{"id":"8c653dbc-41d7-434b-bbe2-42a8541d80e7","arxiv_id":"2411.17418","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A dual fusion network that combines DNA methylation and whole slide image features at both patch and slide levels modestly improves CNS tumor subtyping and survival prediction.","lead":"This paper merges DNA methylation profiles with whole slide tumor images in two fusion stages to classify brain tumor subtypes and predict patient survival. The reported gains over methylation-only models are small and the method is not yet validated outside its own experiments.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The CpG feature count and subset appear to be chosen on the full BRAIN UK cohort before the 2-fold split; without nested feature selection, the reported dual-fusion advantage over early/late fusion may be leakage-driven.","rationale":"The architecture is plausible and the ablations align qualitatively with the claim, so I do not reject the paper. But because the contribution is an incremental design choice and the headline numerical support is a small gap (0.055 F1-Macro over the best single-stage fusion, c-index differences within overlapping standard deviations), evaluation hygiene matters. The manuscript's own Limitations section acknowledges fusion trade-offs but does not mention feature-selection leakage, and no code is released. The reader's conditional verdict is therefore appropriate: the claim can be accepted only after a leakage-free nested feature-selection protocol is demonstrated and the statistical test is fully specified. My read does not move the verdict; it sharpens the reason why conditional acceptance is needed.","tokens_in":15388,"tokens_out":6891,"duration_ms":64680,"concrete_test":"Run a nested 2-fold cross-validation on BRAIN UK: inside each training fold, repeat the CpG selection (intersection of variance/CV/MAD/IQR) and tune the feature count among 4K/8K/10K using only that training fold; apply the chosen features to the held-out fold. Report F1-Macro for ABMIL-MOAD-FNet, early-only, and late-only. If the dual-fusion margin over the best single-stage fusion drops below the reported 0.055 (or changes sign), the central claim is not supported. Also report the exact Wilcoxon procedure with sample counts to confirm p=0.043.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section II-A, before describing the 2-fold split, states: 'we applied variance, coefficient of variation (CV), median absolute deviation, and inter-quartile range, selecting the 8K most variable CpG sites by intersecting these methods. We experimented with clustering 4K CpG sites and Random Forest selection of 10K sites... but 4K reduced performance and 10K showed no improvement, leading us to select 8K CpG sites.' The decision between 4K and 10K is explicitly performance-based, and no nested cross-validation is reported. If the 8K set was tuned on the same 1,504 patients later split into the two evaluation folds, the test folds have informed feature selection. The key evidence for the central claim is Table II: dual fusion F1-Macro 0.745 vs early-only 0.644 and late-only 0.690. All three variants share the selected CpG features, so the magnitude of the dual-vs-single gap can be inflated by feature selection tuned to the full dataset; a leakage-free protocol could shrink or reverse the gap. The paper also does not give enough detail on the Wilcoxon test (what is compared, over how many independent units) to establish that the p=0.043 result is valid; however, the feature-selection leak is the more fundamental threat because it undermines the controlled comparison at the center of the contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MOAD-FNet, a dual-fusion architecture that integrates whole-slide images (WSIs) with DNA methylation or pathway-level omics data. It performs early fusion by concatenating an omic embedding to each WSI patch embedding and processing the result with an MLP, followed by ABMIL attention pooling; it then reintroduces the omic embedding at a late stage through a Multimodal Outer Arithmetic Block (MOAB) that combines outer product, division, addition, and subtraction. The method is evaluated on a private NHNN BRAIN UK CNS tumor dataset with 20 subtypes (reporting F1-Macro 0.745 for the full model vs 0.644 for early-only and 0.690 for late-only fusion) and on TCGA-BLCA and TCGA-BRCA for survival prediction (reporting c-index 0.691 and 0.726, respectively), with ablations comparing fusion stages, backbones, and aggregation methods.","tokens_in":15628,"tokens_out":2790,"duration_ms":29294,"significance":"If the reported gains are leakage-free, the paper makes a useful contribution: it demonstrates that reintroducing omic data at both patch and slide levels can improve multimodal WSI-omics classification