{"id":"84ddc4f8-1ef8-47bd-aab4-670738f981d5","arxiv_id":"2412.16715","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A new 5-billion-cell dataset and a hierarchical cell-cloud transformer match or beat patch-based models for survival and staging on several TCGA cancers.","lead":"The authors built WSI-Cell5B, a dataset of over 5 billion cell-level annotations from 6,998 cancer whole-slide images, and CCFormer, a model that treats each slide's cells as a point cloud to predict survival and cancer stage. The work suggests that cell spatial patterns alone can rival patch-based methods on several cancer types, potentially making WSI analysis cheaper and more interpretable.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cell-level annotation accuracy is the load-bearing assumption: only patch-level tumor/normal is validated, so the claimed SOTA from cell spatial distribution may reflect detector noise rather than true cellular biology.","rationale":"The paper makes a strong contribution: a large-scale cell-annotated WSI dataset and a hierarchical transformer for cell clouds, with extensive experiments across 11 cancer types and honest discussion of failure cases (e.g., KIRC/UCEC, Appendix A.4). The central claim, however, is that cell spatial distribution alone reaches SOTA. That claim is only as strong as the cell-level annotations. The authors validate at patch level (92.7% tumor/normal agreement), but patch-level agreement can mask per-cell type errors, particularly because WSLR (Sec. 3.1) is a self-training loop on pseudo-labels filtered by foundation-model patch votes. If cell-type labels are biased by tissue context, the NIE features and HSP grouping learn those biases, and the reported C-Index/F1 gains may not reflect true cell spatial biology. This is the single most load-bearing assumption: if it fails, both the dataset's utility and the method's conceptual claim collapse. The proposed test—expert per-cell validation on a sample—would settle the matter. Since the concern is concrete and addressable, the reader's CONDITIONAL verdict is appropriate; no verdict change is needed.","tokens_in":19125,"tokens_out":6216,"duration_ms":54172,"concrete_test":"Have two pathologists independently annotate all cells (type and centroid) in a random sample of 50–100 512×512 patches spanning the 11 TCGA cancer types (or align existing cell-level benchmarks such as PanNuke/Lizard to the same tissue). Compute per-cell type F1, confusion matrix, and coordinate error (e.g., detection within 5 pixels) between WSI-Cell5B and the expert annotations. If per-cell type F1 is below ~0.8, or if misclassifications are spatially correlated with tumor/normal regions (e.g., inflammatory cells in tumor stroma preferentially relabeled as neoplastic), the downstream SOTA results cannot be attributed to true cell spatial distribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that learning from cell spatial distribution alone achieves SOTA rests on the premise that WSI-Cell5B's per-cell types and coordinates are trustworthy. The paper validates annotations only at patch level: Appendix A.2 compares a patch-level tumor/normal label derived from a 25% neoplastic-cell threshold against pathologist votes, reporting 92.7% agreement. This is compatible with systematically wrong cell-level labels: a detector that labels all cells in tumor patches as neoplastic and all cells in normal patches as 'other' would achieve similar patch accuracy while misclassifying nearly every inflammatory cell. The WSLR procedure (Section 3.1) amplifies this risk: credible patches are selected by agreement between foundation-model tumor/normal votes and the detector's own aggregate cell assessment, and the detector is then fine-tuned on its own pseudo-labels. That loop can reinforce existing biases rather than correct them, and there is no per-cell ground-truth check. Because NIE (Section 4.1) and HSP (Section 4.2) consume these labels as counts and one-hot encodings, any systematic cell-type bias (e.g., tumor-stroma confounding) would be learned as if it were spatial biology. The survival/staging results in Table 1 and Figure 6 could then be achieved by a model that is effectively reading patch-level tumor proportion through a cell-detection proxy, not by modeling true cell spatial distribution. This directly threatens the paper's central contribution, not merely its quantitative claims.