{"id":"3ed2d690-89a3-414d-9d97-98ebbb1cbaf1","arxiv_id":"2506.18042","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CmFNet combines separate CT and MR feature streams with a cross-modal stream and a hybrid scribble, regularization, and consistency loss to segment tumors and organs from scribble annotations, reporting gains over several baselines on an NPC dataset and competitive results on WORD.","lead":"This paper presents CmFNet, a neural network that segments medical images (CT and MR) using only scribble annotations, thin hand-drawn lines instead of full pixel masks. The authors test it on two medical datasets and report that it beats several existing weakly supervised methods and, in one comparison, a fully supervised method.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MR branch is trained with CT-only scribbles and CT/MR geometry is never registered or resampled; Eq. (4) and Eq. (8) presuppose spatial correspondence that is not established.","rationale":"Good-faith reading: CmFNet is a coherent triple-branch V-Net with CFF/CFE fusion and hybrid losses, and the NPC experiments include useful ablations. The strongest quantitative claim is Table I, and for that claim to hold the scribble mask used in Eq. (4) must be valid for the MR branch. The paper says only that scribbles were drawn on CT images (Section IV-A3), and CT/MR differ in matrix size and pixel spacing, so spatial correspondence cannot be assumed. This is the most load-bearing assumption because it is presupposed by both direct MR-branch supervision and the inter-modal consistency loss, yet it is never stated or validated. I agree with the reader's weakest_assumption. The concern is concrete and testable: check the registration/resampling step and rerun with explicit alignment. Secondary concerns, including the lambda typo in Eq. (7), the abstract overclaim on WORD, and the borrowed fully-supervised baseline numbers, should be fixed but do not change the conditional verdict. If the alignment check fails, the central NPC claim collapses; if it passes, the paper is a solid conditional accept.","tokens_in":16507,"tokens_out":11299,"duration_ms":123997,"concrete_test":"Check the NPC preprocessing pipeline for a CT-MR registration or resampling step. Concretely, take the CT-drawn scribble mask from Section IV-A3, apply the affine transform needed to resample MR to CT space, and compute Dice overlap between the CT tumor mask and the MR tumor contour at the scribble voxels. If the overlap is not near-perfect, L_pCE(s, y_mr) in Eq. (4) uses misaligned labels and Eq. (8) propagates that error; rerunning the full model on explicitly registered CT/MR pairs and comparing DSC would settle whether the 76.75 DSC result is reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central weakly-supervised claim on NPC depends on Eq. (4), which applies the same CT-drawn scribble mask s to the MR branch prediction y_mr as L_pCE(s, y_mr). The data description gives different acquisition matrices and pixel spacings for CT (512x512, 1.27 mm) and MR (384x384, 1.30 mm), and Section IV-A3 says only that scribbles were drawn with ITK-SNAP on CT images. No registration, resampling to a common grid, or alignment check is reported. Without a shared coordinate system, the CT scribble labels are not valid for the MR voxels to which Eq. (4) applies them, so y_mr is trained on mislabeled pixels. Because Eq. (8) then enforces L_imc = alpha_1||y_mm - y_ct||^2 + alpha_2||y_mm - y_mr||^2, the cross-modal output used at inference is pulled toward that noisy branch. The reported 76.75 DSC gain over TV (74.55) and DMSPS (71.58) could therefore reflect an undocumented registration step rather than the method as described, or could be unreproducible. A secondary issue: Eq. (7) lists lambda_ct twice instead of lambda_ct and lambda_mr, and the abstract overstates the WORD result since Table II ranks CmFNet second to DMSPS, but those are not the load-bearing failure mode.