and survival prediction, and it provides a concrete fusion block (MOAB) with ablations against concatenation and Kronecker-product alternatives. The study is also strengthened by comparing against several recent baselines (SurvPath, MCAT, TransMIL, PIBD, MMP) on public TCGA cohorts and by including attention heatmap-based interpretability. However, the central comparison between dual and single-stage fusion on the CNS dataset rests on a feature-selection protocol that appears to use the full dataset before the evaluation split, which could inflate the reported advantage. The survival results also have overlapping standard deviations, so the empirical support for the headline claims is currently incomplete.","major_comments":[{"comment":"The CpG feature selection is described before the 2-fold split and appears to be performed on the full 1,504-patient dataset: the text states that variance, coefficient of variation, median absolute deviation, and inter-quartile range were intersected to select the 8K most variable CpG sites, and that 4K and 10K alternatives were compared on performance before selecting 8K. No nested cross-validation or outer-loop feature selection is described. If the test fold informed the choice of feature subset or its size, the reported F1-Macro values in Tables I and II are optimistically biased. Because all compared fusion variants share the same CpG features, this bias could either inflate or mask the dual-fusion advantage, and the main claim of the paper is therefore not yet supported by a leakage-free protocol. The authors should either describe a nested CV procedure, fix the feature set a priori based on prior work, or re-run the central ablations with feature selection confined to training folds.","section":"Section IV, subtyping results"},{"comment":"The Wilcoxon rank-sum test reported as p=0.043 is under-specified. It is not stated what the units of comparison are (number of folds, number of runs, per-class F1 values, or per-patient predictions), how many such units were used, or whether the comparison against 'all other multimodal models' involved multiple tests and any correction. With only two cross-validation folds, a rank-sum test would have very low resolution. As written, the sentence 'the test yielded a p-value of 0.043, indicating a statistically significant difference' does not establish significance in a defensible way. The authors should specify the test design, sample size, and correction, or refrain from claiming statistical significance.","section":"Table III and Table IV"},{"comment":"The survival-prediction improvements over baselines are generally within one standard deviation of the baseline values. For example, ABMIL-MOAD-FNet on BLCA achieves 0.691±0.069, while SNN achieves 0.671±0.058 and MoME 0.686±0.041; on BRCA the proposed method's 0.726±0.049 is lower than PIBD-MOAB's 0.749±0.062. The text states 'improved survival prediction on TCGA-BLCA and competitive performance on TCGA-BRCA,' but without statistical tests or a clear demonstration that the differences are not noise, the survival claim is not yet supported. The authors should report confidence intervals or paired significance tests across their five runs, or soften the claim accordingly.","section":"Eq. (1)-(2)"},{"comment":"The early fusion definition has a likely error or redundancy: Eq. (1) defines z_ij = [e_ij, o_i], and Eq. (2) then applies f_E([z_ij, o_i]), which concatenates o_i a second time. If this is a typo, the intended input should be f_E(z_ij) or f_E([e_ij, o_i]); if o_i is intentionally duplicated, the text should explain why. As written, the formulation is ambiguous and affects reproducibility of the central early-fusion module.","section":"Eq. (1)-(2)"}],"minor_comments":[{"comment":"The text says the method is aimed at 'survival prediction in both lung and breast cancer,' but the datasets are TCGA-BLCA (bladder urothelial carcinoma) and TCGA-BRCA (breast invasive carcinoma). This should be corrected to bladder and breast cancer.","section":"Section II-B"},{"comment":"There are two rows both labeled 'MMP [22]' with different c-index values (0.628 and 0.635 for BLCA). The authors should distinguish the re-implemented result from the originally reported value, for example by adding an asterisk or a footnote.","section":"Table III"},{"comment":"The caption reads 'TGGA BRCA AND BLCA DATASETS'; this should be 'TCGA.'","section":"Table IV caption"},{"comment":"The ablation table lists 'ABMIL Early fusion with p_ij' and 'ABMIL - MOAB Late fusion with e_ij,' but the rows do not specify whether the late-fusion row uses MOAB on e_ij-derived slide embeddings or some other aggregation. A brief clarifying note would help readers map the ablations to the model description in Section II-B.","section":"Table II description"},{"comment":"The