{"id":"df0b3488-438d-40d5-8a86-24f1da21f40d","arxiv_id":"2501.12085","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"A WSI classification pipeline that clusters patch embeddings with K-means and encodes each cluster with Fisher vectors, evaluated on four benchmark datasets.","lead":"The paper combines K-means clustering with Fisher vector encoding to turn gigapixel pathology images into compact feature vectors for classification. The authors report accuracy gains over several baselines on HER2 scoring, mutation prediction, and metastasis detection tasks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Hyperparameters are explicitly 'decided based on performance for test dataset,' so the reported gains in Tables 1–4 can reflect test-set overfitting rather than genuine superiority; the central accuracy claim is unsupported without validation-controlled reruns.","rationale":"The reader's weakest assumption was the Fisher-vector formula; I agree that equation is nonstandard and unexplained: it writes FV as a concatenation of scaled first and second moments with constants c_j and ĉ_j and fixed π_m=0.2 and σ_m=0.1, but does not show responsibilities, normalization, or a derivation from a GMM. That matters for representational validity. Yet the single most load-bearing condition for the abstract's empirical claim is that reported accuracy measures generalization. Section 3's explicit statement that hyperparameters were decided based on performance on the test dataset directly violates that condition. Even a formally perfect Fisher-vector encoding cannot make test-set-tuned results evidence of 'superior accuracy' on unseen slides. The paper also provides no code, no error bars, no confidence intervals, and no description of a held-out validation set used before test evaluation, so there is no internal way to separate genuine gains from selection artifacts. I do not accuse the authors of deliberate misreporting; a careless sentence or an undisclosed validation split could make the statement harmless. That is exactly why the decisive test is to rerun the protocol with strict validation-only model selection and identical treatment of baselines. If the advantage persists, the central claim would be substantially rehabilitated, and the remaining issue would reduce to the Fisher-vector derivation and reproducibility. As written, however, the evaluation protocol cannot support the claimed generalization, so I would keep the reader's REJECT rather than adjusting it.","tokens_in":4919,"tokens_out":5420,"duration_ms":57711,"concrete_test":"Obtain the authors' code and splits, then rerun the full protocol with a strict three-way split: choose k, m, π_m, σ_m, augmentation parameters, encoder, and classifier using only a validation set; freeze all choices; evaluate once on the held-out test set. Apply the identical protocol to every baseline (AMIL, DFVC, Anand, Sekhar) and recompute Tables 1–4. If the proposed method no longer remains best, or if its test metrics drop substantially relative to the reported values, the 'superior accuracy' claim fails. As a secondary check, independently re-derive the Fisher-vector formula in Section 2 from the standard GMM Fisher-vector definition and verify it is a valid gradient of a log-likelihood.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is empirical: the proposed K-means plus Fisher-vector representation 'demonstrat[es] superior accuracy and scalability.' The most load-bearing condition for that claim is that the reported test-set numbers are honest estimates of generalization. Section 3 states: 'These hyperparameters are decided based on performance for test dataset,' immediately after listing k=10 clusters, m=5 Fisher centers, π_m=0.2, σ_m=0.1, feature scaling range, jitter level, and mixup α. If the test labels were used to select these hyperparameters, or the encoder/classifier choices, then Tables 1–4 compare models chosen after seeing the test answers. The accuracy, AUC, precision, recall, and F1 improvements over AMIL, Anand et al., and DFVC are then evidence of fitting these particular test sets, not evidence that the representation generalizes better. Unlike the Fisher-vector equation's mathematical opacity, which affects interpretability of the representation, this protocol issue alone invalidates the comparison stated in the abstract. A corrected Fisher-vector formula would not fix the evaluation leakage, and no amount of architectural detail can restore the generalization claim without a clean model-selection protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a scalable whole