{"id":"90927145-5e75-4968-9688-5f6f51cbe4f7","arxiv_id":"2505.11921","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DC-Seg trains a single model that aligns anatomical features across MRI modalities with contrastive learning and adds a per-modality segmentation regularizer, yielding higher Dice scores when modalities are missing.","lead":"A brain-tumor segmentation method that separates MRI images into shared anatomy and modality-specific style performs better when some MRI contrasts are missing. It beats previously published unified models on BraTS 2020 and also transfers to white-matter hyperintensity data, though the comparison is based on quoted numbers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's SOTA comparison is not protocol-matched: baselines are quoted from prior papers, so average Dice gaps of ~0.6 to 1 point may reflect preprocessing or training differences rather than the method.","rationale":"The paper's method-level design is coherent: the bidirectional contrastive losses, reconstruction, and segmentation regularizer are well specified, and the ablation in Table 2 shows each component contributes. The code is released, which supports reproducibility of DC-Seg itself. The load-bearing weakness is not inside the method but in the empirical comparison: Table 1 transcribes baseline numbers from papers that were not run under the same protocol, and no variance or significance is reported. The reader identified exactly this in the weakest_assumption. I agree with the conditional verdict: the central SOTA claim should not be accepted as fully established until the baselines are rerun under identical conditions. This is a benchmark-comparability concern rather than an internal inconsistency or a disagreement with field consensus.","tokens_in":8727,"tokens_out":7384,"duration_ms":75474,"concrete_test":"Retrain RobustSeg, RFNet, mmFormer, and M3AE using the released codebases under the exact DC-Seg protocol on the BraTS2020 split of [5] (same preprocessing, 112^3 patches, augmentation, Adam settings, 500 epochs, same evaluation script), with at least three random seeds per model. Compare per-combination Dice and the 15-combination averages; if the average Complete/Core/Enhancing gaps over RFNet or M3AE fall below ~0.5 Dice or do not reach significance in a paired test, the SOTA claim in Table 1 is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is in Section 3: DC-Seg 'significantly outperforms the state-of-the-art methods' on most of the 15 modality combinations. Table 1 supports this by directly referencing baseline numbers from RobustSeg, RFNet, mmFormer, and M3AE under the assertion that 'we use the same data split as [5]'. That assertion is load-bearing: every headline comparison in the abstract and conclusion depends on cross-paper equivalence. The premise is insecure. The original papers use different preprocessing, patch sizes, training budgets, augmentation, and evaluation details; a shared data-split label does not equal an identical protocol. Table 1 reports single Dice values with no error bars or significance tests, and several entries are close (full-modality Complete: RFNet 91.11 vs DC-Seg 90.95; average Complete gap over RFNet is only 0.56 Dice). The WMH comparisons in Table 3 are likewise reported without saying whether baselines were retrained on the private SAHZU-WMH split or with what hyperparameters. If the quoted baselines are not reproducible under DC-Seg's protocol, the claimed superiority may shrink below noise or reverse. The method-level ablation in Table 2 is coherent, but it does not by itself establish the headline SOTA claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents DC-Seg, a missing-modality brain tumor segmentation method that disentangles each input modality into a modality-invariant anatomical representation and a modality-specific representation. It trains these with a sigmoid-based anatomical contrastive loss (Eq. 1), a modality contrastive loss (Eq. 4), a reconstruction loss (Eq. 5), and a per-modality segmentation regularizer (Eq. 7). Experiments on BraTS 2020 and a private WMH dataset compare against RobustSeg, RFNet, mmFormer, M3AE, and GSS. The component ablation in Table 2 supports the usefulness of each loss term, but the state-of-the-art comparisons in Table 1 and Table 3 are based on directly quoted baseline numbers rather than protocol-matched reruns, and no variance or significance information is reported.","tokens_in":8971,"tokens_out":5830,"duration_ms":57606,"significance":"If the empirical claims hold, the paper contributes a clean and plausible combination of bidirectional contrastive disentanglement with reconstruction and a segmentation regularizer. The code release and the component-wise ablation are strengths, and the idea of aligning anatomical representations across modalities while separating modality-specific ones is a reasonable inductive bias for missing-modality segmentation. However, the headline claim that DC-Seg significantly outperforms state-of-the-art methods currently rests on cross-paper baseline comparisons, so the scientific significance cannot be fully assessed until the comparison protocol is controlled and statistical variability is reported.","major_comments":[{"comment":"The claim that DC-Seg 'significantly outperforms the state-of-the-art methods' is based on directly referencing baseline Dice values from prior papers under the assertion 'we use the same data split as [5]'. A shared split label does not imply equal preprocessing, patch