{"id":"51fe5690-6025-470c-93ea-2cc3cec7e5d3","arxiv_id":"2505.08524","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A buffer-free, attention-filtered GMM latent replay method matches or outperforms existing continual learning baselines for domain-incremental whole slide image classification.","lead":"This paper introduces AGLR-CL, a continual learning method for whole slide image analysis that replays synthetic features from past datasets instead of storing images. It reports that the method beats other buffer-free continual learning baselines and usually matches rehearsal-based methods on biomarker and molecular status tasks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GMM replay assumes patch embeddings are independent within a slide; this untested assumption is the weakest link in the claim that synthetic data preserves past-domain signal.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: replay quality depends on the unverified assumption that independently sampling patch embeddings from a class-wise GMM reproduces training signals sufficient for the MIL aggregator. This is the right focal point because it is the central mechanism of the proposed method. The paper's experimental comparisons do not isolate this assumption: buffer-based Replay stores real patches, so its strong performance does not validate the generative model; the attention-filter ablation (Table 4) only varies which patches are used to fit the GMM, not whether within-slide dependencies matter. A direct synthetic-only probe would settle whether the generative replay distribution is faithful enough to support the central claim. One caveat is that the paper also lacks error bars, code, and a comparison to the authors' prior work [15], but those are secondary to the mechanism's validity. The proposed check would provide evidence for or against the independence assumption without requiring access to the authors' implementation, making it a practical and decisive test.","tokens_in":9254,"tokens_out":6099,"duration_ms":70204,"concrete_test":"Train a fresh AB-MIL classifier on synthetic slides generated from GMMs fitted to domain D1 (e.g., PAIP-CRC in sequence a2), using the same class balance and number of slides as D1, and evaluate it on the real D1 test set. Compare against a fresh AB-MIL trained on the real D1 slides. Run 5 seeds and report mean±std AUROC and weighted F1. If synthetic-only accuracy is close to real-only accuracy, the iid-from-marginal assumption is adequate; if it drops substantially (e.g., more than 10 points), the replay distribution lacks within-slide joint structure and the claim of matching buffer-based replay should be tempered.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The core mechanism (Sections 2.3–2.4) synthesizes each past-domain WSI by sampling a patch count from GMMt_count and then drawing that many patch embeddings independently from a class-wise GMMt_emb. This replaces the real joint distribution of a slide's patch set with a product distribution over the marginal embedding density. AB-MIL (Section 2.1) is permutation-invariant, but its output is a function of the whole multiset of patches via an attention-weighted sum, so within-slide composition—relative frequencies of tissue types, correlations among patches from the same microenvironment or stain batch—can carry label information. If that joint structure matters, replayed data misrepresent past domains exactly where no real data are stored. The attention filter (Section 2.3) makes this more acute: GMMs are trained only on the top q=80% of patches, and Table 4 shows that on a5 filtering reduces performance, so discarded patches are not uniformly irrelevant. The paper never measures the fidelity of synthetic slides nor ablates the iid assumption; the buffer-based Replay comparison uses real patches, so it does not test the generative model itself.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AGLR-CL, a buffer-free continual learning method for domain-incremental whole slide image (WSI) classification. The method uses a frozen pathology foundation model to embed patches, an AB-MIL aggregator, and per-episode Gaussian mixture models fitted to attention-filtered patch embeddings and to patch counts. In later episodes, synthetic WSI representations are sampled from these GMMs and combined with current real data, avoiding storage of raw WSIs. The authors evaluate on five domain-incremental sequences spanning MSI, TMB, PR, and HER2 prediction using TCGA, CPTAC, PAIP, and BCNB data, comparing against naive, joint, cumulative, regularization-based, and rehearsal-based baselines. They report ACC, ILM, and BWT based on weighted F1, AUROC, and AUPRC, plus an ablation of the attention-based filtering step. The central claim is that AGLR-CL is the best buffer-free method on most sequences and matches or approaches buffer-based Replay and GDumb.","tokens_in":9458,"tokens_out":5421,"duration_ms":58365,"significance":"If the empirical claims hold, the paper makes a practically relevant contribution: it offers a privacy-preserving way to perform continual learning on WSIs without storing original slides, using a relatively simple generative model in latent space. The choice of clinically meaningful tasks (biomarker and molecular-status prediction), the use of public datasets, and the inclusion of standard continual learning baselines are strengths. The attention-filtering idea is well motivated by the observation that not all patches are informative for the MIL aggregator. The paper also provides an ablation study and attention heatmaps. However, the empirical evidence currently rests on single-run point estimates with no uncertainty quantification, and the core generative assumption of independent patch