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes CmFNet, a 3D weakly supervised segmentation network for cross-modal medical images. The architecture consists of two modality-specific branches (CT and MR) and a shared cross-modal branch, combined with CFF and CFE modules. Training uses a hybrid loss: partial cross-entropy on scribbles, a multi-view CRF loss for intra-modal regularization, and an MSE inter-modal consistency loss. Experiments on a clinical NPC dataset (CT+MR) and the WORD CT dataset compare with weakly and fully supervised baselines. The paper reports that CmFNet achieves 76.75±5.78 DSC and 2.69±0.89 mm ASD on NPC with scribbles, outperforming the listed weakly supervised baselines, and 84.19 mean DSC on WORD, ranking second behind DMSPS; with full annotations on NPC, it reports 78.95 DSC, above the fully supervised baselines listed.","tokens_in":16910,"tokens_out":8849,"duration_ms":85328,"significance":"If the results are reproducible, CmFNet would be a useful contribution to scribble-based cross-modal segmentation. The paper has practical strengths: it reports standard deviations, runs ablations of the fusion modules and loss terms, and includes a hyperparameter sensitivity analysis. The architectural motivation of separating modality-specific and shared features is reasonable, and the loss decomposition is clear. However, the current manuscript leaves a critical spatial-correspondence issue unresolved: scribbles are drawn only on CT, yet the same mask supervises the MR branch, and the geometry between CT and MR is not described. This, together with an imported fully-supervised comparison from earlier work, currently limits confidence in the headline numbers.","major_comments":[{"comment":"Scribble supervision for the MR branch requires spatial correspondence that is not established. Section IV-A3 states that scribbles were manually drawn on CT images only, and the CT and MR volumes have different matrices and pixel spacings (512x512, 1.27 mm vs 384x384, 1.30 mm). Eq. (4) nevertheless applies the same mask s to y_mr, and Eq. (8) uses y_mr as a consistency target for the cross-modal output. If the MR volumes are not registered or resampled to the CT grid, the MR branch is trained on misaligned labels and the IMC loss propagates that noise. The Resampler shown in Fig. 2 is never described in the method or implementation sections. The authors must specify the registration/resampling protocol, describe how the scribble mask is transformed, and provide an alignment check or quantitative evidence that the CT-drawn scribbles are valid for the MR branch. Without this, the reported NPC gains in Table I are not reproducible from the paper as written.","section":"Section IV-A3 / Section III-D1"},{"comment":"The fully-supervised comparison is not self-contained. The manuscript says, 'we referred to the experimental results from [25] on the NPC dataset,' but it does not establish that the data splits, preprocessing, training settings, and evaluation metric computations are identical to those used for CmFNet. Since Table III is the basis for the claim that CmFNet outperforms fully supervised methods with full annotations, the authors need to retrain all baselines under the same protocol or provide a detailed equivalence argument. Without this, the 78.95±5.99 DSC result is not directly comparable to the DA-VNet, V-Net, UNet++, and other listed numbers.","section":"Section IV-C3 / Table III"}],"minor_comments":[{"comment":"Eq. (7) lists lambda_ct twice instead of lambda_ct and lambda_mr; since the sentence explicitly refers to both weights, this is a typo, but the equation should be corrected.","section":"Section III-D2"},{"comment":"The abstract states that the approach outperforms state-of-the-art weakly supervised methods without qualification, but on WORD Table II shows CmFNet ranks second overall behind DMSPS in mean DSC (84.19 vs 86.72). Please revise the abstract and conclusion to say that it is competitive with, or second-best to, DMSPS on WORD.","section":"Abstract / Section IV-C2"},{"comment":"The text says Table III presents a qualitative comparison, but Table III reports quantitative DSC/ASD values; use quantitative instead.","section":"Section IV-C3"},{"comment":"Several table cells contain duplicated numbers in the text (for example, 76.75±5.7876.75±5.78 in Table I and 1.44±2.911.44±2.91 in Table II); clean these formatting artifacts before resubmission.","section":"Tables I-VII"},{"comment":"The implementation section names hyperparameters lambda_1, lambda_2, alpha_1, alpha_2, while Eq. (7) and Fig. 6 use lambda_ct and lambda_mr; align the notation.","section":"Section IV-B"},{"comment":"The row labeled FullySup is described as an upper bound, yet Ours with scribbles (76.75 DSC) exceeds it (73.98 DSC). Clarify what architecture and supervision this row uses and avoid the term upper bound unless the model is identical to the one used for the weakly supervised comparisons.","section":"Section IV-C1 / Table I"},{"comment":"Eq. (1) has an unclosed parenthesis after f_i_mr, and the text defines W as weight when width is intended; correct these presentation