size, training budget, augmentation, or evaluation protocol. The full-modality Complete row illustrates the risk: RFNet is quoted at 91.11 and DC-Seg at 90.95, and the average Complete gap over RFNet is only 0.56 Dice. The same protocol concern applies to the GSS numbers quoted in the text (Complete 87.33, Core 79.38, Enhancing 65.54), which are not even shown in Table 1. Unless the baselines are rerun under the authors' protocol or the comparisons are otherwise demonstrated to be protocol-matched, the headline SOTA claim is not supported.","section":"Section 3, Table 1"},{"comment":"The WMH comparison does not state whether RobustSeg, RFNet, and mmFormer were retrained on the private SAHZU-WMH split or whether their numbers are quoted from BraTS experiments. WMH segmentation involves different modalities and lesion characteristics, so either choice needs explicit protocol details (training split, preprocessing, hyperparameters, augmentation) before the generalization claim in Section 3 can be evaluated.","section":"Section 3, Table 3"},{"comment":"No standard deviations, confidence intervals, or significance tests are reported anywhere. In Table 2, the ablation gaps (e.g., Complete 86.31 without anatomical contrastive learning versus 87.54 for the full model) may be within run-to-run noise, especially with a batch size of 2 and 500 epochs. The paper should report multiple runs and statistical testing before concluding that each component contributes and that the method is superior to the baselines.","section":"Section 3, Tables 1 and 2"},{"comment":"The central disentanglement loss is not fully reproducible because several key hyperparameters are unspecified: the temperature t in Eq. (1) and Eq. (4), the SSIM constants C1 and C2 in Eq. (3), and the modality-dropout probability for the Bernoulli indicator delta in Eq. (5). Section 3 gives only alpha = 0.4. These values should be reported in the implementation details.","section":"Section 2.1, Eqs. (1), (3), (5)"}],"minor_comments":[{"comment":"The SSIM computation on 3D feature maps needs the window size, stride, and assumed data range to be specified, since these choices affect the contrastive loss values and hence the learned representations.","section":"Section 2.1, Eq. (3)"},{"comment":"The phrase '8-bit vector' for the modality representation is confusing; if the intended meaning is an 8-dimensional vector, the notation m_i^j in R^C should be adjusted accordingly.","section":"Section 2.1"},{"comment":"The augmentation description 'Random flips, cropping, and intensity shifts' does not give ranges for these augmentations; specifying them would improve reproducibility.","section":"Section 3"},{"comment":"MedSAM is compared only qualitatively with a bounding-box prompt; reporting a quantitative Dice score for MedSAM would make the dismissal of the foundation-model baseline more convincing.","section":"Section 3, Figure 2"},{"comment":"The text reports GSS numbers in prose rather than in Table 1; adding GSS as a separate row or table would clarify the comparison and make the protocol details easier to verify.","section":"Section 3"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the paper is within scope for a medical imaging venue and the proposed method is promising, but the current form overstates the SOTA claim. The central fix is to make Table 1 and Table 3 protocol-matched, either by rerunning the baselines under the authors' protocol or by reporting variance and significance, and to specify the missing hyperparameters of the contrastive losses. The novelty relative to RFNet is moderate; the main differentiator is the bidirectional contrastive loss applied at the 3D representation level."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read on DC-Seg. The core idea is new: instead of only reconstructing images from fused anatomical features (like RobustSeg or Ouyang), they add two contrastive losses on the full 3D representations -- anatomical contrastive pulls same-subject modalities together via SSIM, modality contrastive pulls same-modality features together via cosine. The sigmoid loss handles the multi-positive setting. That combination I haven't seen in the cited disentanglement papers. The ablation in Table 2 is coherent and shows each component helps, so the method-level claim is plausible.\n\nWhere I'd push back is the headline comparison. The stress-test note is right: Table 1 quotes baseline Dice from RobustSeg, RFNet, mmFormer, M3AE and says \"same data split as [5]\" but that doesn't make the protocols identical. Preprocessing, patch size, training budget, augmentation, evaluation details differ across those papers. Some entries are close -- full-modality Complete: DC-Seg 90.95 vs RFNet 91.11, so DC-Seg is actually lower there; average Complete gap over RFNet is 0.56 Dice. Without error bars or significance tests, \"significantly outperforms\" is not warranted. The WMH table has the same issue: baselines are quoted, not rerun, and the dataset is private.\n\nMinor: the temperature t in the contrastive losses is never given. SSIM constants C1 and C2 are unspecified but presumably standard. Alpha is 0.4.\n\nSo the method is worth knowing and the disentanglement scheme is a real contribution to the missing-modality segmentation subfield. But the empirical evidence for superiority over baselines is weaker than the abstract suggests. The ablation across 15 modality combinations is a decent internal check. I'd send it to peer review, but ask the authors to rerun at least RFNet and one other baseline under their own protocol, report variance (e.g., repeated runs), and state t and SSIM constants. If they can't rerun, they should soften the superiority claims.