{"id":"53b73217-8cfa-46c8-861f-4f7c1215a2e8","arxiv_id":"2412.10946","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A multi-task segmentation model learns all, new, and vanishing MS lesions from heterogeneous MRI datasets using anatomical constraints and lesion-aware augmentation.","lead":"SegHeD+ is a deep learning model that segments multiple sclerosis lesions from brain MRI scans, handling different data formats and three lesion types at once. It combines anatomical constraints and lesion-aware data augmentation and reports strong results on five datasets, though the vanishing-lesion benchmark is simulated.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Vanishing-lesion benchmark is a time-reversal simulation of new lesions; if real vanishing lesions are not symmetric, the VAN results and the LesionMix-driven gains on VAN do not measure clinical vanishing-lesion performance.","rationale":"The reader's weakest_assumption is exactly the concern I identify: VAN is a time-reversal simulation of new-lesion labels, and the paper's Discussion concedes that new and vanishing lesions have different dynamics. This is the single most load-bearing issue because (i) it supports the abstract's 'vanishing lesions' claim, (ii) it drives the largest ablation gain in Table 5, and (iii) it is acknowledged in the text yet still presented as a benchmark. The paper is otherwise a credible engineering contribution: the architecture is clearly described, the losses are mostly defined (modulo the undefined notation in Eq. 2 noted by the reader), the ablations are thorough, and the comparisons to strong baselines are useful. The concern does not warrant rejection, because the authors disclosed the simulation and because the all-lesion and new-lesion results stand on real data. It does warrant a conditional acceptance: the vanishing-lesion results must be reframed as a simulation study, and the claimed 'superior performance in segmenting ... vanishing lesions' must be tempered until validated on real disappearing-lesion data. My recommended adjustment is therefore to keep the reader's CONDITIONAL verdict, with the condition focused on real-data validation of the vanishing-lesion task and removal of the overstatement.","tokens_in":16599,"tokens_out":2603,"duration_ms":18054,"concrete_test":"Obtain or construct a small real longitudinal MS dataset (or a subset of MSSEG-2 with expert-annotated disappearing lesions, or a public longitudinal dataset such as MS2015 with expert re-annotation of lesions present at t1 and absent at t2) and evaluate the SegHeD+ vanishing-lesion head on that held-out data, without any VAN-augmented training data. If the Dice/F1 on real vanishing lesions is markedly lower than the 43.84/46.02 reported on VAN, or if the LesionMix row of Table 5 no longer shows a large gain in the vanishing-lesion column, then the VAN-based numbers overstate the clinical vanishing-lesion capability.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing element is that the model 'simultaneously segment[s] all, new, and vanishing lesions' with 'superior performance', and the ablation (Table 5) attributes a large part of the vanishing-lesion gain to LesionMix (30.97→43.84 on VAN). But VAN is not a real vanishing-lesion dataset: it is created by inverting the timepoints of MSSEG-2 (Section 3.1), so 'vanishing' lesions are literally the new-lesion masks of MSSEG-2 read backwards. This assumes dynamic symmetry between new and vanishing lesions. The authors themselves state in the Discussion that 'Vanishing and newly forming lesions show different lesion dynamics and characteristics [35].' If that is true, the VAN labels do not correspond to clinical disappearance: a lesion that appears between t1 and t2 is not the same biological event as one that disappears, even after time reversal, because of differences in intensity changes, surrounding tissue evolution, and the temporal context of the two scans. Consequently, the reported VAN Dice/F1 numbers, the comparison to SegHeD on VAN, and the LesionMix ablation's vanishing-lesion column all benchmark a synthetic task, not clinical vanishing-lesion segmentation. The central claim of handling heterogeneous real-world data is therefore overstated for the vanishing-lesion task. This is an acknowledged limitation, but it is acknowledged only in the Discussion, while the abstract and contributions present vanishing-lesion segmentation as a demonstrated capability.