slide image (WSI) classification pipeline. Patches extracted from a WSI are embedded with a pretrained CNN/transformer, the patch embeddings are grouped with K-means clustering, and each cluster is summarized by a custom 'Fisher vector' computed from the patch embeddings. These per-cluster vectors are concatenated into a single WSI-level representation and classified by an attention-based multiple instance learning (AMIL) block. Experiments are reported on Warwick HER2 scoring and HER2 status, TCGA-BRCA HER2 status, TCGA-LUAD EGFR mutation prediction, and CAMELYON17 metastasis detection. The abstract claims that the method 'captures local and global tissue structures and yields robust performance ... demonstrating superior accuracy and scalability compared to other approaches.'","tokens_in":5252,"tokens_out":4983,"duration_ms":52903,"significance":"The topic is timely, and the idea of combining clustering with Fisher-vector-style codewords for WSI classification is plausible and worth exploring. The paper includes a relatively broad evaluation across four datasets and compares with several baselines, which is a strength. However, the evaluation protocol is compromised by an explicit statement that hyperparameters were chosen using the test dataset, and the Fisher vector equation is mathematically underspecified and is not a standard Fisher vector. These issues are load-bearing because the reported accuracy gains in Tables 1–4 are the sole evidence for the central claim. If the evaluation were redone with a clean validation protocol and a correctly defined Fisher vector, the method could become a useful contribution, but as presented the evidence does not support the claimed superiority.","major_comments":[{"comment":"The text states: 'These hyperparameters are decided based on performance for test dataset.' This is an explicit admission that test-set labels were used to select k, m, π_m, σ_m, feature scaling range, jitter level, mixup alpha, and the backbone per task. As a result, the accuracy, AUC, and F1 numbers in Tables 1–4 may reflect overfitting to the test sets rather than genuine generalization performance. The central claim of the abstract is unsupported without a clean model-selection protocol (e.g., a held-out validation set or nested cross-validation) and re-reporting of all test metrics after fixing hyperparameters without looking at test labels.","section":"Section 3, 'Hyperparameters' paragraph"},{"comment":"The Fisher vector function is not the standard Fisher vector (the gradient of the log-likelihood with respect to GMM parameters). The constants c_j and ĉ_j are never defined, the responsibilities s_ij are never defined, and fixing π_m=0.2 and σ_m=0.1 as scalar constants is inconsistent with fitting a Gaussian mixture model. The notation '1/n N∑ i=1' also mixes n and N. Because this equation is the core of the proposed representation, the method is not reproducible and its behavior cannot be assessed. Please provide a complete, standard definition of the Fisher vector (e.g., following Sánchez et al., 'Image classification with the Fisher vector: Theory and practice') or derive the formula from a specified probabilistic model.","section":"Section 2, Fisher vector equation"},{"comment":"In Tables 1, 2, and 3, the proposed method uses a different backbone feature extractor (EfficientNetV2-S, RegNetY-3.2GF, MoCo-v2, respectively) than the AMIL or Anand et al. baselines (ResNet-50 in each table). The reported accuracy differences could therefore be caused by the backbone rather than by the K-means and Fisher-vector aggregation. To support the claim that the proposed aggregation method is superior, the authors should compare against baselines using the same backbone features, or include an ablation that holds the backbone fixed while varying only the aggregation method. Table 4 is better in this respect but does not rescue the earlier tables.","section":"Tables 1–3 and Section 3, 'Ablation and validation studies'"}],"minor_comments":[{"comment":"The displayed formula '1/n N∑ i=1 FV(...)' uses both n and N without defining either; this appears to be a typo and should be corrected.","section":"Section 2, equation display"},{"comment":"The text says the concatenated vector is 'input into a permutation-invariant classifier,' but MLP, Swin Tiny, and ConvNeXt are not permutation-invariant, while AMIL is. Please clarify which classifier is used for the reported results.","section":"Section 2, 'permutation-invariant classifier'"},{"comment":"The elbow method is mentioned as determining the optimal number of clusters, but the