\n\nWho reads this? People working on missing-modality medical imaging. It's not a breakthrough, but it's a legitimate step. I'd take it to a reading group to discuss the protocol-matching problem.\n\nRecommendation: engage with it, revise, and re-evaluate. Accept with major revisions.","headline":"New disentanglement idea with a solid ablation, but the SOTA comparison rests on quoted baselines and needs protocol-matched reruns.","tokens_in":9517,"tokens_out":2255,"would_cite":false,"duration_ms":21426,"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":"Explicitly disentangling MRI anatomy from scan-specific style improves brain tumor segmentation when modalities are missing.","keywords":["brain tumor segmentation","missing modalities","multimodal MRI","contrastive learning","disentangled representation","BraTS 2020","white matter hyperintensity","modality-invariant anatomy"],"falsifier":"Retrain RobustSeg, RFNet, mmFormer, M3AE, and DC-Seg on the same BraTS 2020 split with identical preprocessing, patch size, augmentation, epochs, and evaluation script; if DC-Seg no longer leads on most of the 15 modality combinations in whole-tumor, core, and enhancing Dice, the claimed superiority is an artifact of comparison protocol. A cheaper version is to retrain just RobustSeg and RFNet under DC-Seg's exact training protocol and compare with Table 1.","tokens_in":8522,"feed_emoji":"🧠","tokens_out":8598,"duration_ms":82018,"temperature":0.7,"pith_summary":"DC-Seg claims that encoding all available MRI modalities into one shared latent space is the wrong fix for missing-modality brain tumor segmentation, because each sequence carries information the others do not. The method separates every scan into a modality-invariant anatomical representation and a modality-specific code, with two contrastive losses: one pulls anatomical maps of the same patient across modalities together, the other pulls modality codes of the same sequence type across patients together. Reconstruction of the original scans and a regularizer that forces each single modality to segment on its own complete the training. On BraTS 2020 the paper reports higher Dice than the unified-model baselines RobustSeg, RFNet, mmFormer, and M3AE on most of the 15 possible modality combinations, and it reports transfer gains on a private white matter hyperintensity dataset. If correct, a single trained model covers every missing-modality scenario at accuracy comparable to training one dedicated model per scenario.","feed_headline":"Two contrastive losses beat missing-MRI tumor segmentation baselines","feed_subtitle":"A single DC-Seg model separates anatomy from scan style, matching per-scenario models on missing scans.","key_machinery":"The machinery is bidirectional contrastive learning on full 3D image feature maps. The anatomical contrastive loss $\\mathcal{L}_{\\mathrm{ana}}$ uses channel-wise SSIM to pull together the anatomical representations $a^i_j$ of the same subject across modalities and push apart representations from different subjects; the modality contrastive loss $\\mathcal{L}_{\\mathrm{mod}}$ uses cosine similarity on the 8-element modality codes $m^i_j$ to pull together codes of the same sequence type across subjects and push apart codes of different sequences. A reconstruction loss with Bernoulli modality dropout keeps both subspaces informative enough to re-render any input modality, and a weight-shared decoder regularizer forces every single-modality anatomical map to predict the segmentation alone. Together these make the fused anatomical representation $z_i$ robust to arbitrary subsets of missing modalities.","core_discovery":"The paper's central claim is that bidirectional contrastive disentanglement is a better inductive bias for incomplete multimodal segmentation than a shared latent space. Anatomical contrastive learning reinforces that the same anatomy viewed through T1, T1ce, T2, and FLAIR lands on the same feature map, while modality contrastive learning enforces that each sequence's style code is consistent across subjects. The reconstruction branch keeps both subspaces informative enough to regenerate any input modality, and the segmentation regularizer stops the fusion path from leaning on the discriminative sequences T1ce and FLAIR. The reported result is an average Dice of 87.54 on whole tumor, 79.63 on tumor core, and 65.00 on enhancing tumor over all 15 missing-modality combinations, close to the dedicated group-self-support baseline that trains 15 separate models, and a 75.43 average Dice on the two-modality white matter hyperintensity task.","pith_inferences":["A matched-protocol rerun under identical preprocessing, augmentation, training schedule, and evaluation code would determine how much of the reported gap over the prior baselines is the method itself rather than inherited published numbers; the paper's comparison quotes previous results directly.","The same anatomical-versus-modality split could apply to other paired multimodal inputs where one channel carries style, such as CT and PET or RGB and depth, provided a reconstruction branch can be defined.","Varying the dimensionality of the 8-element modality code while holding everything else fixed would test whether the bottleneck is what enforces the separation between anatomical and modality-specific information."],"forward_implications":["A single