{"id":"1dec8631-344f-4051-be07-3bd8980de53c","arxiv_id":"2507.18848","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"PTCMIL learns prompt-token clusters of patches inside a vision transformer, improving whole slide image classification and survival analysis.","lead":"This paper adds small trainable 'prompt' markers to a transformer model that groups microscopic image patches from cancer slides into meaningful clusters. The authors report improved cancer diagnosis and survival predictions on eight datasets, with code released.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PTCMIL's headline 'outperforms SOTA' is not statistically supported: several best results are within one standard deviation of top baselines, with tied kappa and c-index; no paired significance tests or confidence intervals are reported.","rationale":"The reader identifies the raw inner-product affinity in Eq (2) as the weakest assumption, arguing that prompt collapse could make the cluster prototypes uninformative. I agree collapse would be harmful, but the paper includes explicit safeguards—orthogonal initialization, the soft orthogonality regularizer in Eq (5), a moving-average update, and qualitative clustering maps in Fig 2—so the geometry concern is not directly evidenced. The more load-bearing risk is that the reported SOTA margins are statistically indistinguishable from the best baselines on several key metrics. Since the abstract and Sec 3.3 explicitly claim superiority, the absence of paired significance tests, confidence intervals, and kappa uncertainty makes the headline comparative claim unfalsified rather than false. The paper's own tables show ties (PANDA kappa with CTransPath, LUAD c-index) and margins within one standard deviation, which is exactly where statistical testing matters. A concrete re-analysis with paired tests on the released code would settle whether the central claim survives. The reader's conditional verdict is appropriate; no change to the verdict is needed, but the condition should explicitly require significance testing or a softened comparative claim.","tokens_in":9602,"tokens_out":7127,"duration_ms":84682,"concrete_test":"Obtain the released code and reproduce PTCMIL and the top three baselines on the same folds, then run paired DeLong tests for AUC, McNemar or paired bootstrap tests for accuracy/kappa, and paired bootstrap tests for c-index (e.g., 10,000 resamples) on all eight datasets, reporting 95% confidence intervals. If the majority of comparisons against the best baseline are not significant at p<0.05 after multiple-comparison correction, the abstract's 'outperforms' claim should be softened to 'matches or exceeds' and the verdict should remain conditional.","verdict_should_be":"UNCHANGED","load_bearing_attack":"PTCMIL's central claim is that it 'outperforms state-of-the-art methods' on eight datasets (Abstract; Sec 3.3). The supporting evidence is statistically insecure. In Table 1, using UNI features, PTCMIL's TCGA-NSCLC accuracy is 98.60±0.35 versus CLAM's 97.97±0.43 and MambaMIL's 97.97±1.96; Camelyon16 AUC is 99.60±0.34 versus MambaMIL's 99.06±0.77. With CTransPath features, the PANDA kappa is 0.928, identical to ILRA's 0.928, and in Table 2 the LUAD c-index is 0.688, a tie with ILRA's 0.688. The paper reports no standard errors for kappa, no paired tests, no confidence intervals, and no correction for multiple comparisons across eight datasets and two feature extractors. Because several leading margins are within one standard deviation of the best baseline, the load-bearing empirical assertion that end-to-end prompt-token clustering yields a state-of-the-art aggregator is not established as presented. The mechanism may be sound, but the headline comparison is under-powered as reported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PTCMIL, a multiple instance learning aggregator for whole slide images that inserts learnable prompt tokens into a ViT backbone, treats the prompts as cluster prototypes via a soft inner-product assignment (Eq. 2), merges tokens inside each cluster into weighted prototypes, and pools prototype tokens with a class token for classification and survival tasks. Experiments on eight datasets with CTransPath and UNI features report accuracy, AUC, Cohen's kappa, or c-index against multiple prior MIL methods, together with few-shot adaptation results and ablations of the clustering, merging, and pooling modules.","tokens_in":9865,"tokens_out":4812,"duration_ms":55708,"significance":"If the empirical claims hold, PTCMIL offers a conceptually clean end-to-end alternative to two-stage clustering MIL, with limited added parameters and naturally interpretable cluster maps. The authors provide code, systematic ablations of the three architectural components, and visual evidence that the learned clusters correspond to meaningful tissue types. The main contribution is the unification of task-aligned clustering and prediction, and the paper avoids circular evaluation because the representation is trained end-to-end and assessed on external benchmarks. The significance is currently tempered by gaps in the statistical support for the headline state-of-the-art claim and by an uninformative one-class adaptation evaluation.","major_comments":[{"comment":"The central claim that PTCMIL 'outperforms state-of-the-art methods' is not statistically established as presented. Several leading