{"id":"bbe3b25f-e408-4002-9b35-79ff3a96a94a","arxiv_id":"2411.08936","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper proposes clustering patch embeddings into ten cluster means and classifying the mean vectors with attention MIL, reporting 0.83 accuracy on lung EGFR and 0.75 on Camelyon17 metastasis classification.","lead":"This paper tests a way to shrink gigapixel tissue images into ten summary vectors before classifying them, and reports accuracy on lung and breast cancer slide datasets. It matters because if this kind of compression works, whole-slide analysis could run on smaller computers and become more practical in hospitals.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported accuracies may depend on transductive clustering: K-means and the elbow method appear to be fit on all slides, including test slides, before any train/test split; if so, the 0.83/0.75 results are not a valid held-out evaluation.","rationale":"The reader's weakest assumption targeted the nucleus-count threshold in Section 3, which is plausible but less central: even if that threshold excluded informative low-cellularity patches, it would be a preprocessing choice that could be ablated and might affect accuracy in either direction. The clustering-scope ambiguity is more load-bearing because it threatens the validity of every quantitative result in the paper. The text is genuinely ambiguous about whether K-means is fit globally, per slide, or on training data only, and no train/test split or error bars are reported. If global clustering includes test slides, the representation itself leaks test-set information, so the comparison to baseline [21] is not a fair held-out evaluation. If clustering is per slide, the proposed representation lacks a consistent cluster ordering, so the permutation-invariance claim is not established. The paper does have useful elements: multiple backbones are compared, a no-clustering ablation is included, and the idea of summarizing a WSI by a small set of cluster means is coherent. However, none of these positives resolves the central methodological gap. I would not reject outright, because the pipeline could be valid if clustering was train-only and the authors can clarify; but as written, the evidence is insufficient to reach a verdict, so UNVERDICTED is the most honest recommendation. If the requested test shows no accuracy drop, the verdict could move back to CONDITIONAL or ACCEPT depending on the remaining split and baseline details.","tokens_in":8657,"tokens_out":7316,"duration_ms":69290,"concrete_test":"Re-run the pipeline with an explicit split: hold out a test set of slides (and ideally of patients) before any clustering. Fit K-means with k=10 on patch embeddings from training slides only, freeze the centroids, and encode each test slide as the 10 concatenated cluster means exactly as described in Section 3. Report accuracy, precision, and recall on the same test set. Also run the alternative of per-slide clustering with a set-based classifier such as attention MIL over unordered cluster means, and verify permutation invariance by randomizing cluster order. If train-only clustering yields substantially lower accuracy than the reported 0.83/0.75, the headline numbers are inflated by transductive clustering. If the authors can confirm they already fit clustering only on training data and report the exact split, the concern is resolved.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The most load-bearing uncertainty is not the nucleus-count threshold but the scope of the K-means fit. Section 3 says patch embeddings 'were then clustered using the K-means algorithm' and that the elbow method picks k=10 'for our dataset'; Section 4 repeats that clustering is applied to 'the features of both datasets (TCGA and Camelyon-17)'. The paper never states that the K-means centroids, or the elbow curve, are computed on a training split only and then frozen before encoding test slides. If the 10 centroids are learned from patches from all slides, then each test slide's cluster-mean representation is a function of the test-set feature distribution: the centroids encode unlabeled test information. That transductive step can inflate the 0.83 and 0.75 accuracies relative to a proper supervised baseline, and it would undermine the claim that a compact cluster-mean vector faithfully summarizes a single gigapixel slide. If instead clustering is done separately per slide, cluster indices have no semantic correspondence across slides, so concatenating cluster means and feeding them to an MLP or Swin transformer is not a well-defined permutation-invariant representation. Either reading makes the reported numbers uninterpretable without a precise statement of when and how clustering is fit. The low-nucleus threshold is a real modeling choice, but it does not threaten the validity of every number in Tables 1 and 2 in the same way.