DC-Seg model can replace the 15 dedicated models of the group-self-support approach with roughly unchanged whole-tumor Dice, cutting memory and training cost for missing-modality deployment.","The per-modality regularizer makes each sequence alone produce a usable segmentation, so losing discriminative modalities such as T1ce or FLAIR causes graceful rather than catastrophic degradation.","The method needs no generative model to synthesize absent sequences at inference time, removing the extra training costs of modality-synthesis pipelines.","The same loss design transfers from four-modal tumor imaging to two-modal white matter hyperintensity imaging, suggesting the disentanglement is not tuned to a particular set of MR contrasts."],"supporting_citations":[{"why":"Supplies the data split that all baseline results are quoted on, the fusion used to build z_i, and the region-aware fusion baseline RFNet that DC-Seg must beat.","marker":"[5]"},{"why":"Provides the reconstruction-based disentanglement idea and the 8-element modality code convention that DC-Seg adopts for modality-specific representations.","marker":"[10]"},{"why":"Supplies the sigmoid-based contrastive loss formulation that both anatomical and modality contrastive losses are built from.","marker":"[23]"},{"why":"The RobustSeg baseline, which also disentangles content and appearance by reconstruction and is the main shared-latent competitor for the missing-modality comparison.","marker":"[3]"},{"why":"The mmFormer Transformer baseline for incomplete multimodal learning whose published numbers DC-Seg compares against.","marker":"[25]"},{"why":"The M3AE multimodal autoencoder baseline, a large pre-trained unified model that DC-Seg reports surpassing on all three tumor subregions.","marker":"[11]"},{"why":"The group-self-support baseline that trains one model per missing-modality combination; its performance is the target DC-Seg matches with a single model.","marker":"[16]"},{"why":"The BraTS benchmark, source of the 369 multimodal scans and tumor subregion labels used for the main experiments.","marker":"[14]"}],"fun_headline_variants":["Two contrastive losses untangle missing-MRI tumor segmentation","DC-Seg: split brain MRI into anatomy and style for missing scans","Disentangle anatomy from scan style to segment tumors without full MRI","Missing MRI? DC-Seg separates anatomy and style to keep segmentation strong","Bidirectional contrastive learning boosts tumor segmentation with missing modalities"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported advantage over earlier methods assumes that Dice values quoted from their papers are directly comparable to DC-Seg's own runs because all use the same data split as RFNet; different preprocessing, augmentation, training budgets, or evaluation code would make the gaps protocol effects rather than method effects.","fun_headline_variants_meta":{"raw":{"variants":["Two contrastive losses untangle missing-MRI tumor segmentation","DC-Seg: split brain MRI into anatomy and style for missing scans","Disentangle anatomy from scan style to segment tumors without full MRI","Missing MRI? DC-Seg separates anatomy and style to keep segmentation strong","Bidirectional contrastive learning boosts tumor segmentation with missing modalities"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000256,"raw_usage":{"total_tokens":1578,"prompt_tokens":950,"completion_tokens":628,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":539}},"tokens_in":566,"tokens_out":628,"duration_ms":5914,"temperature":1.0,"reasoning_tokens":539,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:44:57.536424+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain RobustSeg, RFNet, mmFormer, M3AE, and DC-Seg on the same BraTS 2020 split with identical preprocessing, patch size, augmentation, epochs, and evaluation script; if DC-Seg no longer leads on most of the 15 modality combinations in whole-tumor, core, and enhancing Dice, the claimed superiority is an artifact of comparison protocol. A cheaper version is to retrain just RobustSeg and RFNet under DC-Seg's exact training protocol and compare with Table 1.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF inter- national conference on computer vision","cited_arxiv_id":null,"evidence_quote":"Supplies the data split that all baseline results are quoted on, the fusion used to build z_i, and the region-aware fusion baseline RFNet that DC-Seg must beat."},{"cited_title":"IEEE Transactions on Medical Imaging41(10), 2953–2964 (2022)","cited_arxiv_id":null,"evidence_quote":"Supplies the sigmoid-based contrastive loss formulation that both anatomical and modality contrastive losses are built from."},{"cited_title":"In: Medical Im- age Computing and Computer Assisted Intervention–MICCAI 2019: 22nd Inter- national Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part III","cited_arxiv_id":null,"evidence_quote":"The RobustSeg baseline, which also disentangles content and appearance by reconstruction and is the main shared-latent competitor for the missing-modality comparison."},{"cited_title":"IEEE Transactions on Image Processing29, 9032–9043 (2020)","cited_arxiv_id":null,"evidence_quote":"The mmFormer Transformer baseline for incomplete multimodal learning whose published numbers DC-Seg compares against."},{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"The group-self-support baseline that trains one model per missing-modality combination; its performance is the target DC-Seg matches with a single model."}],"review_version":1}