margins are within one standard deviation of the best baseline or are ties: in Table 2 the LUAD c-index is 0.688 for both PTCMIL and ILRA, and the BRCA c-index of 0.745 vs PANTHER's 0.729 is within PANTHER's reported standard deviation of 0.08; in Table 1 the CTransPath PANDA kappa is 0.928 for both PTCMIL and ILRA. The paper reports no paired significance tests, no confidence intervals for kappa or c-index, and no correction for multiple comparisons across eight datasets and two feature extractors. Please add paired permutation tests or bootstrap confidence intervals for the head-to-head comparisons, and temper the abstract and Sec. 3.3 wording to the level of support the statistics provide.","section":"Sec. 3.3, Table 1, Table 2"},{"comment":"The in-house prostate adaptation evaluation is uninformative as reported. The dataset is described as '1-class' and contains 749 cancerous slides, and the paper reports only accuracy for this column. On a one-class test set, a constant classifier can achieve high accuracy, so this result does not measure the model's ability to distinguish cancer from normal tissue or to adapt to a new domain. Please evaluate on a class-balanced or multi-class cohort and report sensitivity/specificity, AUC, or balanced accuracy; otherwise, remove the adaptability claim based on this column.","section":"Sec. 3.1, Table 1"},{"comment":"The load-bearing clustering mechanism relies on a raw inner-product softmax over prompt tokens, with no temperature, learned projection, or normalization, and collapse prevention depends on orthogonal initialization, the soft orthogonality regularizer Lreg, and a moving average update. The paper provides no quantitative evidence that the prompts remain diverse and that the assignments are informative after training. Please report cluster assignment entropy or purity, prompt token orthogonality during training, and sensitivity to the number of clusters C, the regularization weight alpha, and the decay factor theta; also compare against a version with a learned projection or temperature in the affinity. Such an analysis would directly support the claim that the prompt-token clustering geometry is stable and meaningful.","section":"Sec. 2.1, Eq. (2), Prompt Updating"}],"minor_comments":[{"comment":"The ablation table is very difficult to interpret because the rows are not explicitly labeled and the checkmark/dash structure is unclear. Please give each row a named configuration, e.g., 'without clustering', 'without merging', 'without prototype pooling', so the reader can map rows to the text in Sec. 3.5.","section":"Table 4"},{"comment":"Two accuracy values in Table 4 appear to have formatting errors ('94 1.87' for TCGA-NSCLC and '95 0.87' for TCGA-CRC); please correct the decimal formatting.","section":"Table 4"},{"comment":"References [18] and [19] are listed as the same PANTHER paper with identical titles and venue; please remove the duplicate and renumber.","section":"References"},{"comment":"The text says 'the cluster index vector a = ... for i in [C]' and later refers to 'cluster i' and 'for cluster i' where the index should be c; please fix the variable naming for clarity.","section":"Sec. 2.2"},{"comment":"The censoring indicator cj is defined as 1 for patients who live longer than the follow-up period and 0 for patients who pass away, which is the opposite of the common convention. Please state the convention explicitly in the text to avoid confusion with the survival likelihood formula.","section":"Sec. 2.3, Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":"The one-class accuracy evaluation is a substantive flaw rather than a mere presentation issue; if the authors cannot obtain a class-balanced adaptation cohort, the adaptability claim in Table 1 and Sec. 3.3 should be removed. The statistical support for the SOTA claim is also weak, so I would ask for paired tests or confidence intervals before accepting the headline. The reference duplication for PANTHER ([18] and [19]) suggests a citation slip that should be checked."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: PTCMIL is a genuinely new MIL aggregator for WSIs — learnable prompt tokens act as cluster centers, patches are soft-assigned by projection, and clusters are merged into prototypes, all trained end-to-end with the downstream head. The mechanism is clean, the ablations are good, and the authors ship a broad evaluation (eight datasets, two feature extractors, few-shot adaptation, code). That said, the 'outperforms SOTA' headline is not established as written: no significance tests, and several reported wins are within one standard deviation of the best baseline, with an outright tie on PANDA kappa (0.928) and LUAD c-index (0.688). I'd still send it to a serious referee, but ask for statistical stiffening before publication.\n\nWhat's actually new: prior clustering ViTs compute pairwise token similarities, and PANTHER does clustering in a separate unsupervised stage. Here the prompts are the cluster centers, the assignment is a softmax over inner products with prompt tokens, and a local transformer then refines each cluster before weighted centroid pooling. The design is efficient and task-aligned, and the ablation table supports each module.