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pipeline for whole-slide image (WSI) classification in which each slide is preprocessed to remove artifacts, detect tissue, and tile into 512x512 patches; patches with fewer than 10 detected nuclei are discarded. Patch features are extracted with pretrained encoders (SimCLR, ResNet-50, EfficientNet, RegNet, ConvNeXT-Tiny, Swin-Tiny), K-means clustering is applied to the resulting patch embeddings with k selected by the elbow method, and each slide is represented by the concatenated mean vectors of the k clusters. This compact representation is then classified with an MLP, a Swin-Transformer, or an attention-based MIL classifier. The paper reports accuracy, precision, and recall for binary EGFR mutation classification on a TCGA lung dataset (best 0.83 accuracy with ResNet-50 + clustering + attention MIL) and binary metastasis classification on Camelyon17 (best 0.75 accuracy), and concludes that the cluster-mean representation matches or exceeds a previously proposed baseline.","tokens_in":8883,"tokens_out":4823,"duration_ms":46004,"significance":"If the reported results are valid, the compact cluster-mean representation would be a practically interesting way to reduce a gigapixel slide to a fixed-size vector while retaining classification performance, and the combination of clustering with permutation-invariant classifiers could be useful for low-resource WSI analysis. However, the paper currently does not establish the validity of its evaluation protocol: the train/test split and the scope of the K-means fit are not specified, no error bars or significance tests are given, and the only comparison is to an unpublished baseline [21] from the same group. The central computational claim is plausible, but the evidence presented is not yet sufficient to support it.","major_comments":[{"comment":"The evaluation protocol is not specified, and this is load-bearing for every number in Tables 1 and 2. The manuscript never states how the data are split into train, validation, and test sets, nor whether the K-means centroids and the elbow-method choice of k are computed on the training portion only. Section 3 says patch embeddings 'were then clustered using the K-means algorithm' and Section 4 says k=10 is set 'for the features of both datasets (TCGA and Camelyon-17)' based on the elbow plot. If the centroids and elbow curve are computed on all slides, including test slides, then each test slide's cluster-mean representation is a function of the test-set feature distribution, and the reported 0.83 and 0.75 accuracies are not valid held-out results. The authors must clarify and, if necessary, redo the experiments with clustering and k-selection performed inside each training fold, then frozen before encoding test slides.","section":"Section 4 and Figure 4"},{"comment":"The preprocessing rule that excludes patches with fewer than 10 nuclei is an untested modeling assumption that can systematically remove informative tissue. The manuscript justifies this by saying 'the Region of Interest is the tumor region,' but the Camelyon17 discussion in Section 5 explicitly concedes that 'regions of high cellularity may not always serve as reliable morphological biomarkers for metastasis detection.' Since the same threshold is applied to both datasets, and no ablation over the threshold is reported, it is unclear whether the reported accuracies are an artifact of discarding stroma, negative-tissue, and low-cellularity tumor patches. The authors should either justify the threshold with data or report an ablation across nucleus-count thresholds.","section":"Section 3 and Tables 1-2"},{"comment":"The quantitative claims are not supported by error bars, confidence intervals, or significance tests, and the baseline comparison is incomplete. The text states that the method gives results 'comparable with the classical approach used in [21],' but no numerical baseline results are reported, and [21] is an unpublished work that shares authors with this paper. The paper also does not compare to standard current WSI classifiers such as CLAM, TransMIL, or ABMIL on either dataset. Given the small dataset sizes (159 TCGA slides and an unspecified number of Camelyon17 slides), the reported differences between configurations could be within noise. The authors should provide split-level statistics and at least one non-self baseline comparison.","section":"Section 5, Tables 1-2"},{"comment":"The manuscript does not define what happens when a slide has no patches assigned to one or more of the k clusters. The concatenated cluster-mean vector is only permutation-invariant if the cluster order is fixed by a single set of centroids learned on a training set; if clustering is instead performed separately for each slide, the cluster indices have no semantic correspondence across slides and the representation is not well-defined. The authors should state explicitly whether K-means is fit globally, per slide, or per dataset, and how empty clusters are handled.","section":"Section 3 and Figure 4"}],"minor_comments":[{"comment":"The abstract contains repeated and informal phrasing (e.g., 'our novel approach' appears twice, and the long final sentence is difficult to parse). The authors should tighten the abstract and avoid marketing-style language.","section":"Abstract"},{"comment":"Reference [21] is described only as '2023' with no arXiv identifier, DOI, or venue. Since it is used as the main baseline and appears to be same-group work, the authors should provide a full citation and disclose the relationship.","section":"Reference [21]"},{"comment":"In the RegNet rows, the classifier is cited as 'Swin-Transformer [11]' but [11] is ConvNeXT; it should be [12]. Also the recall value '0.6' should be '0.60' for consistency with the other entries.","section":"Table 1"},{"comment":"The text says 'we are experimenting with fewer WSIs of the Camelyon17 dataset' but does not state how many slides were used or how the 500-slide dataset was subsampled. The exact number of slides and the split must be reported.","section":"Section 4.0.2"},{"comment":"The description of Swin-Tiny as a classifier on cluster-mean vectors is unclear: Swin-Transformer is designed for grid-structured image inputs, and the paper does not explain how the concatenated cluster-mean vector is reshaped or tokenized for Swin. Please clarify the input format.","section":"Section 3"}],"recommendation":"major_revision","confidential_remarks":"The paper's main baseline [21] is an unpublished prior work by the same group, and this is not disclosed as a self-comparison. Given