List of Papers Browse by Subject Areas Author List
Abstract
Automatically generating pathology reports from gigapixel Whole Slide Images (WSIs) is essential for reducing pathologist workload and improving clinical efficiency. However, the gigapixel resolution and semantic complexity of WSIs expose two critical limitations in existing methods: (1) Visual dilution: Predominant approaches treat WSIs as single-granularity patch sequences, causing fine-grained focal details to be overwhelmed by background noise; (2) Retrieval noise: Existing methods rely on raw corpora with non-diagnostic redundancy, and their fusion mechanisms fail to suppress interference from low-similarity evidence, leading to factual hallucinations. To address these issues, we propose DR-Gen, a Dual-Stream Retrieval-Augmented Pathology Report Generation framework. Inspired by pathologists’ workflow, DR-Gen decouples global tissue context from focal cellular details via two visual streams: a global stream for slide-level architecture and a focal stream for diagnostic regions. A Visual Priming Adapter leverages these decoupled features as visual anchors, grounding generation in slide morphology. To safeguard medical validity, we construct a CAP-protocol-guided diagnostic knowledge bank and propose an Adaptive Confidence Gating module that dynamically reweights evidence by retrieval confidence. Experiments on the PathText (BRCA) benchmark show that DR-Gen consistently outperforms existing methods in both report quality and subtype classification, achieving a relative improvement of 10.4% in BLEU-4. Ablation studies further validate each proposed component. Code is publicly available at https://github.com/jasonlightx/DR-Gen.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4899_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/jasonlightx/DR-Gen
Link to the Dataset(s)
N/A
BibTex
@InProceedings{XioLia_Think_MICCAI2026,
author = { Xiong, Liang AND Zhang, Yi AND Song, HaoJie AND Peng, QianYao AND Cui, ShaoGuo},
title = { { Think Global, Look Focal: Dual-Stream Retrieval-Augmented Pathology Report Generation for Whole Slide Images } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16879},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes DR-Gen, a novel Dual-Stream Retrieval-Augmented Pathology Report Generation framework for gigapixel Whole Slide Images (WSIs). To combat visual dilution and semantic sparsity, the authors design a global-focal visual extraction mechanism that mimics the “think global, look focal” workflow of pathologists, anchored by a Visual Priming Adapter (VPA). Furthermore, to address retrieval noise and factual hallucinations common in existing RAG methods, the paper introduces a College of American Pathologists (CAP)-guided knowledge bank and an Adaptive Confidence Gating (ACG) module to dynamically filter and reweight retrieved medical evidence.
- 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 architectural design is highly intuitive and aligns seamlessly with real-world clinical workflows. Decoupling the visual processing into a global stream (for tissue architecture) and a focal stream (for cellular details) is an elegant solution to the unique challenges posed by gigapixel WSIs.
2.The transition from unstructured raw corpora to a structured, CAP-protocol-guided atomic knowledge bank significantly reduces non-diagnostic redundancy. Coupled with the ACG module, which uses temperature-scaled softmax to calibrate confidence, the model effectively filters out irrelevant priors and curbs factual hallucinations.
3.The experimental results are convincing. The model achieves state-of-the-art performance not only on traditional NLP metrics (e. g. , a 10.4% relative improvement in BLEU-4) but also on clinical classification metrics (F1 score of 0.926 for subtype prediction) and factual correctness ($Fact_{ent}$). The ablation studies clearly isolate and validate the individual contributions of the dual-stream architecture, VPA, and ACG components.
- 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 evaluation is exclusively conducted on the PathText (BRCA) benchmark, which focuses solely on breast cancer. Because morphological features and diagnostic complexities vary drastically across different organs and cancer types (e. g. , lung or prostate cancer), the lack of multi-center or multi-organ cross-validation makes it difficult to assess the true generalizability of the framework.
2.The retention ratio ($\kappa$) for the focal stream is fixed at 0.6.Given that the proportion of diagnostically relevant tissue—such as tumor bed size—fluctuates wildly from slide to slide, a fixed ratio might discard critical micro-morphological details in slides with extensive lesions, or conversely, introduce background noise in slides with extremely localized focal points. An adaptive retention mechanism would make the methodology more robust.
3.The reliance on the CAP protocol to construct the diagnostic memory bank is effective here, but it poses scalability challenges. The authors do not adequately discuss how this approach would translate to rare diseases or pathologies that lack standardized, well-documented reporting templates, which could result in high manual construction costs in broader clinical applications.
