{"id":"656deb6f-6103-4f21-b023-fa1a76608ddc","arxiv_id":"2501.16787","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DyHG, a dynamic hypergraph MIL model with low-rank incidence and Gumbel-Softmax sampling, reports improved whole slide image classification for bone metastasis origin and subtype.","lead":"This paper proposes a dynamic hypergraph neural network for classifying bone metastasis whole slide images, using learnable hyperedges that connect many tissue patches at once. It reports accuracy gains over standard MIL and graph baselines on private bone metastasis datasets and on two public benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Primary evidence for DyHG's SOTA claim depends on an internal dataset split per WSI rather than per patient; without patient-level partitioning, reported gains may reflect slide leakage, and public CAMELYON+ results do not show consistent superiority.","rationale":"The reader's weakest assumption is exactly the concern I find most load-bearing: the internal dataset is split per WSI, not per patient, and no one-slide-per-patient guarantee is given. This threatens the validity of the primary evidence. The paper's central claim is empirical, and Table I is the only place where DyHG consistently leads; the public CAMELYON+ results are mixed, which is a secondary red flag. No technical flaw in the method itself appears fatal: the hypergraph construction is coherent (modulo notation), the ablation studies are consistent with the design choices, and the use of a frozen UNI extractor is shared across baselines. Because the concern is addressable with a standard patient-level split and significance testing, the CONDITIONAL verdict remains appropriate; no change to the reader's verdict is needed.","tokens_in":17478,"tokens_out":7439,"duration_ms":67090,"concrete_test":"Request de-identified patient identifiers for the internal bone metastasis dataset and re-run the full DyHG vs. baseline comparison using a patient-level stratified 5:2:3 split, ensuring all WSIs from one patient are in the same fold. Recompute Table I under this split. If DyHG's accuracy or balanced-accuracy advantage over the best baseline falls below the originally reported margin or reverses, the headline claim fails. As a complementary check, perform a paired bootstrap or McNemar test comparing DyHG with the best baseline on the patient-level test fold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is the evaluation protocol for the internal bone metastasis datasets. Section IV-C states only that 'for each category of the two bone metastasis datasets, we divide the data into training set, validation set, and test set in a ratio of 5:2:3.' There is no mention of patient-level partitioning, no statement that each patient contributes exactly one WSI, and no patient identifier is reported. If multiple slides come from the same patient, random per-WSI splitting places slides from one patient in both training and test folds, letting the model exploit patient-specific staining, sectioning, or scanner artifacts instead of generalizable tumor morphology. The claimed gains in Table I are small (0.16–1.28 percentage points over the best baseline) and the standard deviations overlap (e.g., DyHG 86.32±0.96 vs. CLAMMB 85.04±1.01), so leakage could plausibly account for the entire advantage. This concern is load-bearing because the central claim of significant SOTA outperformance rests almost entirely on Table I. The public results in Table II do not support a blanket superiority claim: on CAMELYON+, DyHG is not the best in accuracy (87.42 vs. Hyper-AdaC 87.67) or balanced accuracy (66.82 vs. PatchGCN 66.99). Thus, unless the internal data are split by patient, the evidence does not establish that DyHG significantly outperforms SOTA baselines.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DyHG, a dynamic hypergraph neural network for whole-slide image (WSI) classification, and applies it to primary bone cancer origin and subtyping classification. The method constructs a hypergraph incidence matrix by applying a low-rank linear transformation to patch embeddings and then refines the assignment with a Gumbel-Softmax sampling strategy; node and hyperedge aggregation follow, and an attention-based MIL pooling produces the slide-level prediction. The manuscript reports experiments on two internal bone metastasis datasets and two public benchmarks (CAMELYON+ and PANDA), compares against embedding-based MIL and graph/hypergraph baselines, and includes ablations, hyperparameter studies, time-efficiency measurements, and attention heatmaps.","tokens_in":17753,"tokens_out":4824,"duration_ms":43682,"significance":"If the results are validated, DyHG would be a useful contribution to WSI analysis by showing that an end-to-end learnable hypergraph construction can capture high-order patch interactions more effectively than static hypergraph baselines. The paper provides explicit equations for the method, a relatively comprehensive set of baselines, and ablation studies that isolate the contributions of Gumbel noise and temperature. The main weakness is that the experimental evidence for the central 'significantly outperforms SOTA' claim is not fully supported: the internal evaluation lacks a patient-level split and significance tests, and the public results do not show consistent superiority across all metrics. The method itself is plausible and