\n\nSoft spots, roughly in order of severity:\n1. Evidence for the central claim is underpowered. Across Table 1 and Table 2, several PTCMIL numbers are within one std of a strong baseline (e.g., UNI Acc on TCGA-NSCLC 98.60±0.35 vs CLAM 97.97±0.43; CTransPath PANDA kappa 0.928 tied with ILRA). No paired tests, no confidence intervals, no multiple-comparison correction. The aggregate pattern is positive, but 'outperforms' is too strong.\n2. C (number of clusters) is selected per dataset and the authors acknowledge future work will seek automatic selection; fine for now, but a sensitivity analysis across C (they show one in Fig. 3) should accompany the main tables, and nested validation would be cleaner.\n3. The in-house prostate set is 1-class, and they report accuracy on it. That number is uninformative. Report something meaningful or drop it.\n4. The raw inner-product assignment in Eq. 2 is fragile; prompt collapse is a real risk. The Gram-Schmidt init, orthogonality regularizer, and moving average are plausible guards, but I'd like to see a prompt-collapse diagnostic or an ablation with random prompt reinit.\n5. Citation inconsistencies: ILRA is cited as [23] in the tables but [20] in Table 3 (a different paper appears as [20]); PANTHER appears twice as [18] and [19] with identical text. Easy fixes, but sloppy.\n\nBottom line: the paper is a credible engineering contribution with a novel mechanism, and the authors are honest about the cluster-count limitation. The math is simple and sound, the code release helps reproducibility. For a computational pathology audience, it's worth reading and worth refereeing. My recommendation: send to a strong reviewer, with the expectation of a revision that adds significance tests, cleans up the metrics, and fixes references.","headline":"Solid new MIL aggregator with a novel prompt-token clustering mechanism, but the SOTA claim needs statistical support.","tokens_in":10403,"tokens_out":4173,"would_cite":true,"duration_ms":40629,"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":"A Vision Transformer with learnable prompt tokens as cluster centers outperforms prior MIL methods on eight whole-slide image datasets, in both classification and survival analysis.","keywords":["multiple instance learning","whole slide image analysis","prompt token clustering","Vision Transformer","computational pathology","survival analysis","prototype pooling","weakly supervised learning"],"falsifier":"Re-run the Camelyon16 and TCGA-NSCLC experiments with the prompt tokens frozen at their orthonormal initial values so that gradients never reshape the cluster assignments; if accuracy and AUC do not drop materially, end-to-end clustering is not the cause of the improvements.","tokens_in":9383,"feed_emoji":"🔬","tokens_out":8970,"duration_ms":95075,"temperature":0.7,"pith_summary":"This paper proposes PTCMIL, a multiple-instance-learning method that reads a whole slide image—a giga-pixel digitized biopsy—as a bag of patch tokens and compresses it into a small set of cluster prototypes. Its central claim is that clustering and slide-level prediction should be optimized together, unlike earlier prototype-based MIL methods that first cluster patches unsupervised and then train a predictor on the prototypes. The method introduces learnable prompt tokens as cluster centers inside a Vision Transformer, assigns each patch to a prompt by inner-product affinity, and lets the classification or survival loss reshape those assignments. The authors report that this end-to-end design outperforms existing MIL aggregators on eight whole-slide datasets and produces clusters that correspond to interpretable tissue types such as tumor, stroma, alveoli, and vessels.","feed_headline":"Prompt-token clustering beats two-stage MIL on eight slide datasets","feed_subtitle":"End-to-end prompt clusters replace unsupervised prototypes, lifting cancer diagnosis and survival prediction.","key_machinery":"The load-bearing object is a set of $C$ learnable prompt tokens, one per cluster, that enter a global Transformer alongside the patch tokens and the class token. After one self-attention pass, each patch token is assigned to a cluster by a softmax over inner products with the prompt tokens (Eq. (2)); the assignment is differentiable, so the downstream classification or survival loss reshapes the clusters. An orthogonality regularizer $L_{\\text{reg}} = \\|\\bar{P}^T \\bar{P} - I\\|^2$ plus a moving-average update keeps the prompt tokens from collapsing, then a shared local Transformer per cluster and weighted centroid merging compress each cluster into one prototype, and the prototypes are pooled with the class token for prediction. This stack is what lets the paper claim that clustering and prediction are optimized jointly rather than in two stages.","core_discovery":"On the paper's own terms, the discovery is that clustering and prediction need not be decoupled in whole-slide image analysis: a small number of trainable prompt vectors inside a Vision Transformer can serve as cluster centroids, the assignment of patches to centroids is differentiable through a softmax inner product, and the task loss therefore shapes the clusters. The paper reports that this end-to-end coupling is also the efficient option, because it avoids computing all pairwise patch similarities over tens of thousands of tokens, and that it yields the highest accuracy, AUC, kappa, and concordance index among the compared methods on the eight datasets tested, for example a Camelyon16 AUC of 99.60 with UNI features and a PANDA kappa