embeddings is not validated. These issues limit the strength of the claims until addressed.","major_comments":[{"comment":"Every reported ACC, ILM, and BWT value is a single-run point estimate, with no seeds, error bars, confidence intervals, or significance tests. Some headline comparisons are close (e.g., a3 weighted-F1 ACC 67.97 for AGLR-CL vs 67.21 for GDumb; a1 AUPRC 38.38 for AGLR-CL vs 45.24 for SI), and the table itself contains cases where a buffer-free baseline beats AGLR-CL. Without variance estimates, the claims that AGLR-CL 'outperforms other buffer-free methods' and 'matches' buffer-based solutions are not statistically supported. Please report results over at least three random seeds per method and sequence and provide paired tests or confidence intervals.","section":"§4, Table 3"},{"comment":"Synthetic WSI generation samples a patch count from GMM_t_count and then draws that many patch embeddings independently from a class-wise GMM_t_emb, replacing the joint distribution of a slide's patch multiset with a product of marginal densities. AB-MIL, as described in §2.1, computes an attention-weighted sum over the multiset, so within-slide composition (e.g., proportions of tissue types, correlations among patches from the same microenvironment or stain batch) can carry label information. The paper does not measure how faithfully synthetic slide-level statistics reproduce the real data, and it does not ablate the independence assumption. In addition, the covariance structure in Eq. (2) is not specified; if full covariances are used in the UNI embedding space (which is 1024-dimensional), the parameter count per component is large and stability is a concern, while diagonal covariances would make the independence assumption even more restrictive. Please add a fidelity analysis of generated versus real embeddings and, if needed, extend the generative model to capture inter-patch dependence.","section":"§2.3–2.4, Eq. (1)–(3)"},{"comment":"The ablation shows that attention-based filtering is not uniformly beneficial. On sequence a5 (TMB), the filtered version is worse than the unfiltered version on weighted-F1 ACC (73.04 vs 73.58), AUROC (57.45 vs 58.60), and on most ILM/BWT entries, with only a small AUPRC improvement. The explanation that 'high variability of morphological alterations' makes discarded patches informative is post hoc and is not tested. Since the attention filter is presented as a central novelty ('ensuring high-quality synthetic samples'), the paper should characterize when filtering helps versus hurts and either adapt q per domain or temper the claim.","section":"§4, Table 4"},{"comment":"The statement 'we introduce domain incremental CL for MIL for the first time' is difficult to reconcile with the closely related prior work cited in the paper, especially the authors' own [15] on continual domain incremental learning for privacy-aware digital pathology and existing WSI continual learning methods such as ConSlide [9] and lifelong histopathology retrieval [25]. Please clarify precisely how AGLR-CL differs from [15] in problem setting and method. If [15] already uses MIL and generative replay, the novelty should be positioned as the attention-filtered latent GMM replay rather than the problem setting.","section":"§1, Contribution (1)"}],"minor_comments":[{"comment":"There are several typographical errors: 'Wholeslideimage' in the abstract, 'datastes' in the Introduction, and 'subsequentially' in §2.4. These should be corrected.","section":"§5 and throughout"},{"comment":"The choice of q=80% and the candidate sets for K are not justified or analyzed. A sensitivity study over q and K would strengthen the paper, particularly because the ablation shows that q has non-negligible effects.","section":"§3, Implementation Details"},{"comment":"The BWT sign convention appears inconsistent with intuition. For example, Naive on sequence a2 shows positive BWT for AUROC (1.27) and AUPRC (6.78) despite being fine-tuned only on the current episode. Please state the exact BWT definition used and explain how positive values should be interpreted.","section":"§4, Table 3"},{"comment":"The attention heatmap analysis is qualitative. A quantitative measure of overlap between high-attention regions and annotated ground-truth regions across sessions would provide stronger evidence that past knowledge is preserved.","section":"§4, Fig. 2"},{"comment":"No code or reproducibility statement is provided. Releasing code (or at least the trained GMM parameters and evaluation pipeline) would significantly increase the reproducibility of the empirical claims.","section":"§3, Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and practically important problem, and the core idea is reasonable. The main risks are the absence of uncertainty quantification in all reported results and the untested independence assumption in the generative replay mechanism. Both are fixable with additional experiments and analysis; if the authors address them, the paper could become a solid contribution to the WSI continual learning literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi — quick take on arXiv:2505.08524. The paper does something real: it shows you can run continual learning for WSI classification without storing old slides, by fitting per-domain GMMs to patch embeddings and counts, filtering patches by attention, and replaying synthetic embeddings. On five domain sequences spanning MSI, TMB, PR, and HER2, the method is the best buffer-free option and lands close to buffer-based Replay/GDumb on several. The ablation shows attention filtering helps on four of five sequences. The experimental scope across organs and centers is good, and the privacy framing is sensible.