{"id":"06f30009-a0bd-4d50-a4dc-4215c32b688b","arxiv_id":"2502.07979","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"M3C2 jointly predicts histology features and three molecular markers from whole slide images and uses their interactions to classify gliomas under WHO 2021 criteria, reporting better accuracy than 13 prior methods on internal and external data.","lead":"A new deep learning system reads microscope images of brain tumor tissue at two magnifications and simultaneously predicts the molecular markers and tissue features used in the 2021 WHO glioma diagnosis, then combines them to classify the tumor. If it holds up, it could let doctors get a modern molecular classification from a routine stained slide without waiting for gene sequencing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Case-level split unconfirmed for multi-regional internal dataset; WSI-level splitting would inflate the reported 5.6% gains.","rationale":"The reader's weakest assumption correctly identifies the most load-bearing verification gap: the internal dataset is multi-regional and multi-WSI per case, yet the paper never states that the split is case-level. If the split is WSI-level, the central empirical claim is invalid regardless of the method's architectural merits, because patient-level leakage would explain the reported gains. I agree with the reader's verdict of CONDITIONAL: the open-source code and supplementary material are not available for inspection in this review, so the split cannot be verified from the manuscript. The concern is concrete and testable, and it does not require rejecting the method's novelty. I considered whether the CPLC-Graph label-prior leak is even more serious, but that leak is secondary: it is a 3x3 co-occurrence matrix that may only slightly bias results, whereas WSI-level splitting would systematically corrupt all comparisons. Therefore the split is the single most load-bearing concern, and the appropriate action is to condition acceptance on verifying the split and re-running with proper patient-level separation.","tokens_in":24477,"tokens_out":4177,"duration_ms":38001,"concrete_test":"Inspect the released repository's data-splitting script to determine whether the internal train/validation/test split is at the case (patient) level or the WSI level. Then re-run the Table 1 comparison with a strict case-level split across at least three random seeds and report mean and standard deviation. If the 78.6% internal accuracy drops below the best baseline or the 5.6% margin collapses within seed noise, the superiority claim fails. Additionally, check whether the CPLC-Graph matrix A is computed from training labels only; if it uses all internal cases, recompute A from the training split and re-evaluate the molecular-marker and glioma classification results.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical superiority on the internal dataset: Table 1 reports 78.6% accuracy versus 73.0% for the best baseline, a 5.6% gain. This claim requires the train/validation/test split to be at the case (patient) level, because the internal dataset merges multi-regional TCGA GBM-LGG and CPTAC into 3,578 WSIs from only 1,054 cases. Section 4.1 does not state that all WSIs from a single case are confined to one fold. If the split is at the WSI level, multiple WSIs from the same patient can appear in both training and test, letting the model memorize patient-specific appearance (staining, sectioning, dominant morphology) and inflate every internal metric, including the 78.6% accuracy, sensitivity, AUC, and F1-score. The external validation (98.1% accuracy) does not rescue the internal comparison because it uses a different dataset with only two classes, as the paper itself notes in Section 4.4. A secondary but related leak is that the CPLC-Graph correlation matrix A (Section 3.4) is said to be 'calculated using our internal large-scale TCGA dataset' with no statement that it was restricted to the training split; using test-label co-occurrence statistics in the graph adjacency and LC loss could further inflate molecular-marker performance. No error bars are reported, so the stated margins cannot be distinguished from seed-level noise even under a correct split.