text says ConvNeXt achieves F1-Macro 0.732±0.012 while Table II reports 0.732±0.012 for 'MOAD-FNet ConvNeXt encoder'; the text says this is 'marginally lower than the pretrained UNI encoder,' but the difference is small relative to the reported standard deviations. This point is not a blocker, but the phrasing overstates the contrast.","section":"Section IV, subtyping results"}],"recommendation":"major_revision","confidential_remarks":"The feature-selection leakage concern in Section II-A is the main threat to the central claim and is fixable in principle (e.g., nested CV or a pre-specified feature set). The paper also builds on the authors' previous MOAB conference paper; while that is not a problem per se, the novelty of the dual-fusion combination would be stronger if the MOAB component were independently evaluated outside its original context, which the ablations in Table IV partially address. The manuscript fits the scope of IEEE TMI and, with the evaluation protocol tightened and the statistical claims corrected, could become a publishable contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: this is a reasonable incremental architecture paper with a genuinely useful new dataset, but the central empirical claim is weakened by a likely data leak in feature selection and by statistical tests that don't actually show what they claim.\n\nWhat's new: combining early fusion (concatenating omics to patch embeddings) with a late MOAB fusion block, applied for the first time to WSI + methylation on a real CNS tumor cohort (NHNN BRAIN UK, 1,504 patients, 20 subtypes). The ablations against early-only, late-only, and different backbones are thorough. The interpretability heatmaps are a nice touch.\n\nWhere it's soft: the big problem is in Section II-A. The 8K CpG set was chosen by comparing 4K, 8K, and 10K performance before the 2-fold split is described, with no nested cross-validation. That means test information likely informed the feature count and subset, inflating every number in Tables I and II. All fusion variants share the features, so the relative dual-vs-single comparison could still be roughly controlled, but the absolute scores and the claim of 'best across all metrics' are not reliable. The Wilcoxon test is under-specified — there's no description of what units go into it, and the actual differences (0.745 vs 0.724, with SDs of 0.025) are within noise. On survival, the c-index improvements overlap with existing baseline standard deviations; on BRCA their model is not the best despite the abstract implying otherwise. Also, the paper overclaims: omics-only SNN has higher precision (0.799) than their 0.769, so 'best across all metrics' is false. Code is not released and data is access-restricted, so reproducibility is limited.\n\nBottom line: the idea of dual fusion is worth reporting, and the BRAIN UK cohort is valuable. But as it stands, the evidence for the central claim is not solid enough to trust the magnitude of the gains. A serious referee should see this, but the paper needs major revision: nest the feature selection inside the CV, report the statistical test properly, fix the overclaims, and release code.\n\nFor whom: researchers working on multimodal pathology fusion, especially those evaluating WSI-omics integration methods. I would not cite it yet, but I'd want to see the revised version.","headline":"Plausible dual-fusion architecture; evaluation leak and weak stats mean the headline gains are not yet credible.","tokens_in":16245,"tokens_out":3123,"would_cite":false,"duration_ms":42844,"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":"Reintroducing the same DNA methylation vector at both patch-level and slide-level fusion stages improves CNS tumor subtyping and survival prediction over either stage alone.","keywords":["dual fusion","whole slide images","DNA methylation","CNS tumor classification","multiple instance learning","survival prediction","multimodal deep learning","outer arithmetic block"],"falsifier":"Run the entire MOAD-FNet pipeline under nested cross-validation, selecting the CpG feature count (4K, 8K, or 10K) inside each training fold only, and compare ABMIL-MOAD-FNet against early and late fusion on the brain-tumor subtyping dataset; if the F1-Macro gap (0.745 versus 0.644 and 0.690) shrinks or reverses, the dual-fusion benefit is partly a feature-selection artifact.","tokens_in":15156,"feed_emoji":"🧬","tokens_out":12624,"duration_ms":102786,"temperature":0.7,"pith_summary":"This paper proposes MOAD-FNet, a dual-fusion network that feeds one DNA methylation embedding into a whole-slide image pipeline twice: first concatenated with each patch embedding and mapped by an MLP before gated-attention pooling, and second reintroduced alongside the pooled slide embedding through an outer-arithmetic