\n\nWhere it falls short, in decreasing order of importance. First, every number in the tables is a single run with no seed variation, error bars, or significance tests. Several margins against baselines are small (e.g., 73.04 vs 72.98 on a5 weighted F1), so the central \"outperforms\" claim rests on noise-prone point estimates. Second, the authors claim the first domain-incremental MIL method without comparing to their own MICCAI 2024 paper [15], which sounds like the direct predecessor. Maybe [15] is not MIL, but the text does not say so; that omission weakens both the novelty claim and the empirical case. Third, the generative model draws patch embeddings independently from a class-wise GMM. Since AB-MIL attends over the whole bag, joint structure among patches within a slide could carry label signal. The paper never tests the fidelity of synthetic slides or the independence assumption. The a5 ablation drop shows the attention filter is not universally harmless, which makes this gap concrete.\n\nNone of these are fatal. The central claim — that buffer-free GMM replay is a competitive privacy-preserving alternative to buffers — holds up at the level of table trends. The method is simple, the experiments are broad, and the paper is honest about the a5 drop. For a subfield paper, this is a useful incremental contribution.\n\nThis is a paper for computational pathology and continual learning people. It deserves a serious referee, but it should not be accepted as-is. Recommended verdict: revise with three fixes — multiple seeds with error bars, a direct comparison to [15], and ideally a replay-fidelity diagnostic that tests the independence assumption. With those, I'd be comfortable citing it.\n\n— Your take?","headline":"Solid buffer-free generative replay for WSI continual learning, but the missing comparison to the authors' own prior MICCAI paper and absent error bars undercut the empirical and novelty claims.","tokens_in":10015,"tokens_out":3917,"would_cite":true,"duration_ms":39667,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that an attention-filtered Gaussian mixture latent replay can match buffer-based continual learning for whole-slide image classification while storing no raw slides.","keywords":["Whole slide image analysis","Computational pathology","Biomarker screening","Continual learning","Domain incremental learning","Generative latent replay","Gaussian mixture models","Multiple instance learning"],"falsifier":"Construct a two-domain experiment where the first domain's class discrimination depends on a fixed inter-patch co-occurrence (for example, two patch types that must appear together in the same slide), and check whether the independent-draw GMM replay retains the first domain while the second is learned. A second decisive test: build a domain whose discriminative patches are systematically low-attention; if AGLR-CL forgets it while the unfiltered variant does not, the attention filter is discarding the signal it needs to replay.","tokens_in":9071,"feed_emoji":"🔬","tokens_out":8875,"duration_ms":81587,"temperature":0.7,"pith_summary":"Whole-slide images are gigapixel-scale and privacy-sensitive, so continually updating a classifier on new hospitals or organs would ideally not keep copies of old slides. This paper introduces a generative replay method that, after each training episode, fits Gaussian mixture models to the patch embeddings and patch counts of that domain and later resamples synthetic slides from those models. It reports that, across five domain-incremental sequences covering four biomarker tasks, this buffer-free approach outperforms other buffer-free continual learning methods and matches methods that store a replay buffer of real slides. The practical significance is that pathology classifiers could be updated across institutions and tissue types without storing raw patient images.","feed_headline":"Synthetic replay without raw slides matches buffer methods","feed_subtitle":"Attention-filtered Gaussian mixture replay curbs forgetting across organs and centers without storing slides.","key_machinery":"The load-bearing object is the GMM family: for each past episode, a class-wise multivariate Gaussian mixture captures the distribution of patch embeddings and a one-dimensional Gaussian mixture captures the distribution of patch counts per slide. To synthesize a past-domain slide, the method samples a patch count from the count GMM and then independently draws that many embeddings from the embedding GMM, forming a feature sequence that is fed to the attention-based MIL aggregator together with current-domain data. The attention-based filtering step—retaining only the top $q\\%$ of patches by attention score before fitting the mixtures—is what shapes the generative model toward diagnostically relevant tissue rather than background.","core_discovery":"AGLR-CL's central claim is that a domain can be compressed into a Gaussian mixture family—per-class multivariate GMMs over patch embeddings plus a univariate GMM over slide patch counts—and that replaying synthetic slides sampled from these mixtures preserves enough of the past domain to prevent catastrophic forgetting. An attention filter keeps only the top 80% of patches by the MIL aggregator's attention scores before the GMMs are fit, so the generative model is built from the tissue that actually drives classification. The paper validates this on microsatellite instability, tumor mutational burden, HER2, and progesterone receptor tasks across sequences that mix organs and centers, reporting that the method surpasses buffer-free baselines and is on par with buffer-based replay while never storing original slides.","pith_inferences":["The independence assumption in the GMM resampling is the natural stress point: if