLEU-1/4, ROUGE-L, METEOR, FieldRecall, BERTScore). These are known to correlate poorly with clinical correctness for long-form pathology reports, and they are increasingly inadequate for evaluating LLM-era generation.
- 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 originally had three reviewers with scores of 2 / 3 / 5.The reviewer who gave a score of 2 has since been removed from the system, leaving only two active reviewers. The negative end of the score distribution has effectively disappeared without a written record, which makes the remaining set of reviews skewed toward acceptance. Late score change. The reviewer originally scoring 3 raised the score to 4 close to the deadline. The change is within the reviewer’s discretion, but combined with point (1) it noticeably shifts the apparent consensus. Quality of Reviewer #1’s review. Reviewer #1 recommends a strong accept (5, “Accept — should be accepted, independent of rebuttal”) but the review itself is very thin: the “main contribution” field appears to be copy-pasted from the paper’s own claims; the weaknesses field is a single sentence (“There are too few comparison methods”); and the justification is “Research hotspots and novel methods.” I do not consider this review well-substantiated, and I would weight it accordingly when making the final decision.
After the rebuttal, both remaining reviewers recommended acceptance, and the previously perfunctory review was rewritten more carefully. Given that the paper is genuinely borderline, I concur with the reviewers and accept the paper, but request that the SAC review the case carefully. Thank you.
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 QCAgent, a WSI report generation framework built upon an audit–retrieve–revise pipeline with multiple large language and vision-language models. The proposed framework is well-motivated and demonstrates meaningful improvements in the quality of medical report generation. Although the adoption of VLMs, such as Patho-R1 and Qwen3-VL-30B-A3B-Thinking, introduces substantial computational overhead, the resulting performance gains and report quality improvements are worthwhile and valuable for the medical imaging community. Overall, the paper presents a practically relevant and technically sound system with promising clinical implications. Therefore, I support acceptance of this paper at MICCAI.