List of Papers Browse by Subject Areas Author List
Abstract
Automated pathology report generation from Whole Slide Images (WSIs) has attracted increasing attention in digital pathology. However, existing methods are predominantly developed under single-organ settings, overlooking the multi-organ scenarios encountered in clinical practice, where organ types typically follow a long-tailed distribution. To address this gap, we identify two critical biases: (1) visual representation bias, where the encoder favors head-class patterns over tail-class discriminative features, and (2) textual decoding bias, where the decoder overfits to head-class narrative patterns, yielding diagnostically unreliable outputs for tail-class organs. To mitigate these two biases, we propose a novel Prior-anchored multi-Organ pathology report Generation framework (PriOrGen). Specifically, a Visual-Prototype Anchored Bottleneck module leverages the information bottleneck principle with learnable anchor representations to selectively retain diagnostically relevant visual information while filtering out head-biased redundancy. Secondly, a Meta-Report Anchored Bank module constructs an organ-specific meta-report anchored bank and retrieves organ-faithful textual priors to steer the decoder away from head-class narrative patterns. Extensive experiments on a multi-organ pathology dataset demonstrate that our method effectively mitigates long-tail biases and achieves superior report generation performance across both head and tail organ categories compared to state-of-the-art methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2990_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{YanFen_PriorAnchored_MICCAI2026,
author = { Yang, Feng AND Liu, Jie AND Pang, Yubo AND Chen, Peilin AND Lyu, Xinheng AND Wang, Shiqi AND Leung, Howard AND Chen, Ping},
title = { { Prior-Anchored Debiasing for Long-Tailed Multi-Organ Pathology Report Generation } },
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 proposed methodology, PriOrGen, investigates a prototype-based approach for generating reports from Whole Slide Images (WSIs). The authors claim that by utilizing a meta-report bank and a query mechanism, they have effectively improved the generation performance for tail classes, addressing the challenges inherent in long-tail distributions.
- 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 method is supported by a rigorous evaluation across 11 different cancer types.
- The paper includes a thoroughly designed ablation study that clearly isolates and justifies the contribution of each individual component.
- 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 necessity of the “meta-report” concept is questionable. Since pathology reports often follow standardized formats, it is unclear if this meta-report is a simple template or a structured representation of specific findings (e. g. , organ-specific mentions, staging). If it aims for the latter, an ontology-based graph approach might be more effective.
-
While prototypes are efficient, they risk losing fine-grained details. Using only 4 prototypes seems insufficient for pathology, where micro-scale patterns (e. g. , Lymphovascular Invasion, LVI) are critical for diagnosis. Compared to methods like PANTHER (using 16 or 32 prototypes), the proposed approach may only capture macro-patterns, which is a significant concern for clinical utility.
-
It is unclear whether the prototypes are organ-specific or universal. If the model relies on organ-specific learning, the claim of addressing “long-tail distributions” across a universal scale may be exaggerated. The authors should clarify the scope and implementation of these prototypes.
-
The authors used a mix of UNI and PLIP. For better consistency and to isolate the contribution of the proposed method, using a single backbone (e. g. , PLIP) throughout the experiments would have been more appropriate.
-
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Although the methodology is well-executed and the results are promising, the paper tends to over-interpret the significance of certain components. Some of the high-level claims lack direct empirical support within the provided results.
- 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 reading the authors’ rebuttal, I find that the authors have responded thoroughly and in good faith, resolving many of the concerns I raised. The clarifications provided are convincing, and I expect that the promised revisions will lead to a meaningfully improved final manuscript. I therefore maintain my original decision to Accept.
Review #2
- Please describe the contribution of the paper
The paper targets pathology report generation in a multi-organ setting where organ types follow a long-tailed distribution, a scenario that has received little attention compared with single-organ report generation. The authors identify two biases induced by the long tail, visual representation bias on the encoder side and textual decoding bias on the decoder side, and propose PriOrGen, a dual-prior framework that addresses both. Experiments covering 11 TCGA cancer types show gains over several baselines, particularly on tail-class organs.
- 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 clear and clinically meaningful. Moving from single-organ to multi-organ pathology report generation is a realistic direction, and the observation that a long-tailed organ distribution induces both visual and textual biases is a reasonable framing that has not been explicitly studied before in this task.
2.The decomposition of the problem into two complementary bias sources is clean, and the proposed dual-prior design is conceptually aligned with these two sources.
3.The paper is clearly written and the figures communicate the proposed pipeline effectively.
- 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.Evaluation in the experiments relies only on BLEU-Mean, METEOR and ROUGE-L. However, these surface n-gram metrics have been shown to correlate poorly with clinical correctness in medical report generation. Since the central claim of the paper is improved diagnostic reliability on tail-class organs, the absence of any clinical or entity-level metric, and the absence of pathologist evaluation, makes it hard to verify that the method actually mitigates the “linguistically fluent but diagnostically erroneous” failure mode that it explicitly targets.