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces SegHeD+, a multi-task 3D V-Net that jointly segments all, new, and vanishing MS lesions from heterogeneous cross-sectional and longitudinal FLAIR data. The method takes up to four inputs (baseline/follow-up images, a baseline all-lesion label when available, and a white-matter mask) and produces four segmentation heads. Training combines a Dice loss with longitudinal, volumetric, and spatial constraint losses plus LesionMix lesion-level augmentation. The model is evaluated on five datasets (MS2015, MS2016, MSSEG-2, MSSEG-2+, and the simulated VAN) with Dice and lesion-wise F1 scores; ablations study each loss component, augmentation, and dataset composition. The paper claims superior performance across all three segmentation tasks and positions SegHeD+ as a unified heterogeneous-data model.","tokens_in":16957,"tokens_out":7339,"duration_ms":69596,"significance":"If the claims are fully supported, SegHeD+ would be a valuable step toward a single model that exploits diverse, partially annotated MS datasets, and the public MSSEG-2 evaluation with challenge-standard metrics is a notable strength. The paper also includes a careful ablation of loss terms and augmentation, and it honestly acknowledges several limitations, including the simulated nature of the vanishing-lesion dataset. However, load-bearing parts of the method are currently not implementable as written (the logical operations in the constraint losses are undefined for continuous predictions), and the vanishing-lesion claims are based on a time-reversal simulation that the authors themselves note is not biologically symmetric to real lesion disappearance. These issues must be resolved or substantially qualified before the empirical contribution is fully established.","major_comments":[{"comment":"The longitudinal and spatial losses apply element-wise logical AND (∧) and XOR (⊕) to network predictions (denoted p and s) and labels/white-matter masks. Predictions are continuous outputs, typically probabilities or softmax values, and no thresholding or differentiable Boolean surrogate is defined anywhere in the manuscript. As written, these losses are not computable, which blocks reproduction and makes the Table 5 ablation of the anatomical constraints impossible to interpret. Please state explicitly how ∧ and ⊕ act on continuous values (e.g., hard thresholding, min/max fuzzy logic, or a differentiable approximation) and justify that the resulting loss is differentiable.","section":"Section 2.2, Eqs. (2) and (4)"},{"comment":"The vanishing-lesion results are evaluated on VAN, a private dataset created by inverting the timepoints of MSSEG-2 so that new lesions become disappearing lesions. The authors state in the Discussion that 'vanishing and newly forming lesions show different lesion dynamics and characteristics,' which directly undermines the validity of this time-reversal simulation as a clinical vanishing-lesion benchmark. Consequently, the VAN Dice/F1 numbers, the comparison to SegHeD on VAN, and the LesionMix ablation column for VAN do not substantiate the abstract's claim of superior performance in segmenting clinical vanishing lesions. This is an acknowledged limitation, but only in the Discussion; the abstract and contributions present vanishing-lesion segmentation as a demonstrated capability. Please temper these claims or explicitly label all VAN results as a synthetic benchmark.","section":"Section 3.1, VAN dataset; Tables 3 and 5; Discussion"},{"comment":"In the paragraph on new-lesion segmentation, the text states that SegHeD+ achieves a Dice of 50.52% and an F1 of 55.02%, 'outperforming all SOTA and heterogeneous methods,' but the immediately following sentence says it scores 'slightly lower than the top methods tailored specifically for the new-lesion segmentation.' Table 3 confirms that MedICL (50.67) and Basaran et al. (51.06) are higher than SegHeD+ on MSSEG-2. This internal contradiction means the abstract's phrasing of 'superior performance in segmenting all, new, and vanishing lesions' is not supported for new lesions. Please correct the wording in the text and align the abstract and conclusion with the actual results.","section":"Section 3.4, Table 3"},{"comment":"The thresholds α_high=1.2 and α_low=0.8 are said to be determined from the longitudinal MS dataset [2], which is MS2015, and MS2015 is also one of the held-out test sets in Tables 3 and 4. If the thresholds were chosen using the full MS2015 dataset, including subjects later used for testing, the reported MS2015 improvements are circular and optimistic. Please clarify whether the thresholds were set using only the MS2015 training split or a separate validation set, and if not, provide an analysis showing that the reported MS2015 results are not inflated by this choice.","section":"Section 2.2, volumetric constraint"}],"minor_comments":[{"comment":"The text reports MS2015 all-lesion Dice and F1 scores of 76.39% and 77.45%, but Table 3 shows 78.57 and Table 4 shows 79.35 for SegHeD+. Please reconcile the numbers in the text with the tables.","section":"Section 3.4, Tables 3 and 4"},{"comment":"The model input includes the first-timepoint all-lesion label y^t1_a, but the text says this input is not used for predicting s^t1_a. Please clarify how the network architecture prevents the label from leaking into the first-head prediction, since the input channel is present in Eq. (1).","section":"Section 2.1, Eq. (1)"},{"comment":"The