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper constructs WSI-Cell5B, a large-scale dataset of 6,998 TCGA WSIs from 11 cancer types with over 5 billion cell annotations (coordinates and three coarse cell types), produced by a PanNuke-pretrained detector refined through a weakly supervised label refinement procedure using foundation models. It also proposes CCFormer, a hierarchical cell-cloud transformer: NIE embeds local and global neighborhood density features per cell, and HSP groups cells by farthest-point sampling and applies semantic-spatial filtering and vector attention in a bottom-up hierarchy. The model is evaluated for survival prediction (C-Index, 5-fold CV, 10 cancer types) and cancer staging (Macro-F1, 2 cancer types) against MIL, graph, and point-cloud baselines, and the dataset is used to construct CPS/MCPS survival-risk metrics with Kaplan-Meier analyses.","tokens_in":19432,"tokens_out":9814,"duration_ms":71687,"significance":"The scale of the proposed dataset is a genuine contribution: to the best of my knowledge, it is the first WSI-level dataset with cell-level annotations linked to clinical endpoints, covering 6,998 WSIs and more than 5 billion cells. The CCFormer architecture is a reasonable adaptation of point-cloud learning to histopathology, and the use of a consistent 5-fold CV protocol across all baselines is a strength. If the cell annotations are reliable, the resource would enable many downstream analyses. However, the paper's headline claim that cell spatial distribution alone achieves SOTA is not supported by the reported numbers, and the lack of direct per-cell validation weakens confidence in the biological interpretation of the results.","major_comments":[{"comment":"The abstract states that 'learning from cell spatial distribution alone can already achieve state-of-the-art (SOTA) performance, i.e., CCFormer strongly outperforms other competing methods.' Table 1 does not support this. CCFormer alone achieves the highest C-Index in 6 of 10 cancer types (BLCA, BRCA, COADREAD, LUAD, PAAD, STAD), but on KIRC ABMIL is substantially better (0.712 vs 0.658), on UCEC ABMIL is better (0.735 vs 0.693), on LUSC WiKG is better (0.635 vs 0.633), and on HNSC PointNet is marginally better (0.650 vs 0.649). Moreover, the combined model CCFormer+MeanPool(Patch) beats CCFormer alone on all cancer types except LUSC, so the best results generally require appearance features in addition to cell clouds. The text in §5.2 and Appendix A.4 itself concedes inferior performance on some cancer types. Please revise the abstract and conclusion to reflect the actual pattern, and report whether the C-Index differences are statistically significant rather than relying on fold means whose standard deviations overlap.","section":"Abstract and §5.2, Table 1"},{"comment":"The central claim that CCFormer succeeds by modeling cell spatial distribution presupposes that the cell coordinates and type labels in WSI-Cell5B are accurate. The only quantitative validation is patch-level: Appendix A.2 reports 92.7% agreement between patch-level tumor/normal labels (derived from a 25% neoplastic-cell threshold) and pathologist votes. A detector that systematically mislabels cell types within tumor and normal patches could still achieve this patch-level agreement. The WSLR procedure (Section 3.1) selects credible patches by agreement between foundation-model votes and the detector's own aggregate cell assessment, then fine-tunes the detector on those patches; this self-training loop can reinforce existing biases rather than correct them. Since NIE (Section 4.1) and HSP (Section 4.2) consume the cell-type counts and one-hot encodings, systematic cell-type errors would be learned as spatial biology. To support the paper's conclusions, please provide per-cell validation (e.g., on a small set of manually annotated regions) with per-class precision/recall and coordinate error, and analyze the effect of label noise on the key survival/staging results.","section":"Section 3.1 and Appendix A.2"},{"comment":"The CPS/MCPS results are presented as evidence that WSI-Cell5B can directly yield clinical indicators. However, the weight vector alpha is chosen per cancer type after the fact (A.5.2), the MCPS boxes are randomly sampled without reporting the distribution of p-values across random seeds or box draws, and no multiple-testing correction or held-out validation is applied. The reported p-value improvements (e.g., HNSC from 8.04e-2 to 1.28e-2) are therefore descriptive rather than confirmatory. Please either pre-specify the metric construction, report repeatability/variance of MCPS, or recast these analyses as exploratory.","section":"Section 3.2 and Appendix A.5.2"},{"comment":"The survival comparisons report mean C-Index over 5 folds with standard deviations, but no paired significance tests are reported. Many of the claimed advantages are within one standard deviation of the best baseline (e.g., HNSC: 0.649±0.052 vs PointNet 0.650±0.032; LUAD: 0.657±0.012 vs PointNet++ 0.645±0.020). In addition, the method has several tuned hyperparameters (lambda_r, N_d, lambda_sim, N_k, L, beta), and A.5.3 states that lambda_sim, learning rate, and dropout are adjusted per cancer; no selection protocol or sensitivity analysis is reported. This makes it difficult to assess whether the reported gains reflect the proposed architecture rather than per-dataset tuning. Please add significance testing and a sensitivity analysis for the key hyperparameters.","section":"Section 5.1, Table 1, and A.5.3"}],"minor_comments":[{"comment":"Table 3 has a typo in the header 'C-Idex' (should be 'C-Index'), and Algorithm 1 begins with 'IUPUT' (should be 'INPUT').","section":"Table 3 and Algorithm 1"},{"comment":"In the second paragraph, 'SPC fails to distinguish patients of HNSC' should read 'CPS fails...'