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes M3C2, a multi-task multi-instance learning framework that jointly predicts WHO 2021 glioma classes, histology features (NMP), and molecular markers (IDH, 1p/19q, CDKN) from multi-magnification whole-slide images. It introduces a multi-scale disentangling module, a co-occurrence probability-based label-correlation graph (CPLC-Graph) with an LC loss, a dynamic confidence constrain (DCC) loss, and a cross-modal gradient modulation (CMG-Modu) strategy. The authors validate on an internal dataset merging TCGA GBM-LGG and CPTAC (3,578 WSIs from 1,054 cases) and an external IvYGAP dataset, reporting state-of-the-art accuracy (78.6% internal, 98.1% external) and improved molecular marker and histology prediction. Code is released.","tokens_in":24734,"tokens_out":3979,"duration_ms":32912,"significance":"If the empirical claims hold, M3C2 would be a valuable contribution to computational pathology, aligning with the WHO 2021 diagnostic paradigm by jointly modeling histology-molecular interactions and predicting markers from WSIs alone. The paper is commendable for its broad experimental scope (13 baselines, ablations, subgroup and magnification analyses, external validation, released code). However, the central claim of superiority on the internal dataset is currently weakened by an unspecified split unit, a potentially leaking co-occurrence matrix, and the absence of uncertainty quantification; these issues need to be resolved before the results can be taken as evidence for the method's advantage.","major_comments":[{"comment":"The internal dataset is described as 3,578 WSIs from 1,054 cases, but the paper never states whether all WSIs from a single case are confined to the same train/validation/test fold; if the split is at the WSI level, the same patient's slides can appear in both training and testing, and the reported internal gains (Table 1, 78.6% vs 73.0%) could be inflated by memorizing patient-specific appearance. The authors must specify the split unit and, if it is not case-level, re-run the experiments with a case-level split.","section":"Section 4.1"},{"comment":"The co-occurrence matrix A is 'calculated using our internal large-scale TCGA dataset' with no statement that the calculation is restricted to the training fold; since A is used both as graph adjacency and as the target of the LC loss (Eq. 5), estimating it from the full cohort (including test cases) would leak aggregate label information into training. Please report how A is computed and confirm it uses only the training split, or provide a cross-validation scheme for A.","section":"Section 3.4, Eq. (4)"},{"comment":"All reported metrics are single-point estimates with no error bars, confidence intervals, or significance tests; the headline margin of 5.6% in accuracy over the best baseline could be within seed-level variation. The authors should report mean plus/minus standard deviation over multiple runs (or at least bootstrap confidence intervals) and a significance test for the main comparisons.","section":"Section 4.3, Tables 1 and 2"},{"comment":"The ablation for LC loss shows a drop from 78.6% to 62.5% accuracy (16.1%), which is unusually large for a loss term and may be a symptom of the leakage in A described above; the authors should re-evaluate this ablation under a training-fold-only A and discuss whether the large effect persists.","section":"Section 4.4, Table 4"}],"minor_comments":[{"comment":"The CMG-Modu ablation text refers to 'Figure 4.4' but the actual figure is numbered Figure 7 in the manuscript; the cross-reference should be corrected.","section":"Section 4.4"},{"comment":"The column header 'w/o GMC-Modu' appears to be a typo for 'w/o CMG-Modu'.","section":"Table 5"},{"comment":"The sentence 'the middle panel of Table 1 shows that shows that M3C2 performs the best' contains a duplicated phrase and should be edited.","section":"Section 4.3"},{"comment":"The claim that this is 'the first attempt to classify cancer via modeling the interaction of histology and molecular markers predictions' is too strong given that the conference version (Wang et al., 2023) already models such interactions; please rephrase to avoid overclaiming.","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":"The split and leakage concerns are the crux. If the internal split is at the WSI level, the paper's headline results are invalid and a full re-run is needed. The authors' acknowledgement of the two-class external dataset is honest, but the main text should present the two-class internal comparison alongside. The unusually large ablation effect of the LC loss makes the leakage concern concrete. I would recommend major revision with a request for case-level split details and retraining with a training-only A."