{"id":"527ecf4a-09e3-4be3-a8ef-eb2f28a8d685","arxiv_id":"2506.15365","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"FedWSIDD distills synthetic slides at each hospital, aggregates them at a server, distributes them to all hospitals, and uses them as extra training data to improve local whole slide image classifiers with heterogeneous models.","lead":"This paper proposes FedWSIDD, a federated learning method for whole slide image classification in which hospitals share small sets of synthetic tissue images instead of model weights. The method is tested on two public breast cancer datasets and reports higher average accuracy than several federated learning baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (3) matches only the mean feature vector of real and synthetic slides, so nothing constrains patch diversity or class separation; the claim that the distilled slides are informative for attention-based MIL models is not established.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: Eq. (3) matches only first moments and does not enforce diversity, class separation, or transferability to other feature extractors. My stress-test agrees that this is the most load-bearing issue. If the synthetic slides collapse to class-mean prototypes, the central mechanism of FedWSIDD is not a distillation of informative WSI content but a class-mean prior, and the reported accuracy advantages would need to be explained by other factors, such as extra data volume or test-set-tuned hyperparameters. The reader's other points (missing code, privacy analysis, hyperparameter selection on the test set) are fixable and support a CONDITIONAL verdict, but they are not as conceptually central. The paper does provide repeated runs, paired t-tests, and consistent trends across two datasets, which is real evidence. However, the absence of a direct check of synthetic slide quality leaves the strongest claim under-supported. A synthetic-only training experiment would settle whether Eq. (3) actually yields useful class-discriminative slides, so I recommend keeping the reader's CONDITIONAL verdict unchanged.","tokens_in":9187,"tokens_out":7164,"duration_ms":78928,"concrete_test":"Using the paper's exact distillation settings (M=10, B=100, frozen ResNet50, Adam, one-shot communication), train CLAM on synthetic slides only for CAMELYON16 and CAMELYON17, then evaluate on the real held-out test slides. If synthetic-only accuracy is near chance or substantially below real-slide-only training, Eq. (3) does not produce class-discriminative synthetic slides. As a complementary diagnostic, compute the average pairwise cosine similarity between patch embeddings within a synthetic slide; values close to 1 would confirm that the mean-matching objective allows patch collapse, undermining the informativeness claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central transfer mechanism is Eq. (3), where synthetic patches are optimized so that the mean feature vector of a synthetic slide matches the mean feature vector of one sampled real slide, per class, in the frozen ImageNet ResNet50 feature space. This objective is invariant to the arrangement and diversity of patch embeddings: a bag of B identical patches whose common embedding equals the target mean is an equally good minimizer as a diverse bag. CLAM, ABMIL, and TransMIL all rely on attention or score pooling over patch embeddings, so they are sensitive to the distribution of patch-level evidence, not just the bag mean. If the optimized synthetic slides collapse toward a single prototype per class, they cannot represent tumour heterogeneity (e.g., micro- vs. macro-metastasis in CAMELYON17) and the reported gains would not follow from the distillation quality itself. Table 3 additionally varies only the MIL head while keeping the shared ResNet50 feature extractor, so the paper's flexibility claim for centres using different feature extractors is untested. The test-set-based selection of M and B in Fig. 2 further means the headline numbers are optimistically biased. The combination of an unconstrained first-moment objective and test-set hyperparameter selection leaves the core informative-synthetic-slide claim without direct support.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FedWSIDD, a federated learning framework for whole-slide image (WSI) classification in which each centre distils a small set of synthetic slides by matching the mean feature vector of real and synthetic patches in a frozen ResNet50 feature space, with stain normalization applied during distillation. The synthetic slides, rather than model parameters, are uploaded to a server, aggregated, and redistributed to all centres, which then train their local MIL models on real plus received synthetic slides. Experiments on CAMELYON16 and CAMELYON17 compare FedWSIDD with standard and personalised federated learning baselines under homogeneous and heterogeneous MIL-head settings, reporting higher average accuracy (CA16 90.1%, CA17 81.2% in the homogeneous case) together with