{"id":"3cd2f8d8-c17a-40a2-be7a-fb0b310af5df","arxiv_id":"2505.09829","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"With only four labeled volumes, adding boundary prediction as an auxiliary task raises left atrium segmentation Dice from 0.587 to 0.774, exceeding semi-supervised baselines that have access to only four unlabeled volumes.","lead":"BoundarySeg adds a boundary-prediction branch to a medical image segmentation network and trains it with extra loss terms, reporting higher Dice on left atrium MRI when only 4 or 8 labeled volumes are available. The paper positions this as a low-data alternative to semi-supervised methods, but the loss equations and the forward-pass trick do not match the prose.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Forward-pass-only (FP) gains in Table 1 have no gradient path in a V-Net without normalization; the reported +FP improvements are mechanistically unexplained and need an ablation or code check.","rationale":"The reader's conditional verdict is appropriate. I agree with the reader's weakest assumption: the forward-pass-only mechanism is the least secure component of the paper. The paper's own text provides the contradiction: unlabeled volumes contribute to feature extraction but do not influence gradient backpropagation through the loss terms. In a V-Net without normalization, forward activations from samples with zero loss do not alter parameters, so Table 1's +FP gains lack a mechanism. This is not a disagreement with external consensus; it is an internal inconsistency that can be resolved by code inspection or a targeted ablation. The main BoundarySeg result—auxiliary boundary prediction without unlabeled data—could survive even if the FP extension is refuted, so I would not reject the paper. However, the absence of released code, the lack of error bars, and the additional inconsistency between Eq. 5 (Dice consistency loss) and the implementation text ('We used MSE loss as consistency loss in Eq 6') mean the reported numbers cannot currently be independently reproduced. Conditional acceptance is the right call: the central multitask idea is plausible and the tables support it, but the FP extension and the loss specification need correction and public artifacts before the claims are fully credible.","tokens_in":8404,"tokens_out":7648,"duration_ms":79735,"concrete_test":"Run the proposed training loop with the unlabeled volumes removed from the forward pass (equivalently, compute the unlabeled branch under torch.no_grad() and then discard its outputs) while keeping the labeled batches fixed, and compare Dice/Jaccard on the LA test split after 6,000 iterations. If the training is truly forward-pass-only in a V-Net without normalization, the parameter updates are identical with and without unlabeled volumes, so the two runs must produce identical metrics; any difference would reveal an undisclosed mechanism, such as batch statistics or a hidden loss, and the paper's +FP explanation would fail.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's Section 2, 'Forward Pass Only for Semi-supervised Training (FP),' states that unlabeled volumes contribute to feature extraction but do not influence gradient backpropagation through the loss terms. For the V-Net used here, with no normalization layers, the optimizer update is θ ← θ − η Σ_{i∈L} ∇_θ ℓ_i(f(x_i), y_i) over labeled samples only; activations from a sample with zero attached loss do not appear in any gradient term. Thus the +FP rows in Table 1 (0.774→0.790 with 4 labeled volumes; 0.845→0.857 with 8 labeled volumes) cannot be produced by the described procedure. Either the implementation uses batch or instance normalization statistics, a pseudo-label or consistency term, or a different effective loss on unlabeled volumes—contradicting the 'exclusively during forward pass' description—or the reported differences are noise or implementation artifacts. The no-unlabeled BoundarySeg result (0.774 at 4 labels) is not directly refuted by this, but the FP claim is one of the paper's stated contributions, and its impossibility is a warning that the reported training procedure is not faithfully described.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BoundarySeg, a multi-task framework for medical image segmentation in low-data regimes. A V-Net backbone produces shared features that feed two single-convolutional-layer branches: one for full organ segmentation and one for organ boundary segmentation, with boundary labels generated by morphological erosion and XOR. The training loss combines Dice losses for the two tasks with a consistency term intended to align boundary and segmentation predictions, plus an optional extension in which unlabeled volumes are passed through the network during the forward pass only. Experiments on the left atrium MRI dataset with 4 and 8 labeled volumes report large improvements over the single-task lower bound (e.g., Dice 0.587 to 0.774 at 4 labels) and competitive or better performance than the semi-supervised baselines BCP, SASSnet, DTC, SS-Net, and UA-MT, as well as an ablation over morphological kernel size and boundary-loss weight.","tokens_in":8674,"tokens_out":4613,"duration_ms":50529,"significance":"If the central claims hold, the method is practically valuable: it is architecturally simple, model-agnostic, avoids unlabeled-data requirements, and adds little computational overhead, while the reported gains in a low-data regime are substantial. The paper also provides useful comparisons against several established semi-supervised baselines on a standard benchmark and includes hyperparameter ablations. However, as written, the central claim of consistency between the two task predictions is not faithfully represented by the stated loss equation, the forward-pass-only extension has no gradient-based mechanism under the described architecture, and the reported results lack variance information. These issues are load-bearing for the stated contributions, so the manuscript needs substantive revision before the empirical claims can be accepted.","major_comments":[{"comment":"The consistency loss does not implement the claimed consistency between the two task predictions. Equation (5) computes DiceLoss(f_theta2(X), L_Boundary * f_theta1(X)), where L_Boundary is the ground-truth boundary label, not a boundary extracted from the predicted full-organ segmentation. The text says the model aligns the predicted boundary with the boundary extracted from