.","section":"Section 3.2"},{"comment":"The alpha values are formatted inconsistently (e.g., [0.0, 0.0, 1.0] vs [0.33, 0.33, 0.33]); please also state explicitly how these values were selected and whether the selection was made before or after seeing the survival data.","section":"Appendix A.5.2"},{"comment":"The paper does not state where WSI-Cell5B will be made available. For a dataset contribution, please include an availability statement with a URL or a clear release plan.","section":"General"},{"comment":"There is a duplicated word in 'spatial spatial distribution' in the description of HSP; please fix the typo.","section":"Section 4.2"},{"comment":"The combination weight beta is defined for CCFormer+MeanPool(Patch), but no value or selection procedure for beta is given.","section":"Appendix A.5.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript describes a potentially valuable resource, but the gap between the reported validation and the central claims is substantial. The dataset release should be a condition of acceptance: without public access to WSI-Cell5B and the annotation pipeline, neither the cell-level annotation quality nor the downstream results can be independently verified. My recommendation is major revision rather than rejection because the weaknesses are addressable: per-cell validation, revised claims, and significance testing can be added within the scope of the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the actual contribution is a 5-billion-cell dataset over 6,998 TCGA WSIs plus a point-cloud transformer (CCFormer) that treats cells as a 2D point set. That is a real, large, mostly new resource, and the architecture (NIE, HSP) is a reasonable way to consume cell coordinates and types. Second, the headline claim in the abstract — that CCFormer 'strongly outperforms' everything from cell spatial distribution alone — is not supported. Looking at Table 1, CCFormer is best on only 6 of 10 cancers, and on KIRC, UCEC, and LUAD it is behind the best patch-based baselines. The strongest results come from combining CCFormer with patch appearance features. The main text is more honest about this, but the abstract overstates it.\n\nWhat's good: the scale of WSI-Cell5B is impressive; the WSLR pipeline for generating pseudo-labels is a practical approach to a real cost problem; and the idea of cell-cloud modeling, while anticipated by Patch-GCN and Ceograph, is taken to full-WSI scale here. The survival risk metrics (CPS/MCPS) are a nice demonstration that cell counts alone carry signal, though the alpha weights look hand-tuned per cancer and the MCPS random boxes have no variance reporting.\n\nThe soft spot is the one the stress test flags: the cell-level labels are never validated directly. Appendix A.2 shows 92.7% patch-level agreement on tumor/normal, but that test is compatible with systematically wrong cell-type labels (e.g., labeling all cells in tumor patches as neoplastic). The WSLR self-training loop — pick credible patches by agreement, fine-tune on those pseudo-labels — can reinforce detector biases, and there is no per-cell ground truth to check. Since NIE and HSP consume these labels as counts and one-hot encodings, any bias becomes spatial biology. This is the main reason to withhold full confidence.\n\nThe paper does acknowledge some limitations in A.4 (e.g., KIRC nuclear grade, BLCA invasion), which is a point in its favor. But the missing per-cell validation, plus the unreleased dataset and code, keep this from being a clean acceptance.\n\nWho it's for: anyone working on WSI representation learning or spatial biology in pathology. It deserves a serious referee: the dataset scale and the architecture are substantial, and the flaws are addressable. My recommendation is to send it to peer review, but require the authors to either release the dataset with per-cell validation on a held-out set or substantially weaken the claims.","headline":"A serious cell-cloud dataset and architecture worth engaging, but the abstract overstates the results and the unvalidated cell-level labels are the load-bearing risk.","tokens_in":20037,"tokens_out":2004,"would_cite":true,"duration_ms":17255,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Cell coordinates and types alone can drive state-of-the-art survival and staging predictions from whole-slide images.","keywords":["whole slide image analysis","cell cloud","cell spatial distribution","survival prediction","cancer staging","cell-level annotations","hierarchical transformer","computational pathology"],"falsifier":"Have pathologists annotate every cell (or a dense grid of patches) on a random sample