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a legitimate extension of the authors' MICCAI 2023 paper. If the data handling checks out, it's a useful step toward WSI-only WHO 2021 glioma classification. But two evaluation details — the case-level split and the source of the co-occurrence matrix — are left ambiguous in the main text, and without them the reported 5.6% gains don't hold up yet.\n\nWhat's actually new: the multi-scale disentangling module, the CPLC-Graph co-occurrence graph, the attention-based hierarchical MIL, and the DCC loss with CMG-Modu gradient modulation. Each component builds on known techniques, but the combination is original, and the ablation study convincingly shows each piece helps. The paper also compares against 13 baselines, includes an external validation set, subgroup analysis by FFPE/frozen, and ships code. The experimental breadth is real, and the external 98.1% on IvYGAP is a decent sanity check even with the two-class caveat.\n\nSoft spots, in order of importance. (1) Section 4.1 merges TCGA GBM-LGG and CPTAC into 3,578 WSIs from 1,054 cases but never states that all WSIs from one patient are confined to a single fold. Since these are multi-regional datasets, a WSI-level split would let the model memorize patient identity and inflate every metric. If the split is case-level, this concern goes away, but it needs to be stated explicitly. (2) The CPLC-Graph's co-occurrence matrix A is 'calculated using our internal large-scale TCGA dataset' with no statement that it's restricted to the training fold. If it isn't, test-label co-occurrence statistics leak into both the graph adjacency and the LC loss. (3) All reported metrics are single-point estimates; no error bars or significance tests. That matters less than the split issue but should be addressed.\n\nThe external validation is good evidence but doesn't rescue the internal comparison — the paper itself notes IvYGAP has only two classes, and the two-class internal comparison lives in the supplementary, which isn't available in the arXiv version.\n\nWho's this for: computational pathologists and neuro-oncology groups working on molecular marker prediction from WSIs. The clinical motivation is strong and the WHO 2021 alignment is a real selling point. It deserves a serious referee. Send it to review, with a request that the authors clarify the split and the co-occurrence matrix construction, and ideally add confidence intervals. If the split is case-level and A is train-only, this is a solid accept.","headline":"Genuine extension with a clinically relevant goal, but the internal split and co-occurrence matrix need verification before the headline numbers can be trusted.","tokens_in":25324,"tokens_out":3859,"would_cite":true,"duration_ms":33003,"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":"Jointly predicting molecular markers and histology features from whole slide images, with explicit modeling of their co-occurrence and interaction, yields 78.6% internal and 98.1% external accuracy on four-class WHO 2021 glioma…","keywords":["Cancer Classification","Molecular Pathology","Digital Pathology","Multi-task Learning","Multi-modal Learning","Multi-scale Modeling","Glioma","Whole Slide Image"],"falsifier":"Re-run M3C2 and all baselines on the internal dataset with a strict patient-level split (all WSIs of a case in the same fold, repeated over several seeds) and recompute the accuracy gap; if the at-least-5.6-point advantage over the best baseline shrinks or disappears, the headline result was inflated by patient-level leakage, and the external IvYGAP result would remain fair only if its split is likewise case-level.","tokens_in":24252,"feed_emoji":"🧠","tokens_out":8322,"duration_ms":60283,"temperature":0.7,"pith_summary":"Cancers like glioma are now classified under WHO 2021 by combining histology features with molecular markers, and this paper argues that digital pathology should follow that paradigm: a single model should predict both kinds of information from whole slide images and explicitly model how they co-occur and interact. To that end it introduces M3C2, a multi-task multi-instance framework that takes 10X and 20X patches, disentangles shared versus task-specific multi-scale features, and couples the molecular and histology prediction streams through a label-correlation graph, a dynamic-confidence loss, and a gradient modulation strategy. On an