next sentence sets k=10 for all datasets; please explain whether the elbow method was used or whether k was set a priori.","section":"Section 3, hyperparameter selection"},{"comment":"The SimCLR reference lists 'Tianyang Chen' as the first author; the correct first author is Ting Chen (Chen, T., Kornblith, S., Norouzi, M., Hinton, G.).","section":"Reference [20]"},{"comment":"The comparison with Sekhar et al. [18] uses the same MoCo-v2 backbone as the proposed method, but the differences between the two methods are not described; adding a brief explanation of how the proposed approach differs from that prior work would strengthen the comparison.","section":"Table 3 and references [16,18]"}],"recommendation":"reject","confidential_remarks":"The explicit statement that hyperparameters were 'decided based on performance for test dataset' is the most serious issue; it invalidates the reported accuracies as estimates of generalization. The Fisher vector equation is also not a standard formulation and is not reproducible. Even if the authors intended 'test dataset' to mean 'validation dataset,' the manuscript as written does not support the claimed superiority. I would suggest the editor consider requesting a significantly revised evaluation before any resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proposes a K-means pre-clustering step on patch embeddings before Fisher-vector coding, then concatenates per-cluster Fisher vectors into a WSI representation. That is a sensible, simple variation of DFVC, and the authors apply it to four clinically relevant tasks. The writing is clear and the empirical scope is adequate for a short paper. But the central accuracy claim does not survive contact with the evaluation protocol.\n\nSection 3 explicitly says hyperparameters are \"decided based on performance for test dataset,\" and the best backbone is selected per task. That means the numbers in Tables 1–4 compare configurations chosen after seeing the test labels. The reported gains over AMIL and DFVC may simply reflect overfitting to those test sets. The stress-test note is right: this is a load-bearing flaw, not a cosmetic one. No amount of architectural detail fixes it without a clean validation-controlled model-selection protocol.\n\nThe Fisher-vector equation is also underspecified. Constants c_j and \\hat c_j are never defined, π_m and σ_m are fixed numbers with no derivation from the GMM fit, and the formula is not the standard Fisher vector. The reader's weakest assumption is on point: the representation is the whole method, and we cannot tell exactly what it computes. Also, the baselines mostly use different backbones from the proposed method, so even the direction of the comparison is muddy. There are no error bars and no code.\n\nWhat the paper does well: the idea is simple and easy to grasp, and trying it on HER2, EGFR, and metastasis tasks is a reasonable first test. But it is incremental, and the missing protocol details make the empirical claims unsupported. I would not cite this as evidence of superiority, and I would not bring it to a reading group except as an example of test-set leakage. A serious referee could ask for a corrected derivation and a clean rerun, but as it stands this is a desk reject rather than a paper that merits referee time.","headline":"The K-means + Fisher-vector idea is a reasonable incremental twist on DFVC, but test-set hyperparameter tuning and an underspecified FV equation sink the reported accuracy claims.","tokens_in":5708,"tokens_out":2248,"would_cite":false,"duration_ms":25458,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes that clustering patch embeddings and encoding each cluster with a Fisher vector yields a compact whole-slide representation that matches or beats attention-based baselines on four pathology classification tasks.","keywords":["whole slide image","K-means clustering","Fisher vector","patch-based features","HER2 scoring","EGFR mutation prediction","lymph node metastasis","digital pathology"],"falsifier":"Compute the standard Fisher vector as the gradient of the GMM log-likelihood with respect to the mixture parameters for the same clusters, replace the paper's formula with it, and rerun the Warwick and TCGA-BRCA experiments. If accuracy does not change, the special formula is not the source of the result; if accuracy drops, the published formula is doing the work and needs a derivation. Also, shuffle cluster assignments before computing the Fisher vectors: if performance stays, clustering is not contributing.","tokens_in":4753,"feed_emoji":"🔬","tokens_out":5036,"duration_ms":47787,"temperature":0.7,"pith_summary":"The