notation p is used in Eqs. (2) and (4) but is never explicitly defined as the model prediction. Please define all prediction variables consistently (e.g., s versus p) at their first use.","section":"Section 2.2"},{"comment":"There are minor language issues, including 'segmentated lesions' in the introduction of Section 2.2 and 'we further penalises' in the spatial-constraint paragraph. These should be corrected during revision.","section":"Section 2.2"},{"comment":"The thresholds α_high and α_low are described as annual rates of lesion-volume change, but MSSEG-2 follow-up scans are 1 to 3 years apart. Please state explicitly whether the thresholds were adjusted for non-annual intervals in the experiments.","section":"Section 3.2 and Section 2.2"},{"comment":"The paper does not include a code or data availability statement. The MSSEG-2+ annotations and the VAN dataset are private, so the multi-dataset results cannot be independently reproduced; please add an explicit statement about what will be released, or describe the annotation/protocol in sufficient detail for replication.","section":"Global"}],"recommendation":"major_revision","confidential_remarks":"The paper builds on a reasonable and well-motivated idea, and the public MSSEG-2 results are the strongest part of the evidence. My main concerns are technical reproducibility (the undefined Boolean operators in the losses) and the overstatement of the vanishing-lesion results relative to the simulated VAN dataset. Both are fixable without a fundamentally new experimental campaign. I do not see grounds for rejection, but the authors should be asked to correct the internal contradiction in the new-lesion paragraph, clarify the threshold selection for α_high and α_low, and temper the abstract and conclusion. If the private annotations cannot be released, a detailed dataset description would still help."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"SegHeD+ is a credible extension of the authors' earlier SegHeD work. The core idea—train one V-Net on heterogeneous cross-sectional and longitudinal datasets, with four outputs for all, new, and vanishing lesions—is sensible, and the ablations show each added component (longitudinal, volumetric, spatial losses, white matter mask, LesionMix) contributes. The public MSSEG-2 new-lesion results (50.52 Dice) are close to task-specific methods and above most SOTA baselines, and the all-lesion gains on MS2015/MS2016 are consistent with the method. The temporal-consistency analysis is a nice addition.\n\nThe main soft spot is the vanishing-lesion benchmark. VAN is made by time-reversing MSSEG-2: new lesions at t2 become 'vanishing' lesions at t1. That simulates disappearance from a single pair of scans, but real vanishing lesions reflect a different clinical and biological process, as the authors themselves note in the Discussion. The abstract's 'superior performance' claim for vanishing lesions is therefore stronger than the evidence supports. The LesionMix-driven jump on VAN (30.97 to 43.84) is real for this synthetic task, but it says little about clinical vanishing-lesion segmentation. This should be reframed as a synthetic benchmark, with the limitation moved up from the Discussion to the abstract and contributions.\n\nSecond, the loss equations use logical AND and XOR on continuous network outputs without defining them. The intended element-wise operations are probably multiplication for AND and absolute difference for XOR, but as written the equations are not reproducible. That is an easy fix.\n\nThird, the volumetric thresholds α_high=1.2 and α_low=0.8 are estimated from MS2015, the same dataset used for evaluation. The sensitivity study helps, but it remains a tuned-on-test element for that dataset.\n\nMinor: MSSEG-2+ and VAN are private; code is not released. Not fatal, but hurts reproducibility.\n\nOverall, the central multi-task claim holds up for all and new lesions. The vanishing-lesion contribution is novel but provisional. The paper is clearly written, cites prior work fairly, and the results on public data are independently grounded. The issues are mechanical and disclosure-related, not signs of fabrication. I would send this to peer review, expecting a major revision that clarifies the loss definitions, tempers the vanishing-lesion claims, and ideally releases the derived datasets or code.","headline":"Solid multi-task extension of SegHeD for MS lesion segmentation, but the vanishing-lesion benchmark is a simulated time-reversal task and is oversold; the public all/new lesion results are credible.","tokens_in":17418,"tokens_out":2826,"would_cite":true,"duration_ms":26229,"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":"SegHeD+ claims that a single multi-task 3D V-Net can simultaneously segment all, new, and vanishing MS lesions from heterogeneous datasets, and that this unified model outperforms task-specific state-of-the-art methods on all-lesion…","keywords":["multiple sclerosis","lesion