of, say, 50 whole-slide images from the dataset; if cell-type agreement with the pipeline is substantially below the reported 92.7% patch-level tumor/normal agreement, or if errors concentrate in specific cell types or tissue regions, the survival and staging results could be artifacts of annotation noise. A quicker check: train CCFormer on the dataset with cell-type labels randomly permuted within each slide while keeping coordinates fixed; if the C-Index does not drop toward chance, the model is exploiting label noise rather than true spatial composition.","tokens_in":18916,"feed_emoji":"🔬","tokens_out":5196,"duration_ms":43065,"temperature":0.7,"pith_summary":"This paper sets out to show that the spatial arrangement of cells alone—their coordinates and basic type labels—carries enough information to drive clinical predictions from histopathology slides. To test this, the authors assemble a dataset of more than five billion cell-level annotations across nearly 7,000 whole-slide images of 11 cancers, and pair it with a model (CCFormer) that treats each slide as a 'cell cloud' and learns hierarchical spatial structure. On survival prediction and cancer staging, the cell-cloud model reaches or beats patch-based methods on most cancer types, suggesting that cell spatial distribution is a viable, cheaper alternative to image-feature pipelines. The paper also shows that simple cell-counting scores stratify patients by survival risk. A sympathetic reader would take the core claim as: cell clouds are an effective slide representation, and the new dataset makes direct cell-level analysis possible.","feed_headline":"5 billion cell labels show layout predicts survival and stage","feed_subtitle":"A cell-cloud transformer beats patch-based methods on most of 11 cancers using only cell coordinates and types.","key_machinery":"Two mechanisms carry the argument. First, Neighboring Information Embedding (NIE) computes, for every cell, local and global density features: the fraction of each cell type appearing within successive radial shells around the cell, normalized locally and globally. This turns each cell's identity into a short vector describing its neighborhood mix, which can distinguish cells of the same type in different microenvironments. Second, Hierarchical Spatial Perception (HSP) repeatedly groups cells into sub-regions via farthest-point sampling, filters each group with a semantic-spatial similarity score, applies vector attention with relative coordinates to update features, and aggregates group-wise; repeating this level by level builds a slide-level representation from local clusters upward. The whole pipeline is trained end-to-end on survival and staging objectives.","core_discovery":"The central claim is that modeling the spatial distribution of cells per whole-slide image—rather than the pixel appearance of patches—can already achieve state-of-the-art performance on survival prediction and cancer staging. Concretely, CCFormer obtains the highest C-Index on 6 of 10 cancer types and competitive results on the rest, and large F1 gains on staging tasks; combining it with simple patch features improves results further. The underlying discovery is that a hierarchical treatment of cell clouds—encoding each cell's local neighborhood statistics and then aggregating spatially up the slide scale—captures clinically meaningful structure that patch-level and graph-of-patch methods miss. The authors further claim that clinical indicators computed purely by counting cell types within boxes (CPS and MCPS) separate high- from low-risk patients with small p-values, so the cell-level annotations themselves carry prognostic value.","pith_inferences":["The paper does not test whether the signal is genuinely semantic: if randomizing cell-type labels while keeping coordinates preserves most of the C-Index, the model would be reading density rather than biology.","The reported 92.7% agreement validates only patch-level tumor-versus-normal status; the finer three-type labels (neoplastic, inflammatory, other) that drive the spatial features remain unvalidated at cell level.","If annotation errors are spatially correlated—for example, concentrated in dense or necrotic regions—the clinical associations could be inflated; synthetic perturbation of coordinates and labels could bound this effect.","Future work could extend the hierarchy to fine-grained cell subclasses where the paper itself notes coarse types limit performance on kidney and bladder cancers."],"forward_implications":["If cell clouds alone suffice for state-of-the-art performance on most cancers, heavy patch-level feature extraction may be bypassed for those downstream tasks.","Combining cell-cloud features with a simple patch mean-pool improves C-Index on all evaluated cancer types except where patch features are weak, implying the two signals are complementary.","Cell-counting scores derived from the dataset can serve as interpretable, model-free clinical