List of Papers Browse by Subject Areas Author List
Abstract
Rapid progress in vision-language modeling has enabled pathology report generation from gigapixel whole-slide images, but most approaches assume static training with simultaneous access to all data. In clinical deployment, however, new organs, institutions, and reporting conventions emerge over time, and sequential fine-tuning can cause catastrophic forgetting. We introduce an exemplar-free continual learning framework for WSI-to-report generation that avoids storing raw slides or patch exemplars. The core idea is a compact domain footprint built in a frozen patch-embedding space: a small codebook of representative morphology tokens together with slide-level co-occurrence summaries and lightweight patch-count priors. These footprints support generative replay by synthesizing pseudo-WSI representations that reflect domain-specific morphological mixtures, while a teacher snapshot provides pseudo-reports to supervise the updated model without retaining past data. To address shifting reporting conventions, we distill domain-specific linguistic characteristics into a compact style descriptor and use it to steer generation. At inference, the model identifies the most compatible descriptor directly from the slide signal, enabling domain-agnostic setup without requiring explicit domain identifiers. Evaluated across multiple public continual learning benchmarks, our approach outperforms exemplar-free and limited-buffer rehearsal baselines, highlighting footprint-based generative replay as a practical solution for deployment in evolving clinical settings.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5125_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{KumPra_FootprintGuided_MICCAI2026,
author = { Kumari, Pratibha AND Reisenbüchler, Daniel AND Bozorgpour, Afshin AND Sadegheih, Yousef AND Choudhary, Priyankar AND Merhof, Dorit},
title = { { Footprint-Guided Exemplar-Free Continual Histopathology 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 paper proposes an exemplar-free continual learning framework for pathology report generation from whole-slide images, where each previously seen domain is compressed into a compact domain footprint and replayed through synthetic pseudo-WSI representations, without storing past slides or report pairs.
- 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 paper addresses a meaningful and practically relevant problem. It studies continual pathology report generation in a realistic deployment setting. This makes the problem formulation clinically relevant and well motivated.
2.The proposed replay mechanism is novel and well aligned with the task. The method compresses each domain into a compact footprint and uses it to synthesize pseudo-WSI representations, while an immediate teacher generates pseudo reports for replay. It provides a unified exemplar-free replay strategy for both the visual and language sides of report generation.
- 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 report-style prototype may not purely capture writing style. Since it is obtained by averaging domain-specific reports in a frozen text embedding space, the resulting vector may be dominated by frequent domain-specific lexical content rather than stylistic conventions.
2.The cross-domain comparability of independently learned domain codebooks is also unclear. Although all patch embeddings lie in the same frozen feature space, each domain codebook is obtained via domain-specific k-means compression. It remains uncertain whether reconstruction errors computed with these different codebooks are directly comparable for reliable domain matching at inference time.
3.More importantly, the validity of pseudo replay may be questionable because both the visual inputs and textual targets are synthetic. The pseudo-WSI representations are generated from compressed statistical summaries rather than real case-level structures, which may compromise semantic plausibility. As a result, the teacher-generated pseudo reports may be noisy or unreliable, raising concerns about whether such distillation can consistently preserve prior knowledge.
- 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?
This paper addresses meaningful problem, namely continual pathology report generation under privacy-constrained settings. The proposed method is overall coherent and technically reasonable, with no obvious fatal flaws, although some components are still relatively coarse and not fully convincing.
- Reviewer confidence
Somewhat confident (2)
- [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 meaningfully improves the paper and addresses my concerns to a reasonable extent.
(1) Regarding the report-style prototype, the authors now clarify that it should be interpreted more broadly as a domain report-language descriptor rather than a pure writing-style representation, which makes the method description more accurate.
(2) Regarding codebook comparability, the rebuttal does not fully establish a calibrated cross-domain distance, but it reasonably reframes the matching score as a relative ranking heuristic in a shared frozen embedding space, which is a more limited and acceptable claim.
(3) Regarding pseudo replay validity, the authors clarify that the goal is not faithful case-level reconstruction but functional latent rehearsal, and the additional empirical evidence—especially the strong ablation showing collapse without footprint-based replay and the competitive performance relative to high-buffer replay methods—supports that this mechanism is practically effective for retention.
Overall, while some of my original concerns are not fully resolved at the mechanism level, the rebuttal makes the scope of the claims more precise and provides enough empirical support to make the method coherent and convincing in its intended setting. I therefore support acceptance.
Review #2
- Please describe the contribution of the paper
The paper describes a footprint-based generative replay framework for WSI-to-histopathology-report generation in a continual-learning, domain-agnostic setup.