issues.","section":"Section III-C / Eq. (1)"},{"comment":"On WORD, CT images are fed into both modality-specific branches, so the method is not cross-modal in that setting; state this explicitly and explain how the conclusions transfer to the cross-modal NPC scenario.","section":"Section IV-C2"}],"recommendation":"major_revision","confidential_remarks":"The spatial-correspondence issue is the main gating question; request a precise description of the Resampler and a validation experiment (for example, an alignment error estimate or an ablation with and without registration) before considering acceptance. The fully-supervised comparison should also be made self-contained, ideally by retraining the baselines under the same data split. In addition, the abstract and conclusion should be aligned with the WORD result, which currently ranks second to DMSPS."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Marc, quick take on CmFNet. It's a solid engineering paper: triple-branch V-Net, CFF/CFE modules, hybrid loss (scribble pCE, multi-view CRF, inter-modal MSE). The NPC results are strong—76.75 DSC with scribbles beats the best baseline by a clear margin, and the ablations show each component contributes. Not a conceptual leap, but a legitimate extension of the weak-supervision and cross-modal fusion literature. The WORD experiments also show it degrades gracefully to single-modality input.\n\nThe main problem is a missing procedural detail that undermines the central claim as reported. Scribbles were drawn on CT images only, yet Eq. (4) applies the same mask to the MR branch prediction y_mr. CT and MR have different matrix sizes and pixel spacings (512x512, 1.27mm vs 384x384, 1.30mm). The paper never describes registration, resampling to a common grid, or an alignment check. Unless the volumes are in anatomical correspondence, the MR branch is trained on mislabeled voxels, and Eq. (8) propagates that noise into the cross-modal output used at inference. This could be a simple omission—clinical CT/MR are often rigidly registered during planning—but as written it's a gap a referee cannot check. If I were the editor I'd ask for the preprocessing details before accepting.\n\nThere are a couple of smaller issues. Eq. (7) lists lambda_ct twice; it should be lambda_ct and lambda_mr. The abstract says the method outperforms SOTA on WORD, but Table II shows CmFNet ranks second to DMSPS. And the fully-supervised comparison imports baseline numbers from their earlier DA-VNet paper; that's fine if the split is identical, but the provenance should be explicit.\n\nOverall, the central idea holds up. The reported gains are plausible given the ablations, and the qualitative figures look consistent. The MR-alignment omission is a reporting error, not evidence of fabrication. I'd send this to a serious referee, primarily to get the registration question answered and the abstract fixed. It will be useful to people building cross-modal weak-supervision systems, but it doesn't change how I'd approach the problem.","headline":"A useful cross-modal fusion paper with strong NPC results, but the MR branch is trained on CT-only scribbles with no described registration, so the central claim is hard to verify as written.","tokens_in":17377,"tokens_out":2727,"would_cite":false,"duration_ms":26139,"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 triple-branch network segments tumors from CT and MR using only scribble labels, matching full supervision.","keywords":["weakly supervised segmentation","scribble annotation","cross-modal fusion","medical image segmentation","nasopharyngeal carcinoma","CT-MR imaging","3D segmentation","abdominal organ segmentation"],"falsifier":"Overlay the CT-drawn scribble masks on the corresponding MR volumes and measure their overlap with the tumor boundary visible in MR; if the masks are systematically offset, the MR branch receives corrupted labels and the reported DSC would drop when MR-native scribbles are used instead. A simpler computation: retrain the MR-specific branch with scribbles drawn directly on MR and compare to the current CT-transferred scribbles.","tokens_in":16328,"feed_emoji":"🩻","tokens_out":6793,"duration_ms":64552,"temperature":0.7,"pith_summary":"The paper tries to show that thin scribble lines drawn on medical images—rather than dense voxel-by-voxel labels—are enough to train accurate 3D segmentation networks when two imaging modalities are available. It proposes CmFNet, a triple-branch network with one branch for CT, one for MR, and a shared cross-modal branch that