ablations of the number of synthetic slides and patches and of stain normalization.","tokens_in":9409,"tokens_out":5927,"duration_ms":57951,"significance":"If the reported results are reproducible, FedWSIDD offers a genuinely interesting communication medium for federated WSI classification: exchanging compact synthetic slides instead of model weights makes the framework agnostic to local MIL architectures and potentially communication-efficient. The paper provides quantitative comparisons with several baselines, standard deviations over five runs, paired t-test p-values, and a stated intention to release code, which are all useful. However, the specific claims around informativeness of the distilled slides, privacy preservation, and flexibility to different feature extractors currently go beyond what the experiments demonstrate.","major_comments":[{"comment":"The distillation objective matches only the mean feature vector of one real slide and one synthetic slide per class. This objective is invariant to permutation and duplication within the synthetic bag, so an optimal solution could be a bag of identical patches whose common embedding equals the target mean. CLAM, ABMIL, and TransMIL all use attention or score pooling over patch embeddings and are therefore sensitive to the distribution of patch-level evidence, not just the bag mean. The paper should demonstrate that the optimized synthetic patches are actually diverse and class-discriminative, for example by reporting pairwise distances among synthetic patch embeddings or by comparing against a baseline that transmits random patches matched only in mean. Without such evidence, the core claim that the synthetic slides are informative for MIL models is not established.","section":"Section 2.3, Eq. (3)"},{"comment":"The headline hyperparameters M=10 and B=100 are selected from the test-set ablation in Fig. 2 and then used to produce the results in Tables 1-3. This constitutes selection on the test set and makes the reported numbers optimistically biased. The authors should either choose hyperparameters on a held-out validation split or report the full grid of Fig. 2 with honest error bars and clearly state that the reported numbers come from the same grid that was used to select them.","section":"Section 3.2 and Fig. 2"},{"comment":"The claim that exchanging synthetic slides 'preserves patient privacy' or 'enhanc[es] privacy' is asserted without any privacy evaluation. There is no membership-inference attack, no model-inversion experiment, no discussion of information leakage through the feature-matching objective, and no differential privacy guarantee. Because the synthetic patches are optimized to match real slide feature means, they may retain sensitive information, so this claim needs concrete empirical or theoretical support before it can be accepted.","section":"Abstract and Section 3.4"},{"comment":"The heterogeneous-model experiment varies only the MIL head (CLAM, TransMIL, ABMIL) while all centres use the same frozen ImageNet ResNet50 feature extractor. The paper's stated flexibility for centres with different feature extractors is therefore untested. The distillation objective in Eq. (3) is defined in the feature space of a specific extractor, so if a centre uses a different feature extractor, the transferred synthetic slides may not be useful. Please either test cross-extractor transfer or soften the flexibility claim to apply only to MIL head heterogeneity.","section":"Section 3.3, Table 3"},{"comment":"The description of the distillation loop is missing details needed for reproducibility. The text states that during each distillation iteration the method iterates through all classes and samples one real slide and one synthetic slide, and that there are 1000 distillation rounds, but it does not specify how many real slides are used per class per iteration, how synthetic patches are initialized and updated per round beyond the Adam learning rate, whether the sampled real slide changes across iterations, or how stain normalization is inserted into the differentiable pipeline. Please provide the full algorithm as pseudocode or pseudocode-like steps.","section":"Section 2.3 and Section 3.2"}],"minor_comments":[{"comment":"The captions of Tables 2 and 3 refer to 'FedDFP', which does not appear in the method list; this appears to be a typo for 'personalised FL methods' or similar.","section":"Tables 2 and 3 captions"},{"comment":"References [17] and [18] are duplicate entries for the same paper (Jia et al., ECCV); they should be merged into a single reference.","section":"References [17] and [18]"},{"comment":"The summation indices in Eq. (3) run from i=0 to T and b=0 to B, which is off by one; they should be i=1 to T and b=1 to B (or 0 to T-1 and 0 to B-1).","section":"Equation (3)"},{"comment":"The label 'w SN（ 90.1%)' uses a full-width parenthesis and a misplaced space; this should be formatted consistently with the rest of the paper.","section":"Figure 3"},{"comment":"The text cites 'stain normalisation [22]' and then 'stain normalisation [10]' in the same paragraph; please clarify which stain normalization method is actually used in the pipeline.","section":"Section 2.3"},{"comment":"The statement 'The code is available at FedWSIDD' provides no URL or repository identifier; please include an actual link.","section":"Abstract"},{"comment":"The phrase 'we are the first to successfully extend this concept to WSI classification' is too strong given that reference [10] already applies dataset distillation to histopathology image classification; please rephrase to acknowledge prior work.","section":"Section 2.3"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the manuscript should be positioned more carefully against reference [10], which already applies dataset distillation to histopathology, and against reference [14]; the 'first' claim in Section 2.3 should be checked. The duplicated reference [17]/[18], the 'FedDFP' caption typos, and the missing code link suggest that the camera-ready needs a careful proofread. The main technical concern is whether the first-moment matching objective in Eq. (3) truly yields informative synthetic slides for attention-based MIL models; this should be addressed with additional analysis or experiments before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"FedWSIDD is worth a look: it exchanges distilled synthetic slides instead of model weights in federated WSI classification, and the empirical comparison against a broad set of FL and pFL baselines on CAMELYON16/17 is actually decent. The tables are run five times, report standard deviations, and include paired t-tests. The gains are only a few accuracy points over the strongest baselines (90.1 vs 88.6 on CA16, 81.2 vs 79.3 on CA17), but they are consistent, and the heterogeneous-MIL-head experiment is a nice addition. The specific combination of dataset distillation with stain-normalized feature matching for federated WSI is new, though it is a direct extension of FedDM/DESA/FedD3 and the authors' own histopathology distillation work [10].\n\nThe soft spots are real but fixable. The biggest one is the distillation objective in Eq. (3): it matches the mean feature vector of a synthetic slide to the mean of one real slide, per class. Nothing in the loss prevents the B synthetic patches from collapsing toward a single prototype per class. CLAM, ABMIL, and TransMIL are attention-based and care about the distribution of patch evidence, not just the bag mean. If the synthetic slides are not diverse, the reported improvement is hard to attribute to the distillation quality itself. The paper gives no visual or distributional evidence that synthetic slides retain intra-class heterogeneity. This is the paper's load-bearing assumption, and it is not demonstrated.\n\nSecond, the hyperparameters M and B are selected using the same test set that produces the headline numbers (Fig. 2). That makes the reported gains optimistically biased; the t-tests are computed after the fact on the same set. Moving the choice to a validation split, or reporting the full grid, would fix this.\n\nThird, the privacy claim is asserted, not examined. Synthetic slides are optimized to match feature means of real slides; that can leak information. A proper privacy discussion or attack evaluation is needed. The code is also not actually accessible—the abstract says it is 'available at FedWSIDD,' which is not a link.\n\nFinally, the flexibility claim is only tested for MIL heads, not feature extractors. All clients share the frozen ResNet50, so the claim that FedWSIDD works with different feature extractors is untested. This is a minor point, but it should be stated honestly.\n\nOverall, the central idea is coherent and the evidence is solid enough to deserve a serious referee. I would send it to review, with the expectation of major revisions. The authors need to address the diversity and transferability of synthetic slides, move hyperparameter selection to validation, add privacy analysis, and provide code. The paper is for people working on FL for pathology and dataset distillation; they will find the synthetic-slide exchange idea useful even if the current evidence is conditional.","headline":"A plausible and well-executed empirical paper on federated WSI classification via synthetic slides, but the distillation objective is under-constrained and the headline numbers are partly tuned on the test set.","tokens_in":9967,"tokens_out":3316,"would_cite":false,"duration_ms":34063,"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":"FedWSIDD claims that federated whole-slide image classification works better when centres exchange distilled synthetic slides instead of model parameters, and reports higher accuracy on two lymph-node metastasis benchmarks.","keywords":["federated learning","whole slide image classification","dataset distillation","multiple instance learning","stain normalisation","CAMELYON16","CAMELYON17","heterogeneous models"],"falsifier":"Train a MIL classifier on one centre's real slides plus FedWSIDD