block. The central claim is that this two-stage reintroduction captures complementary local and global cross-modal interactions, and that single-stage early or late fusion leaves performance on the table. On a 1,504-patient CNS tumor cohort spanning 20 subtypes, ABMIL-MOAD-FNet reaches F1-Macro 0.745, compared with 0.644 for early fusion and 0.690 for late fusion; on the two public cancer cohorts it reaches c-index 0.691 for bladder cancer and 0.726 for breast cancer. If the claim holds, dual fusion is a straightforward design choice for integrating methylation with histology in precision oncology, with attention maps that make the molecular-morphological evidence visible.","feed_headline":"Dual fusion reaches F1 0.745 for brain-tumor subtyping","feed_subtitle":"Reintroducing methylation at patch and slide levels beats early or late fusion alone in brain-tumor subtyping.","key_machinery":"The load-bearing machinery is the MOAD-FNet architecture and its MOAB fusion block. In the early fusion stage, the omic embedding $o_i \\in \\mathbb{R}^{256}$ is concatenated with each UNI patch embedding $e_{ij} \\in \\mathbb{R}^{1024}$ to form $z_{ij} = [e_{ij}, o_i]$; an MLP $f_E$ maps each concatenation to a joint per-patch embedding $p_{ij}$, and ABMIL gated attention pools these into a slide embedding $v_i \\in \\mathbb{R}^{256}$. In the late fusion stage, MOAB takes $v_i$ and $o_i$, appends constants, and computes outer product, outer division, outer subtraction, and outer addition, producing four $257 \\times 257$ interaction matrices that are concatenated along the channel dimension and condensed by a 2D convolution before the classifier. The appended constants preserve the original unimodal features inside the interaction matrices, which is why the block can richly intermingle modalities without losing either one.","core_discovery":"On the paper's own terms, the discovery is that the omic vector should not be used only once. The authors show that when the same methylation embedding is projected onto every whole-slide patch in latent space and then reintroduced at the slide level through the Multimodal Outer Arithmetic Block, the resulting dual-fusion model outperforms both early-only and late-only variants on fine-grained CNS tumor subtyping, and improves or matches the strongest compared methods on survival prediction. They interpret this as evidence that local patch-level interactions and global slide-level interactions carry complementary information, and that an outer-product family of arithmetic operations captures cross-modal correlations more richly than concatenation or Kronecker-product fusion.","pith_inferences":["Editorial inference: if the dual-fusion advantage is robust, single-stage fusion is systematically underusing the omic modality; a direct test is to repeat the comparison with other omic types such as RNA expression or copy-number data, and on other tumor sites.","Editorial inference: the four outer arithmetic operations form a parameter-light bilinear interaction layer; one can probe whether the gain comes from true cross-modal correlation by replacing the methylation embedding with a permuted or zeroed vector and checking that the dual-fusion gap disappears.","Editorial inference: the paper's Limitations section notes that the identity of decisive CpG sites is hard to recover after early fusion; an attribution analysis on the early-fusion MLP could recover per-CpG importance and make the dual-fusion approach more actionable clinically."],"forward_implications":["ABMIL-MOAD-FNet reports the best subtyping result on the 20-subtype CNS tumor dataset, with F1-Macro 0.745, beating early fusion by 0.101 and late fusion by 0.055 in the ablation.","Dual fusion also improves survival prediction, reaching c-index 0.691 on the bladder cancer cohort, the best among compared methods, and 0.726 on the breast cancer cohort, on par with the strongest baseline.","The MOAB aggregation block consistently outperforms concatenation and Kronecker-product fusion in the survival ablation across ABMIL and TransMIL backbones, under both late and dual fusion settings.","The improvement is attributed to the fusion design rather than the WSI encoder: swapping the UNI encoder for ConvNeXt only lowers F1-Macro from 0.745 to 0.732.","Attention scores from the early fusion stage can be visualized as heatmaps that highlight diagnostically relevant patches, linking molecular signals to morphology.","The same dual-fusion scheme also improves survival prediction on the two public cohorts, with c-index 0.691 on bladder cancer and 0.726 on breast cancer."],"supporting_citations":[{"why":"It supplies the gated-attention ABMIL pooling used in the early fusion stage and as the WSI-only baseline.","marker":"[35]"},{"why":"It defines the survival-prediction task, the pathway