well-motivated, but the strength of the empirical claim needs to be scaled back or backed by additional experiments.","major_comments":[{"comment":"The internal datasets are split per category in a 5:2:3 ratio, but the paper does not state whether the split is performed at the patient level or whether each patient contributes exactly one WSI. Since the data are collected from a hospital cohort, multiple slides per patient are plausible; a random per-WSI split could place slides from the same patient in both training and test sets, allowing the model to exploit patient-specific staining, sectioning, or scanner artifacts rather than generalizable tumor morphology. The claimed improvements in Table I are small (0.16–1.28 percentage points over the best baseline) with overlapping standard deviations (e.g., DyHG 86.32±0.96 vs. CLAMMB 85.04±1.01), so this leak could plausibly account for the reported advantage. The authors must either demonstrate that each patient contributes exactly one WSI or re-run the evaluation under a patient-level split and report the patient counts per split.","section":"Section IV-C, Section V-A, Table I"},{"comment":"No statistical significance testing is reported for any comparison. The abstract claims that DyHG 'significantly outperforms state-of-the-art (SOTA) baselines', but Tables I and II only provide means and standard deviations over five seeds. For example, in Table I the accuracy of DyHG (86.32±0.96) overlaps with that of CLAMMB (85.04±1.01), and in Table II the balanced accuracy of DyHG on CAMELYON+ (66.82±2.26) overlaps with PatchGCN (66.99±4.42). The authors should provide paired bootstrap or permutation tests, or confidence intervals, to justify the word 'significantly'.","section":"Section V, Tables I-II"},{"comment":"The public benchmark results do not support the blanket SOTA claim. On CAMELYON+, DyHG achieves lower accuracy (87.42) than Hyper-AdaC (87.67) and lower balanced accuracy (66.82) than PatchGCN (66.99). The paper's observation (1) in Section V.F highlights only specificity and weighted F1, where DyHG does well, and does not acknowledge the metrics where it fails to lead. The central claim should be qualified by dataset and metric, or the paper should offer a concrete explanation for why these two deficits do not undermine the general superiority claim.","section":"Section V.F, Table II"},{"comment":"The bHGFN baseline is described as 'a factorization neural network' but the cited original paper (Di et al., IEEE TIP 2022) is designed for survival prediction, not slide-level classification. The manuscript does not explain how bHGFN was adapted to the classification tasks, despite stating only 'For all baselines, we perform experiments in the same settings.' Without a description of the adaptation (e.g., output head, normalization, training objective), the very low bHGFN accuracy in Table I (73.61 and 87.52) could reflect task mismatch rather than architectural inferiority, making the comparison unfair and unreproducible.","section":"Section IV.B, Table I"}],"minor_comments":[{"comment":"Equation (1) defines H = ReLU(XW1), but the text immediately after refers to 'The resulting H0' and later uses H for both the number of hyperedges and the incidence matrix. Please unify the notation (e.g., use H for the hyperedge count and A or M for the incidence matrix).","section":"Section III-A, Eq. (1)"},{"comment":"The description of the w/o S variant states that the heatmap shows 'invalid weights exceeding 1 or less than 0'. Since H is produced by ReLU in Eq. (1), negative weights are not possible; the actual issue is that the weights are not normalized and can exceed 1. Please correct this wording to avoid confusion.","section":"Section V.C, Fig. 9"},{"comment":"The hyperparameter values (H and τ) used for DyHG on CAMELYON+ and PANDA are not reported. The paper gives H and τ for the two internal tasks but not for the public datasets, though the text says the public experiments were run in 'identical settings'. Please specify these values for reproducibility.","section":"Section IV.C, Table II"},{"comment":"There are typos in the figure labels: 'Aggregration' should be 'Aggregation' and 'Nwtwork' should be 'Network'. Please correct these.","section":"Figure 2"},{"comment":"The footnote contains the misspellings 'Hostpital' and 'Univeristy' in the institutional affiliations. Please proofread the front matter.","section":"Section I"},{"comment":"Observation (2) states that DyHG's standard deviation is 'significantly lower' than other methods, but no significance test is performed on the variances. Please soften this wording or provide a proper test.","section":"Section V.A"}],"recommendation":"major_revision","confidential_remarks":"The main barrier to acceptance is the lack of a patient-level split for the internal datasets; given that the data come from a single hospital, the risk of patient-level leakage is high and could undermine the central empirical claim. The public-benchmark results should be reported with a more nuanced interpretation, and significance testing should be added. I would also note that several baselines (WiKG, the CAMELYON+ benchmark) are from the same research group; that is not a problem per se, but the comparisons should be scrutinized for consistency."