paper proposes a way to turn a gigapixel whole slide image into one compact vector for classification: cut the slide into patches, embed each patch with a pre-trained network, group the embeddings with K-means, and summarize each cluster with a Fisher vector computed from a Gaussian mixture model. The cluster Fisher vectors are concatenated and fed to a classifier. The authors claim this captures both local tissue appearance and global slide-level structure, and they report that it outperforms attention-based multiple-instance learning and other baselines on HER2 score prediction, HER2 status classification, EGFR mutation prediction, and lymph node metastasis detection. A sympathetic reader would care because a compact, permutation-invariant slide representation would remove the need to train on all patches jointly and would make gigapixel pathology images tractable for standard classifiers.","feed_headline":"K-means plus Fisher vectors classifies gigapixel slides","feed_subtitle":"Per-cluster encoding beats attention baselines on HER2, EGFR, and metastasis detection.","key_machinery":"The key machinery is the per-cluster Fisher vector. After K-means groups patch embeddings into K=10 clusters, each cluster's descriptor distribution is modeled with a Gaussian mixture model of m=5 centers, and the Fisher vector function, defined in the paper as scaled first and second moments of each embedding relative to the cluster centers with fixed scale constants c_j and ĉ_j and fixed mixture parameters π_m=0.2 and σ_m=0.1, summarizes the cluster. Concatenating these per-cluster Fisher vectors yields the slide-level representation, which is then classified by an attention multi-instance learning block or other classifier. The role of this machinery is to compress thousands of patch embeddings into a fixed-length vector that is permutation invariant and preserves both the identity of tissue patterns through cluster membership and their variation through the moments.","core_discovery":"The central claim is that clustering patch embeddings before Fisher-vector aggregation makes a whole-slide representation both more informative and more scalable than pooling all patches through an attention mechanism. On the Warwick HER2 challenge the method reaches 0.72 accuracy for three-class scoring and 0.80 accuracy for HER2+/HER2- classification; on TCGA-BRCA HER2 status it reaches 0.86 accuracy; on TCGA-LUAD EGFR mutation prediction 0.84 accuracy; and on CAMELYON17 metastasis detection 0.77 accuracy, in each case matching or exceeding the best compared baseline with the same or a lighter feature extractor. The authors interpret these results as evidence that per-cluster distribution summaries preserve the heterogeneity of tissue while a single concatenated vector remains manageable for a classifier.","pith_inferences":["The fixed GMM constants and the nonstandard Fisher vector formula suggest the method may work as a cluster-weighted moment pooling rather than a true probabilistic Fisher vector; renaming it and fitting the GMM would make the contribution easier to verify and compare.","Because the representation is permutation invariant and compact, it should extend to other gigapixel imaging domains, such as satellite or full-section electron microscopy, wherever patch-level features can be clustered.","A testable extension is to replace K-means with soft clustering or with a variational autoencoder's latent clusters; if accuracy holds, the essential ingredient is the per-cluster moment aggregation rather than the specific algorithm.","The reported gains over attention baselines might be partially attributable to the stronger feature extractors used in the proposed pipeline (e.g., RegNetY, MoCo-v2); an apples-to-apples comparison holding the backbone fixed would identify how much the aggregation alone contributes."],"forward_implications":["A single compact vector per slide means classifier training can be done with ordinary fully connected heads rather than patch-level attention, lowering memory use for large cohorts.","Because the pipeline is agnostic to the patch encoder, improvements in pre-trained encoders should transfer directly to slide-level accuracy without re-architecting the aggregation.","The method addresses multiple diagnostic tasks (HER2 scoring, mutation prediction, metastasis detection) with the same aggregation recipe, suggesting one representation can serve several clinical questions.","Using K-means clusters to group semantically similar regions gives a natural way to inspect which