tokenization, and the TCGA omic data used for evaluation.","marker":"[13]"},{"why":"It provides the MCAT multimodal early-fusion baseline that the paper adapts by feeding it the early-fusion embeddings.","marker":"[10]"},{"why":"It provides the MMP baseline whose survival results MOAD-FNet matches or outperforms on the two cohorts.","marker":"[22]"},{"why":"It supplies the UNI foundation-model encoder used to extract the 1024-dimensional WSI patch embeddings.","marker":"[11]"},{"why":"It establishes DNA methylation-based CNS tumor classification and the 10K CpG selection reference point.","marker":"[9]"},{"why":"It introduces the PIBD baseline and the argument that single-stage fusion can overlook modality-specific information.","marker":"[20]"},{"why":"It is the earlier paper that introduces the MOAB block, reused here as the late-fusion component.","marker":"[28]"}],"fun_headline_variants":["Dual fusion of histology and methylation improves CNS tumor subtyping","Reintroducing omic data at two stages lifts brain-tumor classification","Patch and slide level methylation fusion outperforms single-fusion models","MOAB fusion with omic embeddings enhances tumor subtyping accuracy","Early and late omic fusion boosts CNS tumor diagnosis performance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains assume the 8K CpG methylation feature set was chosen without using the held-out test folds; the paper compares 4K, 8K, and 10K sites before describing the split and does not mention nested cross-validation, so if test data influenced the feature count, the F1 and c-index improvements are optimistic.","fun_headline_variants_meta":{"raw":{"variants":["Dual fusion of histology and methylation improves CNS tumor subtyping","Reintroducing omic data at two stages lifts brain-tumor classification","Patch and slide level methylation fusion outperforms single-fusion models","MOAB fusion with omic embeddings enhances tumor subtyping accuracy","Early and late omic fusion boosts CNS tumor diagnosis performance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1377,"prompt_tokens":968,"completion_tokens":409,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":318}},"tokens_in":584,"tokens_out":409,"duration_ms":5765,"temperature":1.0,"reasoning_tokens":318,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:07:51.765441+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the entire MOAD-FNet pipeline under nested cross-validation, selecting the CpG feature count (4K, 8K, or 10K) inside each training fold only, and compare ABMIL-MOAD-FNet against early and late fusion on the brain-tumor subtyping dataset; if the F1-Macro gap (0.745 versus 0.644 and 0.690) shrinks or reverses, the dual-fusion benefit is partly a feature-selection artifact.","supporting_citations":[{"cited_title":"Attention-based deep multiple instance learning,","cited_arxiv_id":null,"evidence_quote":"It supplies the gated-attention ABMIL pooling used in the early fusion stage and as the WSI-only baseline."},{"cited_title":"Modeling dense multimodal interactions between biolog- ical pathways and histology for survival prediction,","cited_arxiv_id":null,"evidence_quote":"It defines the survival-prediction task, the pathway tokenization, and the TCGA omic data used for evaluation."},{"cited_title":"Multimodal co-attention trans- former for survival prediction in gigapixel whole slide images,","cited_arxiv_id":null,"evidence_quote":"It provides the MCAT multimodal early-fusion baseline that the paper adapts by feeding it the early-fusion embeddings."},{"cited_title":"Multimodal prototyping for cancer survival prediction,","cited_arxiv_id":null,"evidence_quote":"It provides the MMP baseline whose survival results MOAD-FNet matches or outperforms on the two cohorts."},{"cited_title":"Towards a general-purpose foundation model for computational pathology,","cited_arxiv_id":null,"evidence_quote":"It supplies the UNI foundation-model encoder used to extract the 1024-dimensional WSI patch embeddings."},{"cited_title":"Dna methylation- based classification of central nervous system tumours,","cited_arxiv_id":null,"evidence_quote":"It establishes DNA methylation-based CNS tumor classification and the 10K CpG selection reference point."},{"cited_title":"Prototypical infor- mation bottlenecking and disentangling for multimodal cancer survival prediction,","cited_arxiv_id":null,"evidence_quote":"It introduces the PIBD baseline and the argument that single-stage fusion can overlook modality-specific information."},{"cited_title":"Moab: Multi-modal outer arithmetic block for fusion of histopathological images and genetic data for brain tumor grading,","cited_arxiv_id":null,"evidence_quote":"It is the earlier paper that introduces the MOAB block, reused here as the late-fusion component."}],"review_version":1}