fuses and enhances features from both. A hybrid loss combines partial cross-entropy on scribbled pixels, a multi-view conditional random field regularizer, and mean-squared-error consistency between branch predictions to stop the model overfitting to the few labelled pixels. On a clinical nasopharyngeal carcinoma dataset the scribble-trained model reports 76.75 Dice, above the best weakly supervised baseline and above a fully supervised V-Net; the authors claim it also beats fully supervised methods when given full masks. A sympathetic reader would take the paper's claim to be that cross-modal fusion is the ingredient that lets weak supervision approach full supervision.","feed_headline":"Scribble labels plus cross-modal fusion rival full supervision","feed_subtitle":"On NPC tumors, the method reaches 76.75 Dice with CT-only scribbles and also handles single-modality CT organs.","key_machinery":"The load-bearing mechanism is the triple-branch architecture plus a three-term loss. Two V-Net-style branches process CT and MR separately with different down-sampling depths (three stages for CT, four for MR), preserving detail in one and context in the other. Their feature maps are combined by the Cross-Modal Feature Fusion (CFF) module, a concatenation-plus-convolution, and the Cross-Modal Feature Enhancement (CFE) module, which aligns the third-layer features and increases receptive field before the shared decoder. Supervision is $L_{\\mathrm{total}} = L_{\\mathrm{ssl}} + L_{\\mathrm{imr}} + L_{\\mathrm{imc}}$: partial cross-entropy restricted to scribbled pixels, a multi-view conditional random field loss applied in axial, sagittal, and coronal planes, and mean-squared-error consistency pulling the cross-modal prediction toward each modality branch. The CFF module is deliberately simple; the paper argues the architecture itself, not complex attention, makes fusion effective.","core_discovery":"The paper's central claim is that cross-modal fusion, not a more elaborate attention mechanism, carries most of the benefit in scribble-supervised segmentation. On its clinical NPC dataset, CmFNet with scribble annotations reaches 76.75±5.78 DSC and 2.69±0.89 mm ASD, exceeding the best weakly supervised baseline DMSPS (71.58±7.73 DSC, 4.70±1.81 mm ASD) and the fully supervised V-Net (73.98±7.28 DSC). With full masks the same network reports 78.95±5.99 DSC, ahead of all compared fully supervised methods including DA-VNet. On the single-modal WORD dataset, where the same CT volume is fed into both branches, the method reaches a mean DSC of 84.19, second to the task-specific DMSPS baseline but with stronger surface-distance results on several organs. The authors interpret these results as evidence that modality-specific branches preserve each image type's details while the shared branch aligns them, and that the three-term loss prevents the overfitting that sparse scribble labels usually cause.","pith_inferences":["If the CT/MR alignment assumption holds, the same triple-branch recipe should transfer to other paired modalities such as PET/CT or T1/T2 MRI; the paper only claims future work, this is an extension.","The WORD experiment suggests the cross-modal branch behaves as a feature-sharing regularizer rather than a true modality encoder, since identical CT content is fed into both branches; this interpretation goes beyond the authors' framing.","A testable extension the paper does not report: compare CT-transferred scribbles against MR-native scribbles for the MR branch; a large gap would indicate alignment noise rather than cross-modal learning drives part of the gain."],"forward_implications":["Scribble-only training on paired CT and MR can match or beat a fully supervised V-Net on NPC tumor segmentation, so annotation cost could drop substantially in radiotherapy workflows.","The same network works on single-modal CT data by feeding the same volume into both branches, giving competitive abdominal organ segmentation on WORD (mean DSC 84.19), though below the task-specific DMSPS baseline.","Ablations show each added component contributes: CFE raises DSC from 73.76 to 75.96, adding CFF reaches 76.75, and the hybrid loss raises scribble-only training from 68.20 to 76.75, so the gains are attributed to the fusion and regularization together.","When full masks are available, the same architecture reports 78.95 DSC, ahead of the compared fully supervised methods, suggesting the cross-modal design does not depend on weak supervision to help."],"supporting_citations":[{"why":"V-Net: supplies the backbone for the