past-domain discriminative signal lives in inter-patch dependencies or spatial layout, replay quality would degrade; a variant that conditions each sampled embedding on its neighbors would test this directly.","The attention filter trades off against low-attention informative tissue; the TMB sequence already shows a decline when filtering is applied, suggesting an adaptive retention rate per domain could recover that lost signal.","The latent-replay scheme should transfer to other gigapixel multi-instance domains such as satellite imagery or full-slide cytology, where patch counts and patch features carry the label signal and raw data cannot be kept.","A formal privacy analysis—for instance, bounding how much a GMM snapshot reveals about any individual slide—would turn the absence of raw storage into a quantifiable guarantee."],"forward_implications":["Domain-incremental WSI classification can be done without a replay buffer, eliminating the main storage and privacy barrier to continual learning in clinical settings.","Attention-guided filtering before GMM fitting improves retention compared to unfiltered GMM replay on four of the five tested domain sequences.","The method retains knowledge across heterogeneous shifts (organ, center, mixed) and across four biomarker tasks, indicating the mechanism is not limited to one tissue or stain.","Buffer-free generative replay reaches accuracy comparable to buffer-based replay, which the paper identifies as the practical upper bound for these continual learning sequences."],"supporting_citations":[{"why":"Supplies the frozen patch encoder whose embeddings the method models with GMMs.","marker":"[3]"},{"why":"Defines the attention-based MIL aggregator that produces the attention scores used for filtering and slide-level predictions.","marker":"[10]"},{"why":"Provides the EM algorithm used to fit the Gaussian mixture parameters.","marker":"[4]"},{"why":"Gives the Bayesian Information Criterion used to select the number of GMM components.","marker":"[8]"},{"why":"The buffer-based experience replay baseline the method is compared against and seeks to match.","marker":"[20]"},{"why":"The GDumb rehearsal baseline, another buffer-based method in the comparison.","marker":"[19]"},{"why":"The EWC regularization baseline for catastrophic forgetting, which the method outperforms.","marker":"[12]"},{"why":"The Learning-without-Forgetting baseline, a regularization approach that does not store past data.","marker":"[16]"},{"why":"The Synaptic Intelligence baseline, another buffer-free regularization method in the comparison.","marker":"[24]"}],"fun_headline_variants":["Synthetic replay without raw slides matches buffer performance","GMM synthetic replay prevents forgetting in WSI tasks","Attention-filtered latent replay for privacy-aware lifelong WSI learning","No buffers, no raw slides: generative replay for domain shifts","Synthetic latent replay keeps knowledge across organs and centers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything hinges on the assumption that sampling a patch count and then drawing patch embeddings independently from a class-wise Gaussian mixture faithfully reproduces the training signal of a past domain, which means any task-relevant structure living in correlations between patches or in low-attention tissue would be missing from replay.","fun_headline_variants_meta":{"raw":{"variants":["Synthetic replay without raw slides matches buffer performance","GMM synthetic replay prevents forgetting in WSI tasks","Attention-filtered latent replay for privacy-aware lifelong WSI learning","No buffers, no raw slides: generative replay for domain shifts","Synthetic latent replay keeps knowledge across organs and centers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000515,"raw_usage":{"total_tokens":2471,"prompt_tokens":890,"completion_tokens":1581,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":1502}},"tokens_in":506,"tokens_out":1581,"duration_ms":12749,"temperature":1.0,"reasoning_tokens":1502,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:52:07.918188+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a two-domain experiment where the first domain's class discrimination depends on a fixed inter-patch co-occurrence (for example, two patch types that must appear together in the same slide), and check whether the independent-draw GMM replay retains the first domain while the second is learned. A second decisive test: build a domain whose discriminative patches are systematically low-attention; if AGLR-CL forgets it while the unfiltered variant does not, the attention filter is discarding the signal it needs to replay.","supporting_citations":[{"cited_title":"Journal of the royal statistical society: series B (method- ological) 39(1), 1–22 (1977)","cited_arxiv_id":null,"evidence_quote":"Provides the EM algorithm used to fit the Gaussian mixture parameters."},{"cited_title":"The computer journal41(8), 578–588 (1998)","cited_arxiv_id":null,"evidence_quote":"Gives the Bayesian Information Criterion used to select the number of GMM components."},{"cited_title":"Advances in Neural Information Processing Systems32 (2019)","cited_arxiv_id":null,"evidence_quote":"The buffer-based experience replay baseline the method is compared against and seeks to match."},{"cited_title":"In: Computer Vision–ECCV 2020: 16th Euro- pean Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16","cited_arxiv_id":null,"evidence_quote":"The GDumb rehearsal baseline, another buffer-based method in the comparison."},{"cited_title":"In: International conference on machine learning","cited_arxiv_id":null,"evidence_quote":"The Synaptic Intelligence baseline, another buffer-free regularization method in the comparison."}],"review_version":1}