the small dataset sizes and the absence of any held-out evaluation description, the reported accuracy numbers should not be taken at face value until the authors specify and re-run the evaluation with clustering confined to training folds. A careful revision with proper cross-validation and standard baselines is needed before this is suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this paper has a genuinely useful idea buried under an inflated abstract, and the most important experimental detail is missing. The authors propose clustering patch embeddings from a whole slide into 10 cluster means, then feeding those means to an attention MIL or transformer classifier. If that works, a gigapixel slide becomes a 10-vector representation, which would be a real efficiency win. They run a broad ablation across six encoders and three classifiers on two public datasets, and report accuracy numbers around 0.83 for TCGA Lung EGFR and 0.75 for Camelyon17. The ablation itself is the main strength; they also honestly note that high cellularity is not a reliable biomarker for metastasis in their Camelyon17 discussion, which contradicts their own nucleus-count threshold.\n\nThe soft spot is not the threshold, it's the clustering protocol. The paper says the features are clustered with K-means and k=10 is chosen by the elbow method 'for our dataset' — but it never says whether the centroids and the elbow curve are computed on a training split only. If the clustering is fit on all slides including test slides, then the test representations encode unlabeled test information, and the reported accuracies are not a held-out evaluation. If clustering is done per slide, the cluster indices have no correspondence across slides, so concatenating the cluster means into a single vector is not a well-defined permutation-invariant representation. Either reading makes Tables 1 and 2 uninterpretable without a precise statement of when and how clustering is fit.\n\nThe rest of the evaluation is thin: no train/validation/test split details, no error bars or significance tests, and no comparison to standard WSI baselines like CLAM or TransMIL. The only baseline is [21], which shares authors with this paper. The abstract overclaims ('breakthrough', 'significant advancement') for what is an incremental efficiency trick with modest accuracy gains.\n\nWho is this for? Someone working on compact WSI representations might want to test the idea, but not rely on the numbers as published. The paper deserves a serious referee because the underlying question is valid and the ablation is substantial, but only with the condition that the authors clarify the clustering scope and provide a proper held-out evaluation with error bars and real baselines. If the transductive concern is confirmed, the headline results are not salvageable. As it stands, I would not cite the reported numbers.","headline":"The efficiency idea is worth testing, but the paper never states whether clustering is fit on test slides, which puts the headline accuracies in question.","tokens_in":9472,"tokens_out":3325,"would_cite":false,"duration_ms":28324,"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":"Clustering patch embeddings compresses a gigapixel whole-slide image into ten cluster-mean vectors while preserving slide-level classification accuracy on EGFR mutation and metastasis prediction.","keywords":["whole slide images","patch embeddings","K-means clustering","permutation invariance","multiple instance learning","digital pathology","EGFR mutation prediction","Camelyon17"],"falsifier":"Re-run the full pipeline on Camelyon17 with the nucleus-count threshold varied from zero (keeping all tissue patches) to, say, 30, and observe whether the reported 0.75 accuracy changes; a material shift would show the preprocessing filter, not clustering, is carrying the result.","tokens_in":8364,"feed_emoji":"🔬","tokens_out":9035,"duration_ms":77524,"temperature":0.7,"pith_summary":"The paper tries to establish that a gigapixel whole-slide image can be reduced to a compact set of cluster-mean embedding vectors without sacrificing slide-level classification accuracy. The recipe is to extract patch embeddings with a pretrained encoder, cluster them into ten groups with K-means, replace each cluster by its mean vector, and feed the concatenated means to a permutation-invariant classifier. On TCGA Lung EGFR mutation prediction, the best configuration reaches 0.83 accuracy, and on Camelyon17 metastasis detection it reaches 0.75, comparable to the baseline. If this holds, gigapixel slides become tractable for memory-limited analysis while retaining clinically relevant signal.","feed_headline":"Whole-slide images shrink to 10 cluster vectors, accuracy intact","feed_subtitle":"Clustered patch embeddings match baseline accuracy on EGFR and metastasis classification at a fraction of the memory.","key_machinery":"The load-bearing object is the K-means cluster-mean representation of a slide: every patch embedding is assigned to one of ten clusters, and each cluster is summarized by its mean vector, giving a single vector of ten times the embedding dimension. Attention-based multiple instance learning then treats the ten cluster means as a bag of instances and learns an attention weight per cluster, which both aggregates them permutation-invariantly and indicates which clusters drive the prediction. This combination of compact cluster summaries with learned attention over them is what carries the argument that a gigapixel image can be compressed without losing classification accuracy.","core_discovery":"The central claim is that a whole-slide image's diagnostic content can be captured by the geometry of its patch-embedding distribution rather than by the full