synthetic slides from another centre, then swap the feature extractor from ResNet50 to a vision transformer; if the accuracy gain over local-only training disappears, the heterogeneous-flexibility claim fails. More directly, compute the average pairwise distance between synthetic slides of the same class generated for one centre: if it is near zero while real same-class slides show large spread, the mean-matching objective has collapsed intra-class diversity and the synthetic set cannot represent tumour heterogeneity.","tokens_in":8950,"feed_emoji":"🩺","tokens_out":5436,"duration_ms":53452,"temperature":0.7,"pith_summary":"This paper proposes that the right unit of exchange in federated whole-slide image classification is not the model update but a compact set of synthetic slides. Each centre distils its real slides into synthetic patches by matching mean feature vectors, sends the synthetic slides to a server, which redistributes the union to all centres for local training. On CAMELYON16 and CAMELYON17, the scheme reports higher average accuracy than standard federated methods (FedHisto, FedDyn, MOON, FedImpro, FedMut) and personalised federated methods (FedDM, FedProto, DESA, FedD3, FedDGM, SGPT), in both homogeneous and heterogeneous local-model settings. The authors argue that this makes FL practical for resource-heterogeneous pathology centres and that transmitting synthetic images rather than patient-derived patches offers a privacy benefit.","feed_headline":"Synthetic slides beat weight-sharing in federated slide AI","feed_subtitle":"FedWSIDD ships distilled synthetic slides between hospitals and reports higher accuracy on CAMELYON16/17.","key_machinery":"The load-bearing object is the synthetic slide: a bag of $B=100$ learnable $64\\times64$ patches, optimised to match the class-conditional mean of real patch features after stain normalisation. Matching happens in the embedding space of a shared ImageNet-pretrained ResNet50, not in pixel space and not against a particular MIL classifier, which is what makes the synthetic slides consumable by any MIL head (CLAM, ABMIL, TransMIL). The communication mechanism is the set union of these synthetic slides; because the server never touches model weights or gradients, centres can run different architectures and still share information.","core_discovery":"The central claim is that a feature-space dataset distillation objective, augmented with stain normalisation, can compress a local WSI dataset into a small number of synthetic slides that carry enough class-discriminative information to improve every participating centre's MIL classifier. In Eq. (3), each distillation step samples one real and one synthetic slide per class, normalises both, and minimises the squared distance between the mean feature vector of the real patches and the mean feature vector of the synthetic patches. The optimised synthetic slides are transmitted instead of model weights; the server aggregates them by set union and redistributes them. Empirically the paper reports 90.1% average accuracy on CAMELYON16 and 81.2% on CAMELYON17 under homogeneous CLAM models, and 84.8% and 77.6% respectively under heterogeneous models, with p-values below 0.05 against the baselines it compares.","pith_inferences":["Editorial inference: the privacy benefit is asserted from the synthetic nature of the images but not formally demonstrated; a membership-inference attack on the transmitted synthetic slides would be a concrete way to test whether patient information is actually protected.","Editorial inference: the one-shot protocol leaves multi-round communication unexplored; iterative rounds of distillation could either amplify the gains or introduce drift, and the paper does not address that trade-off.","Editorial inference: the mean-feature matching objective is sensitive to the choice of stain normalisation; swapping Macenko normalisation for a learned normalisation method would likely change the distilled slides' quality and could be a testable extension.","Editorial inference: the paper does not quantify communication cost in bytes against model-weight exchange; a direct comparison of transmission size versus FedHisto or FedProto would clarify the practical bandwidth advantage."],"forward_implications":["If each centre receives synthetic slides distilled from other centres' data, local MIL training on the merged set improves over local-only training and over existing federated methods, so collaboration can happen without exchanging weights or gradients.","Because the synthetic patches are optimised in the feature space of a fixed pretrained extractor rather than against one classifier, the same transmitted slides can be plugged into CLAM, ABMIL, and TransMIL, enabling heterogeneous local models.","A single round of communication, preceded by 1000 local distillation rounds, is sufficient for the reported gains, meaning FedWSIDD can operate with one-shot communication.","Since only synthetic images move between centres, the authors claim the protocol desensitises