internal set of 3,578 WSIs from 1,054 cases, M3C2 reaches 78.6% accuracy in four-class glioma classification, at least 5.6 percentage points above thirteen baselines, and 98.1% accuracy on an external cohort without fine-tuning. If correct, this would mean that predicting the biomarkers themselves, rather than only the final label, and letting the two prediction streams guide each other, is a practical route to AI diagnostics that match the current clinical workflow.","feed_headline":"Joint histology-marker model lifts glioma accuracy by 5.6 points","feed_subtitle":"One pass predicts WHO 2021 glioma class, molecular markers, and histology, then hits 98.1% on external slides.","key_machinery":"The load-bearing mechanism is a four-part hierarchy. A multi-scale disentangling module splits 10X (tissue-level) and 20X (cellular-level) patch features into shared, molecular-specific, and histology-specific components using a disentanglement loss. An attention-based hierarchical multi-task multi-instance learning (AHMT-MIL) backbone pools up to 2,500 patches per slide per magnification to produce slide-level embeddings. A co-occurrence probability-based label-correlation graph (CPLC-Graph) network, with adjacency entries $A^i_j = \\frac{1}{2}(p(F^i|F^j)+p(F^j|F^i))$ computed from the internal dataset, ties the three molecular markers together, and a label-correlation (LC) loss aligns the cosine similarity of their output features to that matrix. Finally, a cross-modal interaction module couples the histology and molecular streams: a dynamic confidence constrain (DCC) loss forces the top attention patches of IDH wild-type and NMP to overlap, while cross-modal gradient modulation (CMG-Modu) projects each task's gradients perpendicular to the other's depending on whether NMP is present.","core_discovery":"On the paper's own terms, the central discovery is that a whole-slide-image model can outperform separate predictors by treating histology features and molecular markers as jointly predicted, interacting targets rather than as inputs or independent tasks. The empirical claim is that the four WHO 2021 glioma classes can be classified at 78.6% internal accuracy, at least 5.6 points higher in accuracy, sensitivity, and F1 than any of thirteen compared methods, and 98.1% accuracy on an external cohort, while simultaneously predicting IDH mutation, 1p/19q co-deletion, CDKN homozygous deletion, and the histology feature of necrosis and microvascular proliferation with AUCs of 95.0%, 90.9%, 76.0%, and 98.1%. The interaction itself is the point: removing either the molecular or the histology branch drops internal accuracy by about ten points, and ablating any of the cross-modal coupling mechanisms degrades performance further.","pith_inferences":["If the internal gains survive a strict patient-level split, treating biomarker prediction as an auxiliary task with explicit cross-modal coupling could be adopted as a general regularizer for WSI classifiers beyond glioma.","The CPLC-Graph's co-occurrence matrix is computed once from the merged TCGA set; on a new population these probabilities should be re-estimated, and the reported gains may depend on that prior.","The 98.1% external accuracy is on a two-class subset (glioblastoma versus high-grade astrocytoma); an external cohort containing all four classes would be a stronger test of the full WHO 2021 scheme.","The DCC loss couples only IDH and NMP, so the paper's claim that the mechanism generalizes to other histology-molecular pairs is plausible but untested by the reported experiments."],"forward_implications":["A single whole-slide-image pipeline can replace separate molecular-sequencing and histology workflows for the four WHO 2021 glioma classes, reducing cost and turnaround time.","The same joint-prediction-plus-interaction recipe is directly applicable to other cancers whose classification now depends on histology-molecular pairs, such as endometrial, renal, and thyroid tumors.","Multi-scale input (10X tissue-level plus 20X cellular-level) with disentangled shared and task-specific features outperforms single-magnification training and simple feature concatenation.","Explicit co-occurrence modeling through the CPLC-Graph and LC loss improves each molecular marker's prediction, not just the final glioma label.","The cross-modal interaction mechanisms (DCC loss and CMG-Modu) are each necessary: removing any one of them lowers internal accuracy by multiple points."],"supporting_citations":[{"why":"Defines