segmentation","heterogeneous data","multi-task learning","longitudinal MRI","vanishing lesions","new lesions","LesionMix augmentation"],"falsifier":"Train SegHeD+ on a real longitudinal MS dataset with expert-annotated disappearing lesions and compare VAN-trained vs real-trained performance on those real cases; also directly measure whether time-reversed new-lesion masks overlap with true vanishing-lesion masks. If the overlap is low or performance drops, the simulation assumption fails.","tokens_in":16379,"feed_emoji":"🧠","tokens_out":4402,"duration_ms":36302,"temperature":0.7,"pith_summary":"The paper tries to show that one segmentation model can learn from MS brain MRI datasets that differ in format (single scan vs. longitudinal) and in what is annotated (all lesions, new lesions, or vanishing lesions), and can simultaneously produce all three lesion maps. The intended payoff is that clinical and research data no longer need to be harmonized into one annotation protocol before they are useful. SegHeD+ combines a multi-head 3D V-Net with three anatomical loss terms — longitudinal, volumetric, and spatial constraints — and a lesion-level augmentation method called LesionMix that synthesizes extra training examples, including rare vanishing lesions. On five datasets it reports higher Dice and F1 scores than task-specific state-of-the-art methods for all-lesion segmentation, comparable performance for new-lesion segmentation, and a first benchmark for vanishing lesions.","feed_headline":"One model segments all, new, and vanishing MS lesions","feed_subtitle":"Trained on five mixed-format datasets, SegHeD+ beats task-specific rivals for all-lesion and vanishing-lesion segmentation.","key_machinery":"The load-bearing mechanism is a four-headed 3D V-Net trained with a composite loss. Three of the loss terms encode anatomical knowledge: a longitudinal constraint enforces that new lesions are absent at baseline and present at follow-up while vanishing lesions satisfy the reverse; a volumetric constraint penalizes total lesion volume changes beyond thresholds $\\alpha_{\\text{high}}$ and $\\alpha_{\\text{low}}$ (1.2 and 0.8 for annual change); and a spatial constraint penalizes predictions outside the white matter mask, which is also fed as an input channel. Around this, LesionMix performs lesion populating and inpainting to synthesize new and vanishing lesions, rebalancing the scarce vanishing-lesion labels and enlarging the training set from 144 to 400 images.","core_discovery":"SegHeD+ is a single multi-task segmentation model that takes up to four inputs — baseline and follow-up FLAIR scans, an all-lesion label for the baseline scan when available, and a white matter mask — and outputs four segmentations: all lesions at each timepoint, new lesions at follow-up, and vanishing lesions at follow-up. The method's key claim is that combining heterogeneous training data with constraints derived from MS domain knowledge makes the model generalize better than task-specific models trained on homogeneous data. The paper reports that SegHeD+ outperforms the compared state-of-the-art methods on all-lesion segmentation for both longitudinal (MS2015: 78.57 Dice) and cross-sectional (MS2016: 85.18 Dice) benchmarks, achieves the best new-lesion Dice among heterogeneous methods (MSSEG-2: 50.52), and sets a vanishing-lesion benchmark at 43.84 Dice on the synthesized VAN dataset. It also shows temporal consistency: predicted lesion volumes track ground truth across four timepoints with higher Pearson correlation than baselines.","pith_inferences":["The VAN dataset treats 'vanishing' as exact time-reversal of 'new,' an assumption the paper itself flags as questionable; a real longitudinal dataset with actual disappearing-lesion annotations would be the decisive test and would likely change both training signal and benchmark numbers.","Because LesionMix is the main driver of vanishing-lesion gains, increasing the fraction of inpainted (vanishing) examples or using real vanishing-lesion patches could close the remaining gap between new and vanishing Dice scores.","The multi-task, heterogeneous-data recipe is not MS-specific; the same input-channel and constraint pattern could be carried over to other diseases where lesions appear, disappear, or change volume across timepoints (e.g., stroke or tumor follow-up).","The reported 'superior' claim is task-dependent: for new-lesion segmentation on MSSEG-2, some dedicated methods still score higher; the paper's own framing is that a unified model trades a small new-lesion gap for gains across all other tasks."],"forward_implications":["A single SegHeD+ model can replace separate models for cross-sectional, longitudinal, new-lesion, and vanishing-lesion segmentation, so heterogeneous multi-site data can be pooled without re-annotation.","The anatomical constraints improve temporal consistency of lesion volume estimates, which matters for tracking disease progression across follow-up