2.The reported improvements are small and their statistical reliability is not well established. Tail-class absolute gains are also modest and come from very small test sets. Results are reported from a single run without standard deviations, and the significance test uses n=11 organs as independent units rather than repeated runs, which provides limited statistical power. Under this protocol the improvements are not convincingly above noise.
3.No comparison against long-tail learning methods is included.
4.The capacity of VPAB appears insufficient and is not analyzed. Only K=4 prototypes are used to cover diagnostic patterns across 11 histologically heterogeneous organ types, and no sensitivity analysis on K is reported.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a clinically relevant and under-explored problem, and the bias decomposition together with the corresponding dual-prior design is well motivated. However, the empirical support for the main claims is weak in its current form.
- 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.
Although my concerns have not been fully addressed, given that the overall completeness of this paper is relatively high and the author has made efforts to supplement necessary experimental results in the rebuttal, I am inclined to accept it.
Review #3
- Please describe the contribution of the paper
The paper proposed the PriOrGen, a novel pathological image report generation model that addresses the long-tail distribution issue. The proposed method first uses a visual-prototype anchored bottleneck module to filter redundant pathological patches and reduce the influence of long-tailed visual distribution. It then uses a Meta-Report Anchored Bank module to retrieve organ-level meta-report templates. The organ-level meta report ensures the model is not biased to the text distribution of the major organ classes. PriOrGen is evaluated on the ML-Path dataset with 11 different cancer types and an extremely biased distribution. The proposed method shows a uniform improvement in both head and tail classes and demonstrates balanced behavior across different organs and cancer types.
- 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 proposed method is reasonable and well-validated. The PriOrGen addresses the distribution bias issue from both visual and text aspects. VPAB successfully compensates for scarce tail-class visual representations, while MRAB independently guards against textual overfitting during the decoding phase.
2.The ablation experiment is well organized and helpful. The paper not only evaluated the contribution of the VPAB and MRAB module, but also demonstrated the influence of hyperparameters like $\rho$ and $\mu$.
3.Code is provided with an anonymous link in the paper.
- 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 method is largely built on existing methods, such as LLMs like Gemini3-Pro and PLIP. While these methods are well-recognized in the field and are generally considered standard, it is still possible that these methods are not reliable in some aspects. A discussion about such influence should be included.
2.While the paper has evaluated the proposed method against multiple baselines, it would still be great if the proposed method were compared with a long-tail distribution-oriented method. Even simply comparing against a vanilla baseline trained with organ-wise balanced sampling can further highlight the contribution of the proposed method.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The reviewer is curious about the model size and training cost (in terms of GPU-hours) for the proposed method. Providing this information can help improve understanding of the model’s behavior.
- 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?
Overall, this is a solid paper with clear writing, detailed evaluation, and non-trivial contribution. The paper’s arguments are properly discussed and validated through the experiments. Thus, the reviewer would like to recommend direct accept for this work.
- 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 rebuttal has largely addressed my concerns. I would encourage the authors to update the paper in the camera-ready version to reflect these new results.
Author Feedback
We thank the meta-reviewer and reviewers for recognizing the clinical importance, clear presentation, comprehensive evaluation, and ablations. We conducted additional analyses for major concerns below:
Meta-reports concept (R1Q1/MetaQ1). As shown in Fig. 2, meta-reports are designed at the organ level. They summarize recurring report templates and common organ-specific expressions in training reports, including diagnostic terminology, lesion descriptions, staging/margin phrases, and report-style context. Thus, they provide the decoder with an organ-faithful textual prior rather than a rigid output template. An ontology graph is complementary, but it mainly encodes predefined concepts and relations; it does not directly capture the empirical language distribution needed to mitigate textual decoding bias.
Prototype number (R1Q2/R1Q3/R2Q4/MetaQ2). VPAB prototypes are learned across organs. VPAB employ these prototypes to guide patch filtering rather than replacing fine-grained evidence. After prototype scoring, 40% of WSI patches are retained for decoding. We will further clarify these. Here, we further conduct K-sensitivity: Prototype_K / BLEU-MEAN / ROUGH-L / METEOR 2: 0.265 / 0.175 / 0.294 4: 0.273 / 0.181 / 0.305 6: 0.258 / 0.173 / 0.293 8: 0.254 / 0.168 / 0.286 K=4 is empirically optimal. Larger K adds redundant anchors rather than improving micro-pattern capture.