the conference baseline that M3C2 extends and must beat; the paper reports a 5.6-point accuracy gain over it.","marker":"Wang et al. (2023)"},{"why":"Supplies the WHO 2021 diagnostic criteria that define the four glioma classes and the molecular markers (IDH, 1p/19q, CDKN) the model predicts.","marker":"(Louis, 2021)"},{"why":"Source of the TCGA GBM-LGG dataset that forms part of the internal training and evaluation data.","marker":"(Tomczak et al., 2015)"},{"why":"Source of the CPTAC dataset that is merged with TCGA to form the internal dataset.","marker":"(Verdugo et al., 2022)"},{"why":"Source of the IvYGAP external validation cohort on which the model reaches 98.1% accuracy without fine-tuning.","marker":"(Eberhart and Bar, 2020)"},{"why":"Provides the clinical co-occurrence fact (over 95% of NMP-positive patients are IDH wild-type) that motivates the DCC loss.","marker":"(Alzial et al., 2022)"},{"why":"Charm, a WHO 2021 glioma classifier used as a state-of-the-art baseline in the comparison.","marker":"(Nasrallah et al., 2023)"},{"why":"Deepglioma, another state-of-the-art WHO 2021 glioma method used as a baseline.","marker":"(Khalighi et al., 2024)"}],"fun_headline_variants":["Glioma model jointly predicts histology and markers, beats 13 rivals","Joint histology-marker model: 78.6% internal, 98.1% external glioma accuracy","One model, two domains: histology and markers jointly lift glioma accuracy","Cross-modal model for glioma: 5.6-point accuracy gain over separate tasks","Histology and molecular markers modeled together: 5.6-point boost for glioma"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains assume that the internal train/test split keeps every slide of a patient in the same fold, but the paper merges multi-regional TCGA and CPTAC sets (3,578 WSIs from 1,054 cases) without ever stating that the split is at the patient level.","fun_headline_variants_meta":{"raw":{"variants":["Glioma model jointly predicts histology and markers, beats 13 rivals","Joint histology-marker model: 78.6% internal, 98.1% external glioma accuracy","One model, two domains: histology and markers jointly lift glioma accuracy","Cross-modal model for glioma: 5.6-point accuracy gain over separate tasks","Histology and molecular markers modeled together: 5.6-point boost for glioma"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001144,"raw_usage":{"total_tokens":4782,"prompt_tokens":1018,"completion_tokens":3764,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":3656}},"tokens_in":634,"tokens_out":3764,"duration_ms":22509,"temperature":1.0,"reasoning_tokens":3656,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T11:14:34.479319+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run M3C2 and all baselines on the internal dataset with a strict patient-level split (all WSIs of a case in the same fold, repeated over several seeds) and recompute the accuracy gap; if the at-least-5.6-point advantage over the best baseline shrinks or disappears, the headline result was inflated by patient-level leakage, and the external IvYGAP result would remain fair only if its split is likewise case-level.","supporting_citations":[{"cited_title":", author Price, S","cited_arxiv_id":null,"evidence_quote":"Defines the conference baseline that M3C2 extends and must beat; the paper reports a 5.6-point accuracy gain over it."},{"cited_title":", year 2021","cited_arxiv_id":null,"evidence_quote":"Supplies the WHO 2021 diagnostic criteria that define the four glioma classes and the molecular markers (IDH, 1p/19q, CDKN) the model predicts."},{"cited_title":", author Puerto, I","cited_arxiv_id":null,"evidence_quote":"Source of the CPTAC dataset that is merged with TCGA to form the internal dataset."},{"cited_title":", author Bar, E.E","cited_arxiv_id":null,"evidence_quote":"Source of the IvYGAP external validation cohort on which the model reaches 98.1% accuracy without fine-tuning."},{"cited_title":", author Renoult, O","cited_arxiv_id":null,"evidence_quote":"Provides the clinical co-occurrence fact (over 95% of NMP-positive patients are IDH wild-type) that motivates the DCC loss."},{"cited_title":", author Zhao, J","cited_arxiv_id":null,"evidence_quote":"Charm, a WHO 2021 glioma classifier used as a state-of-the-art baseline in the comparison."},{"cited_title":", author Reddy, K","cited_arxiv_id":null,"evidence_quote":"Deepglioma, another state-of-the-art WHO 2021 glioma method used as a baseline."}],"review_version":1}