scans.","LesionMix augmentation is what mainly lifts vanishing-lesion segmentation (from 35.23 to 43.84 Dice) and could be applied to other rare lesion types.","The method sets a reference benchmark for vanishing-lesion segmentation, a task previously unaddressed by public challenges.","The baseline all-lesion label is an optional input channel, so the model still works when such labels are missing from a dataset."],"supporting_citations":[{"why":"Supplies the nnU-Net baseline and self-configuring segmentation framework that SegHeD+ is compared against and extends with heterogeneous input handling.","marker":"[13]"},{"why":"Supplies LesionMix, the lesion-level augmentation method for populating and inpainting lesions.","marker":"[9]"},{"why":"Supplies CoactSeg, the heterogeneous-data baseline that SegHeD+ extends with vanishing-lesion outputs and volumetric/spatial constraints.","marker":"[15]"},{"why":"Is the prior SegHeD method; the comparison shows the gain from LesionMix augmentation.","marker":"[26]"},{"why":"Provides the MS2015 longitudinal dataset with all-lesion labels and informs the volume-change thresholds.","marker":"[2]"},{"why":"Provides the MS2016 cross-sectional dataset and the lesion-wise F1 evaluation protocol.","marker":"[3]"},{"why":"Provides the MSSEG-2 new-lesion challenge dataset, expert baselines, and evaluation metrics.","marker":"[4]"},{"why":"Supplies SynthSeg for extracting white matter masks robustly in the presence of lesions.","marker":"[27]"},{"why":"Supplies the V-Net backbone architecture used by SegHeD+.","marker":"[28]"}],"fun_headline_variants":["SegHeD+ beats task-specific models on all MS lesions","One network for all, new, and vanishing MS lesions","SegHeD+: unified MS segmentation from mixed data","Single model excels across MS lesion benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that a vanishing lesion can be faithfully simulated by reversing the timepoints of a new-lesion dataset, so that lesions appearing at the second scan become lesions disappearing from the first scan.","fun_headline_variants_meta":{"raw":{"variants":["SegHeD+ beats task-specific models on all MS lesions","One network for all, new, and vanishing MS lesions","SegHeD+: unified MS segmentation from mixed data","Single model excels across MS lesion benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1463,"prompt_tokens":971,"completion_tokens":492,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":429}},"tokens_in":587,"tokens_out":492,"duration_ms":5074,"temperature":1.0,"reasoning_tokens":429,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:27:32.777682+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train SegHeD+ on a real longitudinal MS dataset with expert-annotated disappearing lesions and compare VAN-trained vs real-trained performance on those real cases; also directly measure whether time-reversed new-lesion masks overlap with true vanishing-lesion masks. If the overlap is low or performance drops, the simulation assumption fails.","supporting_citations":[{"cited_title":"nnU-Net: a self-configuring method for deep learning-based biomed- ical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the nnU-Net baseline and self-configuring segmentation framework that SegHeD+ is compared against and extends with heterogeneous input handling."},{"cited_title":"LesionMix: A lesion-level data augmentation method for medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies LesionMix, the lesion-level augmentation method for populating and inpainting lesions."},{"cited_title":"CoactSeg: Learning from heterogeneous data for new multiple sclerosis lesion segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies CoactSeg, the heterogeneous-data baseline that SegHeD+ extends with vanishing-lesion outputs and volumetric/spatial constraints."},{"cited_title":"Longitudinal multiple sclerosis lesion segmentation: Resource and challenge,","cited_arxiv_id":null,"evidence_quote":"Provides the MS2015 longitudinal dataset with all-lesion labels and informs the volume-change thresholds."},{"cited_title":"Objective evaluation of multiple sclerosis lesion segmentation using a data management and processing infrastructure,","cited_arxiv_id":null,"evidence_quote":"Provides the MS2016 cross-sectional dataset and the lesion-wise F1 evaluation protocol."},{"cited_title":"MICCAI 2021 MSSEG-2 challenge quantitative results,","cited_arxiv_id":null,"evidence_quote":"Provides the MSSEG-2 new-lesion challenge dataset, expert baselines, and evaluation metrics."},{"cited_title":"SynthSeg: Segmentation of brain MRI scans of any contrast and resolu- tion without retraining,","cited_arxiv_id":null,"evidence_quote":"Supplies SynthSeg for extracting white matter masks robustly in the presence of lesions."},{"cited_title":"V-net: Fully convolutional neural networks for volumetric medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the V-Net backbone architecture used by SegHeD+."}],"review_version":1}