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the dynamic hypergraph construction idea is worth a look, but the evaluation as written doesn't establish the SOTA claim. The core novelty—learning a low-rank incidence matrix and refining it with Gumbel-Softmax—is clean, and the ablation shows the sampling adds value. That part is solid.\n\nThe method is clearly described. Using a pretrained feature extractor, low-rank projection to hyperedge logits, then differentiable assignment, hypergraph convolution, attention pooling. The time-efficiency comparison against K-NN-based construction is plausible. The internal datasets are new and clinically relevant, and the attention heatmaps align with pathologist ROIs.\n\nThe soft spots are real and load-bearing. The internal split is per WSI, not per patient. Section IV-C says only that data are divided 5:2:3 per category. If multiple slides come from one patient, training and test sets share patient-specific staining or scanner artifacts, and the reported gains—most under a point, with overlapping standard deviations—could be leakage. The paper never states each patient contributes exactly one slide. That has to be fixed before any performance claim is taken seriously.\n\nSecond, the headline 'significantly outperforms SOTA' is undercut by the authors' own public-data results. On CAMELYON+, DyHG is behind Hyper-AdaC in accuracy and behind PatchGCN in balanced accuracy. On PANDA it is best but by small margins. With no significance tests, calling these results 'significant' is unsupported. A paired test or confidence interval would help.\n\nThird, no code or data are released. For a method paper, that's not fatal, but it makes the empirical claims harder to verify.\n\nThe notation could also use a pass: the incidence matrix is denoted H and the number of hyperedges is also H, which confused me more than once.\n\nWho is this for? Computational pathology researchers working on WSI-level representation, especially hypergraph methods. It deserves a serious referee, but the verdict should be conditional: require patient-level splitting, significance testing, and either code/data or softened claims.\n\nI'd treat it as a credible method paper, not a definitive result.","headline":"A plausible dynamic hypergraph construction method, but the empirical claims are undercut by a possible patient-leakage split and unsupported 'significant' wording.","tokens_in":18328,"tokens_out":2862,"would_cite":true,"duration_ms":25425,"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":"The paper claims that a dynamic, learnable hypergraph representation of whole-slide images improves classification of primary bone cancer origin and subtype compared with attention-based MIL and static graph/hypergraph baselines.","keywords":["bone metastasis","whole slide image","hypergraph neural network","dynamic hypergraph construction","multiple instance learning","Gumbel-Softmax","low-rank approximation","weakly supervised classification"],"falsifier":"Hold out entire patients—all of their slides in one fold—and rerun the two bone-metastasis tasks; if DyHG's accuracy edge over CLAMMB and CLAMSB disappears or reverses, the dynamic hypergraph's reported advantage is not generalizable beyond patient-specific artifacts.","tokens_in":17253,"feed_emoji":"🦴","tokens_out":6463,"duration_ms":52956,"temperature":0.7,"pith_summary":"This paper tries to establish that a whole-slide image is better modeled as a hypergraph whose structure is learned during training than as a bag of independent patches or a graph with pairwise edges. On bone-metastasis slides, the proposed DyHG predicts the primary cancer origin and the tumor subtype from slide-level labels alone, and the paper reports that it outperforms attention-based MIL and static graph/hypergraph baselines on accuracy, balanced accuracy, specificity, and weighted F1. The method replaces hand-built K-NN or K-means hypergraphs with a low-rank learnable incidence matrix refined by Gumbel-Softmax sampling, so the hypergraph can be optimized end-to-end and scaled to slides with tens of thousands of patches. If this holds, it gives pathologists a weakly supervised tool that ties together distant tumor regions without per-patch annotations.","feed_headline":"Hypergraph model tops baselines on bone-metastasis slides","feed_subtitle":"The dynamic hypergraph learns patch groupings end-to-end and beats static-graph rivals on origin and subtyping tasks.","key_machinery":"The central object is the learnable hypergraph incidence matrix $\\mathbf{H} \\in \\mathbb{R}^{N \\times H}$; a hyperedge is a subset that can connect more than two patches at once, so a single hyperedge can directly tie distant patches that share biological relevance. DyHG obtains $\\mathbf{H}$ from patch embeddings $\\mathbf{X}$ via a low-rank projection $\\mathbf{W}_1$, reducing parameters from $N \\times H$ to $d \\times H$, and then applies Gumbel-Softmax per patch to produce a differentiable soft assignment across hyperedges. This incidence matrix drives the hypergraph convolution—node aggregation $\\mathbf{H}^\\top \\mathbf{X}$ followed by hyperedge aggregation $\\mathbf{H}\\mathbf{E}$—so all downstream messages depend on learned groupings, and the whole pipeline trains end-to-end from slide labels.","core_discovery":"The central claim is that representing a WSI by a dynamic hypergraph—where each hyperedge can connect any number of patches and the assignment of patches to hyperedges is learned, not fixed by coordinates or clustering—captures high-order biological