- 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 introduces a novel design that constructs compact domain footprints in a frozen patch-embedding space, which are used for pseudo-WSI and report generation within an exemplar-free generative replay framework. The setting is practical and addresses a realistic clinical deployment constraint.
-
The method achieves strong performance, which is comparable to the upper bound results obtained under non-CL settings
-
- 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 method relies heavily on pseudo WSIs and pseudo reports for replay, but their quality is not directly evaluated. How good are these pseudo samples, and is there any analysis linking their quality to final performance?
2.It is also unclear whether the method assumes the model is already well-trained on each domain before moving on. If not, how do you ensure stable footprints and reliable replay?
3.The footprint/codebook representation may be biased toward dominant patterns. How well does it capture rare but clinically important patterns (e.g., small lesion regions)? Since patch embeddings are quantized and subsampled, I suppose there is a risk of losing rare findings during replay. Is there any mechanism to address this?
4.I am not fully convinced by the need for domain-agnostic inference. In practice, domain information (e.g., organ or source) is usually available. How reliable is footprint matching for selecting the correct domain? If you eventually select a “seen” prototype for report generation, how well does it generalize to totally unseen domains?
- 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?
The paper presents a strong and practical approach with solid results. The handling of rare yet clinically critical patterns remains unclear and is the main factor affecting my recommendation.
- 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.
I thank the authors for their rebuttal. The response adequately addresses my main concerns. I therefore maintain my acceptance recommendation.
I note that some of my concerns are acknowledged by the authors but deferred to future work, and encourage the authors to discuss these limitations more explicitly in the camera-ready version.
Review #3
- Please describe the contribution of the paper
The main contribution of the paper “Footprint-Guided Exemplar-Free Continual Histopathology Report Generation” is the development of a storage-efficient, exemplar-free continual learning framework for whole-slide image (WSI) report generation that avoids retaining raw slides or patch features. It achieves this by summarizing each new domain into a visual footprint to synthesize pseudo-WSIs for generative replay, and by extracting domain-specific text style prototypes to dynamically condition the language model. This approach prevents catastrophic forgetting and adapts to varying reporting conventions across different hospitals without requiring explicit domain identifiers during inference.
- 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.By compressing each domain’s whole-slide images into a compact footprint, the method completely circumvents the need to retain raw WSIs or large feature archives. It also provided a way to mitigate storage cost and privacy concerns. 2.The framework handles shifts in hospital reporting conventions by extracting a lightweight text-style descriptor for each domain and utilizing it to condition the language model via style-prefix tokens.
- 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 generative replay mechanism synthesizes pseudo-WSIs by independently sampling patch features from a discrete codebook based on domain-specific histograms. This “bag-of-words” approach may discard the spatial topology between tissue patches, which are often critical for accurate histopathological diagnosis and comprehensive report generation. 2.The compression of a WSI images to a footprint may lead to information loss, especially when it is used for report generation that can have hallucination when the provided information is not complete. 3.How the model deal with the difference between WSI from different sites that can have visual variations? Can the footprint generation module deal with this kind of shift?
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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?
Although the paper has proposed an applicable way to deal with WSI continual learning issues, there are some weaknesses, as shown in the weakness part, which cannot be ignored.