Backbone consistency (R1Q4/MetaQ3). UNI and PLIP play different roles. UNI is the main histology visual encoder for VPAB because of its strong pathology representation. PLIP is used only as the vision-language retrieval index in MRAB, matching selected patches to PLIP-encoded meta-reports. We added PLIP-only experiment here: Backbone / BLEU-MEAN / ROUGH-L / METEOR PLIP-only: 0.257 / 0.170 / 0.288
Clinical correctness/statistics (R2Q1/R2Q2/MetaQ4). We further added clinical metrics comparing with second best baseline: Method/ FactENT [1] / DocLens recall [2] / DocLens precision [2] Ours: 0.514 / 0.3626 / 0.4522 BiGen: 0.501 / 0.3519 / 0.4011 These support better diagnostic content beyond surface overlap. For reliability, we conduct report-level tests vs BiGen: B-M p=0.0153 and METEOR p=2.52e-05 are significant; ROUGE-L p=0.1929 is not significant, which is same as Table 1.[1] Miura Y, Zhang Y, Tsai E, et al. Improving factual completeness and consistency of image-to-text radiology report generation[C]//Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021: 5288-5304.[2] Xie Y, Zhang S, Cheng H, et al. Doclens: Multi-aspect fine-grained medical text evaluation[C]//Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024: 649-679. Long-tail benchmarking (R2Q3/R4Q2/MetaQ5). Since other long-tail learning methods are not designed for generation task, we added organ-wise balanced sampling (OWBS) as a standard long-tail baseline. Methods / BLEU-MEAN / ROUGH-L / METEOR OWBS: 0.249 / 0.175 / 0.289 Ours: 0.273 / 0.181 / 0.305 This suggests simple resampling is insufficient, and the dual visual/text priors are needed.
Reliability/cost (R4Q1). Gemini-3 Pro is only used offline to construct the meta-report bank and does not participate in training or inference; potential unreliability is filtered through manual verification. PLIP is a mature pathology vision-language model validated on multiple benchmarks. We will add these experiments/clarifications to strengthen the evidence for PriOrGen’s claims.
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 targets multi-organ report generation, which is quite uniquely position compared to single-organ. Specifically, the authors try to solve report generation for multi-organ settings, where organ type follows a long-tailed distribution. The authors propose PriorGen - a dual-prior framework that addresses biases introduced by long tail (visual and textual biases). The authors have evaluated and benchmark on 11 TCGA cancer types.
All the reviewers have agreed that this is an important problem setting that has been overlooked before and is well motivated clinically. The paper is also clearly written and easy to follow, with a rigorous evaluation across 11 cancer subtypes and detailed ablation studies.
Before making a final decision, certain questions require clarification, and I suggest the authors to lay particular emphasis on the below comments from the reviewers.
1.Meta-report concept (R#1): Whether this is really necessary? Why not do an ontology-guided report?
2.Number and nature of prototypes (R#1, R#2): Only 4 prototypes are required? Does it only capture macro-level structures? Were more prototypes tried or was there a sensitivity study done? And what did these prototypes actually learn? Were they organ-specific?
3.Consistent backbone (R#1): The authors should have used consistent UNI or PLIP throughout the experiments? Any reason why two different backbones were tried?
4.Text evaluation metrics (R#2): BLEU, METEOR don’t provide the best correlation with clinical correctness. Did the authors try any other metrics or qualitative interpretation?
5.Benchmarking with long-tailed (R#2,R#4): Standard long-tailed benchmarking is missing.
- 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 PriorGen which performs report generation for multi-organ settings, where organ type follows a long-tailed distribution.
All the reviewers agreed that this is an important problem setting that has been overlooked before and is well motivated clinically during the first round of review. Certain clarifications and comments were raised regarding the inconsistency of backbones, selection of prototypes, and meta-report concepts.
The authors have clarified during their rebuttal, although albeit by also adding additional experiments, which should not be taken into consideration for decision-making. Nevertheless, the authors have argued convincingly about their choice of backbones, meta-reports, long-tailed benchmarks and other issues that were raised. This is a nice contribution to the field of histopathology analysis.
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.
The paper addresses an important and clinically relevant problem of multi-organ pathology report generation under long-tail distributions. Reviewers appreciated the clear motivation, the well-structured dual-prior design, and the practical formulation of mitigating both visual and textual bias in long-tail settings. The proposed framework is technically sound and supported by comprehensive experiments. Several concerns were initially raised regarding the interpretation of the meta-report representation, the limited number of prototypes, the lack of clinical evaluation metrics, and missing comparisons with long-tail learning methods. However, the rebuttal provided additional clarifications and supplementary experimental evidence that adequately addressed the reviewers’ main concerns. Multiple reviewers noted that the revised explanations and additional analyses substantially improved the paper and supported acceptance. Overall, I support acceptance.
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.
This paper initially received divergent scores from reviewers. Following the rebuttal, however, the reviewers reached a consensus in favor of acceptance. Given the reviewers’ unanimous support for acceptance after the rebuttal, I recommend this paper be accepted.