modality-specific branches and the fully supervised baseline in Table III.","marker":"[24]"},{"why":"DMSPS: the strongest weakly supervised baseline on both NPC and WORD, the method must beat.","marker":"[5]"},{"why":"ScribbleSup: source of the partial cross-entropy loss used for scribble supervision and a baseline.","marker":"[10]"},{"why":"WORD: provides the public CT dataset, split, and scribble annotations for abdominal organ evaluation.","marker":"[31]"},{"why":"DA-VNet: fully supervised cross-modal NPC method used as the main comparison in Table III.","marker":"[25]"},{"why":"ITK-SNAP: the tool used to draw the scribble masks on CT images.","marker":"[33]"},{"why":"PA-Seg: supplies the conditional random field pairwise-energy loss reused as intra-modal regularization.","marker":"[8]"}],"fun_headline_variants":["Cross-modal fusion lifts scribble-supervised segmentation past full supervision","Scribble-supervised cross-modal fusion: small tumors, big gains","Cross-modal net turns scribbles into full-supervision-level accuracy","Weakly supervised with scribbles: cross-modal net beats fully supervised on NPC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"All three branches are trained with the same scribble mask, but the paper says that mask was drawn on CT images only; for the MR branch to receive correct supervision, the CT and MR volumes must be aligned pixel-for-pixel, and the paper does not describe any registration or alignment check.","fun_headline_variants_meta":{"raw":{"variants":["Cross-modal fusion lifts scribble-supervised segmentation past full supervision","Scribble-supervised cross-modal fusion: small tumors, big gains","Cross-modal net turns scribbles into full-supervision-level accuracy","Weakly supervised with scribbles: cross-modal net beats fully supervised on NPC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000889,"raw_usage":{"total_tokens":3891,"prompt_tokens":1053,"completion_tokens":2838,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":669,"completion_tokens_details":{"reasoning_tokens":2761}},"tokens_in":669,"tokens_out":2838,"duration_ms":19569,"temperature":1.0,"reasoning_tokens":2761,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:55:34.409063+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Overlay the CT-drawn scribble masks on the corresponding MR volumes and measure their overlap with the tumor boundary visible in MR; if the masks are systematically offset, the MR branch receives corrupted labels and the reported DSC would drop when MR-native scribbles are used instead. A simpler computation: retrain the MR-specific branch with scribbles drawn directly on MR and compare to the current CT-transferred scribbles.","supporting_citations":[{"cited_title":"V-net: Fully convolutional neural networks for volumetric medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"V-Net: supplies the backbone for the modality-specific branches and the fully supervised baseline in Table III."},{"cited_title":"Dmsps: Dynamically mixed soft pseudo-label supervision for scribble-supervised medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"DMSPS: the strongest weakly supervised baseline on both NPC and WORD, the method must beat."},{"cited_title":"Scribblesup: Scribble- supervised convolutional networks for semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"ScribbleSup: source of the partial cross-entropy loss used for scribble supervision and a baseline."},{"cited_title":"Word: A large scale dataset, bench- mark and clinical applicable study for abdominal organ segmentation from ct image,","cited_arxiv_id":null,"evidence_quote":"WORD: provides the public CT dataset, split, and scribble annotations for abdominal organ evaluation."},{"cited_title":"3d reconstruction-oriented fully automatic multi-modal tumor segmentation by dual attention-guided vnet,","cited_arxiv_id":null,"evidence_quote":"DA-VNet: fully supervised cross-modal NPC method used as the main comparison in Table III."},{"cited_title":"User-guided 3d active contour segmentation of anatomical structures: significantly improved efficiency and reliability,","cited_arxiv_id":null,"evidence_quote":"ITK-SNAP: the tool used to draw the scribble masks on CT images."},{"cited_title":"Pa-seg: learning from point annotations for 3d medical image segmentation using contextual regularization and cross knowledge distillation,","cited_arxiv_id":null,"evidence_quote":"PA-Seg: supplies the conditional random field pairwise-energy loss reused as intra-modal regularization."}],"review_version":2}