patient information, while noting that the synthetic nature reduces explainability of the shared data.","The reported p-values below 0.05 across many baseline comparisons suggest that the accuracy gains are unlikely to be pure chance under the paper's fixed-split experimental protocol."],"supporting_citations":[{"why":"Supplies the dataset distillation paradigm that FedWSIDD adapts from whole-image datasets to whole-slide images.","marker":"[28]"},{"why":"Provides the feature-space dataset distillation approach for histopathology that the local synthetic-slide objective builds on.","marker":"[10]"},{"why":"Supplies the stain normalisation method integrated into the distillation objective to reduce H&E appearance variation.","marker":"[22]"},{"why":"CLAM is the MIL classifier used for the homogeneous-model experiments and part of the heterogeneous model pool.","marker":"[21]"},{"why":"FedHisto is the slide-level federated pathology baseline, representing the standard weight-sharing approach FedWSIDD is designed to improve upon.","marker":"[20]"},{"why":"A prior federated dataset distillation method via synthetic anchors that FedWSIDD compares against and claims to outperform.","marker":"[14]"},{"why":"CAMELYON16 is one of the two evaluation datasets, for breast cancer lymph-node metastasis detection.","marker":"[3]"},{"why":"CAMELYON17 is the other evaluation dataset, with five centres and four-class lymph-node status classification.","marker":"[2]"},{"why":"FedProto is a personalised federated baseline sharing class prototypes, used as a comparison for heterogeneous-model performance.","marker":"[26]"},{"why":"ABMIL is one of the MIL classifiers in the heterogeneous model pool used to simulate centres with different architectures.","marker":"[16]"}],"fun_headline_variants":["FedWSIDD: Distilled slides replace weight sharing in WSI","Synthetic slides, not weights, win federated WSI tasks","Federated WSI gets a distillation upgrade: send slides","FedWSIDD: improving histopathology with dataset distillation","Distill your WSI data for better federated learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on Eq. (3)'s assumption that matching the mean feature vector of synthetic patches to the mean feature vector of real patches, per class, produces synthetic slides that are class-discriminative and useful to every centre's MIL model; nothing in the objective enforces intra-class diversity, inter-class separation, or transferability to feature extractors other than the shared ResNet50.","fun_headline_variants_meta":{"raw":{"variants":["FedWSIDD: Distilled slides replace weight sharing in WSI","Synthetic slides, not weights, win federated WSI tasks","Federated WSI gets a distillation upgrade: send slides","FedWSIDD: improving histopathology with dataset distillation","Distill your WSI data for better federated learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00065,"raw_usage":{"total_tokens":2985,"prompt_tokens":952,"completion_tokens":2033,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":1946}},"tokens_in":568,"tokens_out":2033,"duration_ms":18485,"temperature":1.0,"reasoning_tokens":1946,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:35:44.066354+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a MIL classifier on one centre's real slides plus FedWSIDD synthetic slides from another centre, then swap the feature extractor from ResNet50 to a vision transformer; if the accuracy gain over local-only training disappears, the heterogeneous-flexibility claim fails. More directly, compute the average pairwise distance between synthetic slides of the same class generated for one centre: if it is near zero while real same-class slides show large spread, the mean-matching objective has collapsed intra-class diversity and the synthetic set cannot represent tumour heterogeneity.","supporting_citations":[{"cited_title":"In: IEEE International Symposium on Biomedical Imaging (ISBI)","cited_arxiv_id":null,"evidence_quote":"Supplies the stain normalisation method integrated into the distillation objective to reduce H&E appearance variation."},{"cited_title":"Medical image analysis76, 102298 (2022)","cited_arxiv_id":null,"evidence_quote":"FedHisto is the slide-level federated pathology baseline, representing the standard weight-sharing approach FedWSIDD is designed to improve upon."},{"cited_title":"JAMA318(22), 2199–2210 (2017)","cited_arxiv_id":null,"evidence_quote":"CAMELYON16 is one of the two evaluation datasets, for breast cancer lymph-node metastasis detection."},{"cited_title":"IEEE Transactions on Medical Imaging38(2), 550–560 (2019)","cited_arxiv_id":null,"evidence_quote":"CAMELYON17 is the other evaluation dataset, with five centres and four-class lymph-node status classification."},{"cited_title":"In: Proceedings of the AAAI Conference on Artificial Intelligence","cited_arxiv_id":null,"evidence_quote":"FedProto is a personalised federated baseline sharing class prototypes, used as a comparison for heterogeneous-model performance."}],"review_version":2}