indicators for survival stratification.","The dataset enables evaluating cell spatial distributions across entire slides, which existing patch-level cell datasets cannot support."],"supporting_citations":[{"why":"Supplies the pretrained cell detection and classification model used to produce preliminary annotations.","marker":"[48]"},{"why":"Provides the nuclei-annotation dataset on which the cell detector is pretrained.","marker":"[13]"},{"why":"Sources the whole-slide images and clinical endpoints used to build the dataset.","marker":"[17]"},{"why":"One of the foundation models whose patch-level tumor/normal votes drive the weakly supervised refinement.","marker":"[32]"},{"why":"Another foundation model used in the voting ensemble for label refinement.","marker":"[19]"},{"why":"Graph-of-patches survival baseline that also frames whole-slide images as point sets; CCFormer is compared against it.","marker":"[4]"},{"why":"Standard attention-based multiple-instance-learning baseline for survival prediction.","marker":"[20]"},{"why":"Point-cloud hierarchy method that inspires the grouping design and serves as a baseline.","marker":"[39]"},{"why":"Strong point-cloud transformer baseline that CCFormer outperforms.","marker":"[57]"}],"fun_headline_variants":["Cell clouds beat pixel patches for cancer prognosis","5B cell labels power SOTA survival and staging models","Hierarchical cell transformer wins on cancer staging alone","Cell distribution alone predicts survival and stage"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The cell-level labels and coordinates in the dataset, produced by an automated detection pipeline refined with foundation-model votes, are accurate enough that measured clinical performance reflects real spatial biology rather than systematic annotation errors.","fun_headline_variants_meta":{"raw":{"variants":["Cell clouds beat pixel patches for cancer prognosis","5B cell labels power SOTA survival and staging models","Hierarchical cell transformer wins on cancer staging alone","Cell distribution alone predicts survival and stage"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00027,"raw_usage":{"total_tokens":1670,"prompt_tokens":1034,"completion_tokens":636,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":650,"completion_tokens_details":{"reasoning_tokens":578}},"tokens_in":650,"tokens_out":636,"duration_ms":6589,"temperature":1.0,"reasoning_tokens":578,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:20:03.028788+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Have pathologists annotate every cell (or a dense grid of patches) on a random sample of, say, 50 whole-slide images from the dataset; if cell-type agreement with the pipeline is substantially below the reported 92.7% patch-level tumor/normal agreement, or if errors concentrate in specific cell types or tissue regions, the survival and staging results could be artifacts of annotation noise. A quicker check: train CCFormer on the dataset with cell-type labels randomly permuted within each slide while keeping coordinates fixed; if the C-Index does not drop toward chance, the model is exploiting label noise rather than true spatial composition.","supporting_citations":[{"cited_title":"Dpa-p2pnet: Deformable proposal-aware p2pnet for accurate point-based cell detection","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained cell detection and classification model used to produce preliminary annotations."},{"cited_title":"Toward a shared vision for cancer genomic data","cited_arxiv_id":null,"evidence_quote":"Sources the whole-slide images and clinical endpoints used to build the dataset."},{"cited_title":"A visual- language foundation model for computational pathology","cited_arxiv_id":null,"evidence_quote":"One of the foundation models whose patch-level tumor/normal votes drive the weakly supervised refinement."},{"cited_title":"Quilt-1m: One million image-text pairs for histopathology","cited_arxiv_id":null,"evidence_quote":"Another foundation model used in the voting ensemble for label refinement."},{"cited_title":"Whole slide images are 2d point clouds: Context-aware survival prediction using patch-based graph convolutional networks","cited_arxiv_id":null,"evidence_quote":"Graph-of-patches survival baseline that also frames whole-slide images as point sets; CCFormer is compared against it."},{"cited_title":"Attention-based deep multiple instance learning","cited_arxiv_id":null,"evidence_quote":"Standard attention-based multiple-instance-learning baseline for survival prediction."},{"cited_title":"Pointnet++: Deep hierarchical feature learning on point sets in a metric space","cited_arxiv_id":null,"evidence_quote":"Point-cloud hierarchy method that inspires the grouping design and serves as a baseline."},{"cited_title":"Point transformer v3: Simpler faster stronger","cited_arxiv_id":null,"evidence_quote":"Strong point-cloud transformer baseline that CCFormer outperforms."}],"review_version":1}