set of patch features. After clustering all patch embeddings into ten clusters and replacing each cluster by its mean vector, the concatenated ten mean vectors form a single global representation of the slide. Feeding this representation to attention-based multiple instance learning, which learns weights over the cluster means, makes the pipeline permutation-invariant and yields accuracy comparable to the classical approach: 0.83 for EGFR mutation classification on TCGA Lung and 0.75 for metastasis detection on Camelyon17. The paper also reports an ablation showing that clustering is not always beneficial, and it concludes that no single encoder-classifier pair is universally optimal.","pith_inferences":["An implication the paper leaves implicit is that the cluster-then-aggregate recipe is not tied to pathology: any gigapixel or high-resolution imaging task with slide-level labels could use the same compression, as long as the feature distribution is clusterable.","A testable extension would be to replace the elbow-chosen value of ten clusters with learned or swept cluster counts and check whether accuracy is stable; the fixed K is a heuristic, not an optimized parameter.","Because cluster means are built without labels, the representation could be pretrained on unlabeled slides and then reused for multiple downstream slide-level tasks, which would help in low-annotation clinical settings."],"forward_implications":["A gigapixel slide can be stored and processed as a single vector of roughly ten times the embedding dimension, cutting memory requirements relative to keeping all patch features.","Slide-level labels alone suffice, since clustering is unsupervised and the classifier learns from slide labels through multiple instance learning.","Attention weights over cluster means give a readout of which tissue patterns drove a decision, which could guide pathologist review by pointing back to the clusters and patches of interest.","The optimal encoder-classifier combination differs by dataset, so deploying the method means choosing the pair empirically for the task rather than assuming one backbone is best."],"supporting_citations":[{"why":"ResNet-50 is the feature extractor whose clustered embeddings achieve the best reported accuracy on both datasets.","marker":"[8]"},{"why":"Attention MIL is the permutation-invariant classifier that operates on cluster-mean vectors and produces the top results.","marker":"[18]"},{"why":"SimCLR provides the self-supervised encoder whose clustered features reach 0.81 accuracy on TCGA Lung.","marker":"[5]"},{"why":"Swin Transformer is used both as an encoder and as a classifier baseline, and its ablation shows clustering can reduce accuracy.","marker":"[12]"},{"why":"It supplies the classical EGFR classification baseline that the cluster-based pipeline is claimed to match or outperform.","marker":"[21]"},{"why":"It is the source of the TCGA Lung dataset used for EGFR mutation classification.","marker":"[19]"},{"why":"It is the source of the Camelyon17 dataset used for metastasis detection.","marker":"[20]"},{"why":"It supplies the clustering-based WSI representation idea that the method adapts, supporting the choice of cluster centroids as slide descriptors.","marker":"[13]"},{"why":"It provides the tissue-detection and nucleus-counting routines used to select high-cellularity patches in preprocessing.","marker":"[34]"},{"why":"It is the elbow-method reference used to justify the fixed choice of ten clusters.","marker":"[37]"}],"fun_headline_variants":["Ten cluster vectors capture whole-slide diagnostic info","Whole-slide images compressed to 10 vectors, accuracy holds","Cluster means replace thousands of patches, match baseline accuracy","Permutation-invariant WSI classification via 10 cluster embeddings","Slide-level diagnosis from 10 cluster means, not thousands of patches"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes that the diagnostically relevant regions are exactly the high-cellularity patches, since it discards every patch with fewer than ten detected nuclei before clustering.","fun_headline_variants_meta":{"raw":{"variants":["Ten cluster vectors capture whole-slide diagnostic info","Whole-slide images compressed to 10 vectors, accuracy holds","Cluster means replace thousands of patches, match baseline accuracy","Permutation-invariant WSI classification via 10 cluster embeddings","Slide-level diagnosis from 10 cluster means, not thousands of patches"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001083,"raw_usage":{"total_tokens":4483,"prompt_tokens":853,"completion_tokens":3630,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":3548}},"tokens_in":469,"tokens_out":3630,"duration_ms":25124,"temperature":1.0,"reasoning_tokens":3548,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T21:31:27.084996+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the full pipeline on Camelyon17 with the nucleus-count threshold varied from zero (keeping all tissue patches) to, say, 30, and observe whether the reported 0.75 accuracy changes; a material shift would show the preprocessing filter, not clustering, is carrying the result.","supporting_citations":[{"cited_title":"Cluster-to-conquer: A framework for end-to-end multi-instance learning for whole slide image classification","cited_arxiv_id":null,"evidence_quote":"It supplies the clustering-based WSI representation idea that the method adapts, supporting the choice of cluster centroids as slide descriptors."},{"cited_title":"Integration k-means clustering method and elbow method for identification of the best customer profile cluster","cited_arxiv_id":null,"evidence_quote":"It is the elbow-method reference used to justify the fixed choice of ten clusters."}],"review_version":1}