- 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 authors have identified a critical bottleneck in computational pathology—the paradigm mismatch when applying standard radiological vision-language models to semantically sparse, gigapixel WSIs. The proposed DR-Gen framework addresses this with a logically sound, clinically inspired dual-stream design and a robust confidence-gated retrieval mechanism. The methodology is sound, and the experimental results demonstrate meaningful improvements in both linguistic fluency and clinical accuracy. The score is held back from being higher primarily due to the limitation of validating the model on only a single, organ-specific dataset, and the lack of provided source code for immediate peer verification.
- 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 presents two main contributions. First, it proposes a dual-stream visual modeling framework that integrates global and focal representations to address the visual dilution problem in gigapixel Whole Slide Images (WSIs). By explicitly decoupling slide-level tissue structure from fine-grained diagnostic regions, the approach enables more effective multi-scale modeling and reduces the risk of diagnostically relevant cues being overwhelmed by background noise. Second, it introduces a structured knowledge-enhanced retrieval mechanism together with a confidence-aware fusion module. A CAP-guided atomic diagnostic knowledge bank is constructed to provide higher-quality medical priors, and an Adaptive Confidence Gating (ACG) module is designed to dynamically reweight retrieved evidence based on confidence scores, thereby mitigating retrieval noise and improving the factual consistency of the generated reports.
- 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 motivation is well-grounded and clearly articulated. Given the gigapixel scale and extreme semantic sparsity of Whole Slide Images (WSIs), the need to jointly model global tissue structure and focal diagnostic regions is both intuitive and clinically meaningful. The proposed dual-stream design aligns well with the real diagnostic workflow of pathologists (i.e., “global overview followed by focal inspection”), making the formulation both principled and practically relevant. 2.The experimental evaluation is thorough and convincing. The paper includes comprehensive ablation studies that systematically validate the contribution of each component, including the dual-stream decomposition, VPA, and ACG modules. In addition, the authors provide hyperparameter analyses that offer further insight into the robustness and sensitivity of the method, strengthening the empirical credibility of the approach. 3.the paper is well-written and clearly structured. The technical design is presented in a logical and coherent manner, and the connection between motivation, method, and experimental validation is easy to follow. This clarity helps readers understand both the intuition and the implementation details of the proposed framework.
- 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 proposed framework is relatively complex, involving multiple components such as dual-stream modeling, VPA, and ACG. While each module is justified, the overall design introduces additional implementation and computational overhead, and it is unclear whether the performance gains warrant this level of complexity. 2.The experimental evaluation is limited to a single dataset (PathText-BRCA). This raises concerns about the generalizability of the method to other pathology domains, where data distributions and reporting styles may differ significantly.
- 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?
Overall, this paper presents a well-motivated and thoughtfully designed approach to pathology report generation. The proposed dual-stream framework aligns well with the characteristics of gigapixel WSIs, and the integration of a confidence-aware retrieval mechanism is both intuitive and practically meaningful. The experimental evaluation is reasonably thorough, with ablation studies and hyperparameter analyses that help validate the effectiveness of the individual components. The paper is also clearly written and easy to follow. That said, there are still some limitations. The overall framework is relatively complex, involving multiple interacting modules, and it is not entirely clear whether the performance gains fully justify this level of design complexity. In addition, the evaluation is limited to a single dataset, leaving questions about generalizability to other pathology domains. Despite these concerns, the paper provides a solid contribution with a reasonable level of novelty and empirical support, and therefore is slightly 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.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The paper proposes DR-Gen, a dual-stream retrieval-augmented framework for pathology report generation. The core innovation lies in decoupling visual processing into global (tissue architecture) and focal (diagnostic cellular details) streams, combined with a confidence-gated retrieval mechanism to suppress non-diagnostic noise.
- 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.Clear motivation The identification of Visual dilution and Retrieval noise is clear and convincing. 2.Dual-Stream mechanism To deal with the visual dillution, this paper proposed a dual-stream mechanism, emulating the ‘think global, look focal’ clinial practice of pathologists. 3.Strong results The Factent metric (0.519) and high classification F1 indicate that the model preserves factual medical entities, which is critical for clinical utility. The qualitative example (Fig. 2) effectively illustrates the reduction of hallucinations compared to baseline.
- 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 CAP-protocol-guided knowledge bank is mentioned as constructed from “atomic diagnostic statements,” but the paper lacks details about the bank construction.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Clear and strong motivation. The dual-stream mechanism significantly improves the model’s performance.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
All three reviewers agree on Weak Accept. The paper addresses a well-motivated problem with a clinically inspired dual-stream design and demonstrates solid empirical gains. No major consensus concerns.