of 0.937. It further reports that the resulting clusters line up with recognizable tissue types and that a few-shot transfer variant is more accurate than baseline adaptations.","pith_inferences":["An untested stress test: replace the raw inner-product affinity in Eq. (2) with a learned projection and temperature; if the gains survive, end-to-end prototype supervision, not the specific prompt geometry, is the active ingredient.","The prompts could serve as a stain or institution harmonization layer: because they are few, shared, and task-trained, finetuning them on a small local cohort might adapt a slide model across scanners without touching the feature extractor, which goes beyond the paper's few-shot experiments.","The paper's stability across cluster counts of 3 to 9 suggests that cluster number could be chosen automatically, for example by a validation-based search or by prototype redundancy, which the authors list as future work."],"forward_implications":["The two-stage design of prototype-based MIL—unsupervised clustering first, supervised prediction second—is not necessary; a single ViT can cluster and predict under one loss.","A model trained on one cancer type can be adapted to another with about twenty labeled slides by finetuning only the classifier and prototypes, so the approach suits low-data clinical settings.","The same prototype representation serves classification and survival analysis, meaning one pretrained aggregator can be redirected to time-to-event tasks without architectural changes.","Learned clusters correspond to interpretable tissue types, so slide-level predictions come with a spatial map of where each cluster fires, which is useful for pathology review."],"supporting_citations":[{"why":"Supplies the learnable prompt-token mechanism that PTCMIL repurposes as cluster centers.","marker":"[12]"},{"why":"Defines the two-stage unsupervised prototyping pipeline PTCMIL is designed to replace and serves as the main interpretive contrast.","marker":"[18]"},{"why":"Provides the correlated ViT-based MIL baseline and the motivation for using global self-attention over patches.","marker":"[17]"},{"why":"Contributes the CLAM baseline and the data-processing toolbox used to crop patch features and evaluate classification.","marker":"[14]"},{"why":"Produces the UNI patch embeddings used as input features for the first set of reported results.","marker":"[6]"},{"why":"Produces the CTransPath patch embeddings used for the second set of reported results.","marker":"[22]"},{"why":"Gives the initialization scheme, with Gram-Schmidt orthogonalization, that the paper relies on to prevent cluster collapse.","marker":"[8]"},{"why":"Provides the learnable averaging weights for token merging that PTCMIL adapts for prototype computation.","marker":"[16]"},{"why":"Supplies the survival-analysis loss and discrete-time hazard formulation used for the survival experiments.","marker":"[7]"}],"fun_headline_variants":["End-to-end prompt clusters outclass two-stage MIL","Trainable prompts replace unsupervised MIL clusters","Prompt-token clustering lifts AUC across eight slide datasets","PTCMIL: prompts become clusters for better WSI analysis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole argument rests on the hope that ordinary similarity scores between patch tokens and prompt tokens, after one pass through the Transformer, divide a slide's tissue into stable and meaningful groups; if those groups collapse or drift, the prototypes become uninformative and the reported gains vanish.","fun_headline_variants_meta":{"raw":{"variants":["End-to-end prompt clusters outclass two-stage MIL","Trainable prompts replace unsupervised MIL clusters","Prompt-token clustering lifts AUC across eight slide datasets","PTCMIL: prompts become clusters for better WSI analysis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000715,"raw_usage":{"total_tokens":3193,"prompt_tokens":906,"completion_tokens":2287,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":2225}},"tokens_in":522,"tokens_out":2287,"duration_ms":18932,"temperature":1.0,"reasoning_tokens":2225,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:30:47.691235+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Camelyon16 and TCGA-NSCLC experiments with the prompt tokens frozen at their orthonormal initial values so that gradients never reshape the cluster assignments; if accuracy and AUC do not drop materially, end-to-end clustering is not the cause of the improvements.","supporting_citations":[{"cited_title":"In: European Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Supplies the learnable prompt-token mechanism that PTCMIL repurposes as cluster centers."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Defines the two-stage unsupervised prototyping pipeline PTCMIL is designed to replace and serves as the main interpretive contrast."},{"cited_title":"In: Proceedings of the thirteenth international conference on ar- tificial intelligence and statistics","cited_arxiv_id":null,"evidence_quote":"Gives the initialization scheme, with Gram-Schmidt orthogonalization, that the paper relies on to prevent cluster collapse."},{"cited_title":"In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision (ICCV)","cited_arxiv_id":null,"evidence_quote":"Supplies the survival-analysis loss and discrete-time hazard formulation used for the survival experiments."}],"review_version":1}