tissue patterns drive a prediction, since each cluster corresponds to a set of patches."],"supporting_citations":[{"why":"Supplies the deep Fisher vector coding method for WSI classification and the Fisher vector idea this paper adapts to per-cluster encoding.","marker":"[4]"},{"why":"K-means clustering is the grouping step that partitions patch embeddings into clusters.","marker":"[3]"},{"why":"Provides the GMM-based Fisher vector formulation and the fixed parameters π_m=0.2 and σ_m=0.1 used in the encoding.","marker":"[5]"},{"why":"Attention-based multiple instance learning is both a baseline and the final classifier block in the proposed pipeline.","marker":"[12]"},{"why":"The neural-network baseline on Warwick HER2 tasks and TCGA-BRCA against which the proposed method is compared.","marker":"[16]"},{"why":"The MoCo-v2-based baseline on TCGA-BRCA that the proposed method beats, showing a gain on the same backbone.","marker":"[18]"},{"why":"MoCo-v2 is the self-supervised backbone used for the TCGA-BRCA comparison.","marker":"[19]"},{"why":"ResNet-50 is the backbone for the AMIL baseline comparisons.","marker":"[14]"}],"fun_headline_variants":["Clustered Fisher vectors beat attention on gigapixel slides","K-means plus Fisher encoding scales WSI classification","Per-cluster Fisher vectors improve whole slide image classification","Efficient gigapixel slide representation via cluster-aware encoding","Cluster-then-encode: scalable WSI classification without attention"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole representation depends on the specific Fisher-vector formula and the fixed GMM parameters in Section 2; if that formula is not a faithful summary of each cluster's patch distribution, the method is not actually doing Fisher-vector aggregation and its reported gains would need a different explanation.","fun_headline_variants_meta":{"raw":{"variants":["Clustered Fisher vectors beat attention on gigapixel slides","K-means plus Fisher encoding scales WSI classification","Per-cluster Fisher vectors improve whole slide image classification","Efficient gigapixel slide representation via cluster-aware encoding","Cluster-then-encode: scalable WSI classification without attention"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000182,"raw_usage":{"total_tokens":1287,"prompt_tokens":896,"completion_tokens":391,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":310}},"tokens_in":512,"tokens_out":391,"duration_ms":4771,"temperature":1.0,"reasoning_tokens":310,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:31:19.886159+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the standard Fisher vector as the gradient of the GMM log-likelihood with respect to the mixture parameters for the same clusters, replace the paper's formula with it, and rerun the Warwick and TCGA-BRCA experiments. If accuracy does not change, the special formula is not the source of the result; if accuracy drops, the published formula is doing the work and needs a derivation. Also, shuffle cluster assignments before computing the Fisher vectors: if performance stays, clustering is not contributing.","supporting_citations":[{"cited_title":"Deep fisher vector coding for whole slide image classification","cited_arxiv_id":null,"evidence_quote":"Supplies the deep Fisher vector coding method for WSI classification and the Fisher vector idea this paper adapts to per-cluster encoding."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the GMM-based Fisher vector formulation and the fixed parameters π_m=0.2 and σ_m=0.1 used in the encoding."},{"cited_title":"Attention-based deep multiple instance learning","cited_arxiv_id":null,"evidence_quote":"Attention-based multiple instance learning is both a baseline and the final classifier block in the proposed pipeline."},{"cited_title":"Deep learning to estimate human epidermal growth factor receptor 2 status from hematoxylin and eosin-stained breast tissue images.Journal of Pathology Informatics, 11(1):19, 2020","cited_arxiv_id":null,"evidence_quote":"The neural-network baseline on Warwick HER2 tasks and TCGA-BRCA against which the proposed method is compared."},{"cited_title":"HER2 and FISH Status Prediction in Breast Biopsy H&E-Stained Images Using Deep Learning","cited_arxiv_id":"2408.13818","evidence_quote":"The MoCo-v2-based baseline on TCGA-BRCA that the proposed method beats, showing a gain on the same backbone."},{"cited_title":"Deep residual learning for image recognition","cited_arxiv_id":null,"evidence_quote":"ResNet-50 is the backbone for the AMIL baseline comparisons."}],"review_version":1}