interactions that ordinary MIL and pairwise GNNs miss. The paper constructs the incidence matrix as $\\mathbf{H} = \\mathrm{ReLU}(\\mathbf{X}\\mathbf{W}_1)$, a low-rank map from patch embeddings to hyperedge logits, and makes the discrete assignment differentiable with Gumbel-Softmax sampling. A two-step hypergraph convolution aggregates node features into hyperedges and then back into nodes, and a global attention pool produces the slide-level embedding for classification. On the two internal bone-metastasis datasets DyHG reports the highest scores in all four metrics on both tasks, and on CAMELYON+ and PANDA it reports competitive or best results, with lower standard deviation than most baselines.","pith_inferences":["The dynamic incidence matrix could be inserted as a plug-in representation layer before any MIL aggregator, not only the attention pool used here; the paper does not test that modularity.","A direct patient-level split would separate the contribution of the learned hypergraph from any patient-specific staining or section artifacts; the paper's random slide-level split does not rule those out.","The learned hyperedges could be inspected as candidate tumor-microenvironment motifs: if the same hyperedge repeatedly collects patches from separate metastatic foci, its patch contents could serve as a data-driven dictionary for pathologists.","The ablation suggests the low-rank structure alone carries much of the benefit, since removing sampling entirely still beats removing only Gumbel noise; a follow-up could isolate exactly which part of the sampling term adds signal."],"forward_implications":["Learning the hypergraph structure from patch embeddings removes the need to precompute K-NN or K-means hypergraphs, so WSI classification can train end-to-end without costly offline clustering.","Because the construction time is roughly flat in the number of patches, the method scales to slides with more than 30,000 patches where clustering-based hypergraph construction grows exponentially.","If the reported gains are real, dynamic hypergraph aggregation is a stronger inductive bias than pairwise graphs for tasks where tumor regions are sparse and dispersed across the slide.","The same architecture transfers to other weakly supervised pathology tasks: the paper reports best or competitive results on CAMELYON+ and PANDA without task-specific changes."],"supporting_citations":[{"why":"Supplies the low-rank strategy that reduces the parameter cost of learning the hypergraph incidence matrix.","marker":"[29]"},{"why":"Supplies the Gumbel-Softmax estimator used to make discrete hyperedge assignments differentiable.","marker":"[30]"},{"why":"Supplies Otsu thresholding, used to identify valid tissue areas before patch cropping.","marker":"[57]"},{"why":"Supplies the frozen UNI feature extractor that produces the initial patch embeddings used by all methods compared.","marker":"[58]"},{"why":"Defines the attention-based MIL baseline and the attention aggregation mechanism DyHG adapts for slide-level pooling.","marker":"[16]"},{"why":"Provides the CLAM baselines and the clustering-constrained attention formulation that DyHG is compared against.","marker":"[17]"},{"why":"Provides the bHGFN hypergraph baseline whose random-sampling construction DyHG is designed to outperform.","marker":"[54]"},{"why":"Provides the Hyper-AdaC adaptive clustering hypergraph baseline, the strongest static hypergraph comparator.","marker":"[55]"},{"why":"Supplies the CAMELYON+ public benchmark used to test generalizability.","marker":"[61]"},{"why":"Supplies the PANDA public benchmark used to test generalizability on slides with few patches.","marker":"[62]"}],"fun_headline_variants":["Dynamic hypergraph learns high-order interactions in bone metastasis slides","Bone metastasis analysis gets a dynamic hypergraph boost","Hypergraph learns patch groupings to top bone cancer origin and subtyping","Dynamic hypergraph captures complex biological interactions for slide analysis","Beyond pairwise GNNs: dynamic hypergraph for bone metastasis slides"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains assume that randomly splitting slides into training and test sets does not leak patient-specific information when several slides come from the same patient.","fun_headline_variants_meta":{"raw":{"variants":["Dynamic hypergraph learns high-order interactions in bone metastasis slides","Bone metastasis analysis gets a dynamic hypergraph boost","Hypergraph learns patch groupings to top bone cancer origin and subtyping","Dynamic hypergraph captures complex biological interactions for slide analysis","Beyond pairwise GNNs: dynamic hypergraph for bone metastasis slides"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000712,"raw_usage":{"total_tokens":3231,"prompt_tokens":1002,"completion_tokens":2229,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":2145}},"tokens_in":618,"tokens_out":2229,"duration_ms":14307,"temperature":1.0,"reasoning_tokens":2145,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T10:43:15.872136+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out entire patients—all of their slides in one fold—and rerun the two bone-metastasis tasks; if DyHG's accuracy edge over CLAMMB and CLAMSB disappears or reverses, the dynamic hypergraph's reported advantage is not generalizable beyond patient-specific artifacts.","supporting_citations":[],"review_version":1}