- 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 author has clearly addressed my concern, given its still has some limitations, adding more details to the future work would be promising
Author Feedback
We thank all the reviewers. Meta: Mechanism concerns (topology, rare pattern, hallucination, replay fidelity) are addressed in R3-C1, R2-C3, R3-C2, R1-C3+R2-C1.The key empirical bound is Tab.2: our exemplar-free model matches high-buffer ER/DER on real WSIs, which bounds the content the footprint can lose. Requested CL medical-report comparison is in Tab.3+2 (ProgPrompt, CMRG-LLM). R1-C1: r^t is a domain report-language descriptor rather than pure style. Averaging frozen text embeddings captures phrasing and frequent finding terms, both helping the LM match domain t’s conventions. Tab.4 shows a consistent r^t gain on top of FR, confirming useful conditioning beyond the visual footprint. We will rephrase r^t in Sec.2.R1-C2 & R2-C4: Codebooks are per-domain supports in shared frozen-CONCH space, not aligned vocabularies. QE(t,X) is a relative ranking (Sec.2.5), not an absolute distance, and Tab.2 is produced under this rule. Domain-agnostic inference is desirable in domain-incremental CL literature, as test-time domain ID may be unavailable/insufficient in continual CPath deployment. For substantially distinct unseen domains, footprint matching serves only as a nearest-learned-footprint fallback. We do not claim zero-shot generalization beyond seen CL domains. HS-D (disjoint organs/centers, Tab.1-2) is the closest in-paper test. Calibrated QE is future work. R1-C3 & R2-C1: Pseudo WSIs-reports are designed for functional replay, not exact sample reconstruction. Their role is to provide latent rehearsal signals to preserve prior-domain knowledge. Tab.4 (removing FR collapses all metrics) and Tab.2 (exemplar-free matches high-buffer ER/DER without storing real WSIs) confirm that synthetic pairs support retention. R2-C2: Footprints are data statistics in frozen spaces (codebook from CONCH, prototype from BioGPT), so they don’t depend on VLM convergence and so don’t drift after extraction. Each episode uses the same training budget (reproducible stopping, not assumed convergence). Pseudo-replay supplements to current episode data training, competitive to ER/DER which replay real data (Tab 2). Further, assuming teacher noise compounds, BWT/ILM would degrade along the stream, however our results in Tab.2-4 show stability. R2-C3: CONCH embeddings are semantic, so rare morphologies sit on distinct manifold modes, and K=64 leaves room for rare codes alongside dominant tissue. We store diverse 50 slide histograms, so atypical compositions survive as slide-level entries and are sampled directly at replay. Rare-finding evaluation needs per-finding annotations not in public WSI-report datasets, and rare-aware sampling is future work. R3-C1: Our model consumes WSIs as an unordered set of patch embeddings (no positional encoding), so the model never uses spatial topology. Spatial awareness is contested implicitly in the literature, e.g., Shao et al (“Do MIL Models Transfer?”) report set-based AB-MIL outperforming spatially-aware TransMIL. Topology-aware replay is future work. R3-C2: The footprint never substitutes for a WSI at inference: the model consumes the real WSI’s embeddings, and the footprint only (i) seeds pseudo-replay and (ii) selects style prefix via argmin-QE. Compression loss hence cannot drive inference-time hallucination. Pseudo-replay is latent functional rehearsal, not pixel-true generation, and the empirical cost is bounded by Tab.4 (removing FR collapses AVG/ILM, BWT) and Tab.2.More general LM hallucination is out of scope for this CL contribution. R3-C3: Proposed method covers site shifts: when a new domain is observed, its patch distribution is compressed into a footprint and later used for replay, allowing model to rehearse prior distributions rather than relying on encoder invariance; needed since site shifts persist in CPath embeddings [cf. Thiringer et al.]. Details in Tab.1+2, REG (5 scanners), PathText/TCGA (multi-site), and HS-C/HS-D (cross-institution) cover site shifts. We will make this clearer in Sec. 3.1.
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 main weakness is that the central mechanism is not yet fully validated. The paper assumes that morphology codebooks, slide-level co-occurrence summaries, and patch-count priors are sufficient to reconstruct useful pseudo-WSI representations. However, the reviewers reasonably question whether such footprints preserve spatial topology, rare morphologies, and clinically decisive microscopic patterns. The paper also relies heavily on automatic text-generation metrics and qualitative examples, while clinical correctness, hallucination risk, and pathologist-level evaluation are not sufficiently addressed. The experimental table suggests competitive results against several CL baselines, but the interpretation remains limited without stronger validation of pseudo-replay fidelity and closer comparison to continual medical report-generation or latent-replay methods.
- 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.
All three reviewers agree accept. The rebuttal addressed the comments.
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 a relevant privacy- and storage-constrained continual pathology report generation problem, with a coherent exemplar-free footprint replay strategy and strong empirical support. Although some concerns remain regarding pseudo-replay fidelity, rare-pattern preservation, and domain matching, the rebuttal clarified the intended scope and reframed these as limitations rather than fatal flaws. I therefore recommend 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.
The paper addresses a clinically meaningful and practically important problem of continual histopathology report generation under exemplar-free constraints. Reviewers appreciated the novelty of the footprint-guided replay framework, which provides a unified mechanism for replaying both visual and textual knowledge without storing prior WSIs or reports. The proposed approach is well motivated from both privacy and storage perspectives and demonstrates strong empirical performance across continual learning settings. Several concerns were initially raised regarding the validity of pseudo replay, the interpretation of the report-style prototype, and potential information loss from compressed footprints. However, the rebuttal clarified the scope of the claims, provided additional empirical justification, and adequately addressed the reviewers’ main concerns. Multiple reviewers noted that the revised explanations made the framework more coherent and convincing in its intended setting. Overall, despite some remaining limitations that could be further explored in future work, the paper presents a novel and practically valuable contribution to continual multimodal pathology learning. I support acceptance.