the full segmentation, but the equation instead multiplies the ground-truth boundary by the predicted segmentation. In addition, the Implementation Details section states that an MSE loss with weight 0.3 was used as the consistency loss in Eq. (6), which contradicts the Dice form in Eq. (5). Because the abstract and introduction emphasize consistency between task predictions as the source of additional supervision, the paper must state the actual loss used, define how the predicted boundary is extracted from the predicted segmentation, and rerun or relabel the reported experiments under that definition.","section":"Section 2, Eqs. (5)-(6) and Implementation Details"},{"comment":"The forward-pass-only mechanism is mechanistically unexplained for the architecture described. The paper states that unlabeled volumes contribute to feature extraction but do not influence gradient backpropagation through the loss terms, and that this allows the model to capture statistical properties of unlabeled volumes. In a V-Net without normalization layers, however, an unlabeled volume whose activations appear in no loss term has no path through which its forward pass can change the network weights: the optimizer update sums gradients only over the labeled samples that contribute to the loss. The reported +FP gains in Table 1 (0.774 to 0.790 at 4 labels; 0.845 to 0.857 at 8 labels) therefore cannot arise from the described procedure. The authors should either remove the FP claim, provide the actual mechanism (e.g., batch-normalization statistics, pseudo-labeling, or an additional consistency term on unlabeled data), or include an ablation such as random-label forward passes to demonstrate whether the gain is a real effect or noise.","section":"Section 2, 'Forward Pass Only for Semi-supervised Training (FP)'"},{"comment":"The central empirical claims are stated without any measure of variability. The paper reports that performance metrics are averaged over three independent runs with different random seeds, but no standard deviations, confidence intervals, or significance tests are given. Several decisive comparisons involve small gaps, such as the 0.016 Dice difference between BoundSeg and BoundSeg + Lcons at 4 labels, and the 0.012 Dice difference between BoundSeg and SS-Net with 4 labels and 4 unlabeled volumes, so it is impossible to determine whether these differences exceed run-to-run noise. The authors should report per-run results or error bars and, where appropriate, a paired significance test over the fixed test set.","section":"Table 1 and Section 3, Implementation Details"},{"comment":"The hyperparameters r and lambda are selected using ablation experiments on the LA dataset itself, and the manuscript does not state whether the ablation is performed on a training/validation split or on the same test set used in Table 1. If the selection is made on the test set, the reported results are optimistically biased. The paper should clarify the experimental protocol and, ideally, select hyperparameters on a validation split before evaluation on the held-out test set.","section":"Section 3, Ablation Studies and Table 2"}],"minor_comments":[{"comment":"The notation L_Boundary * f_theta1(X) is ambiguous: f_theta1(X) is not explicitly defined as a sigmoid/softmax probability map versus a hard prediction, and the elementwise product is not described in words.","section":"Section 2, Eq. (5)"},{"comment":"The caption refers to red and blue results, but the printed table does not show colors; the descriptions should identify rows by method name or another visible marker so the table is self-contained.","section":"Section 3, Table 1 caption"},{"comment":"Reference [1] is cited as the V-Net architecture, but the cited paper describes a road-extraction VNet, not the volumetric V-Net commonly used in medical image segmentation; the authors should cite the original 3D V-Net if that is the backbone actually used.","section":"References"},{"comment":"The outputs f_theta1(X) and f_theta2(X) are used inside DiceLoss without specifying whether the Dice loss is computed on raw logits, probabilities, or thresholded predictions; this should be stated for reproducibility.","section":"Section 2, Eqs. (2)-(4)"},{"comment":"The sentence 'it's outputs' should read 'its outputs,' and the paper would benefit from a consistent notation for the consistency-loss weight (lambda_cons is defined in Eq. (6) but not restated with its chosen value in the text).","section":"Section 3, Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"The core idea—using boundary prediction as an auxiliary segmentation task—is reasonable and likely useful, and the external baseline comparisons are a strength. My main concern beyond the technical discrepancies is that the paper's headline claims are stated more strongly than the evidence supports: the consistency loss equation does not match the text, the FP mechanism is not credible as described, and the small performance gaps are reported without variance. If the authors can correct the loss definition, provide a real mechanism or remove the FP claim, and add error bars, the paper could become a solid contribution. I do not see grounds for rejection, but the revision needs to be substantive rather than cosmetic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the core idea—add a boundary branch trained with Dice, using erosion/XOR labels—is a reasonable low-data trick, but as written two claims don't hold up: the forward-pass-only semi-supervised boost and the consistency loss as defined. The single-dataset result is suggestive but not yet reliable.\n\nWhat's good: it's a genuinely simple extension of the multi-task program already in Peri-Loss, DTC, and SASSnet. The boundary branch gives a real improvement over the single-task baseline (0.587→0.774 Dice at 4 labels), and they ran three seeds plus ablations for kernel size and lambda, which is more than many preprints do. The low-data comparison against semi-supervised baselines with matched labeled counts is useful.\n\nWhere it falls apart: First, the consistency loss. Eq. 5 says Dice(f_θ2(X), L_Boundary ⊙ f_θ1(X)); the implementation text says MSE loss was used. Those are different. More importantly, that product is not a consistency between two task predictions—it's a masked Dice on the boundary branch using ground-truth boundary and predicted segmentation. It doesn't do what the abstract claims. Second, the forward-pass-only (FP) results. For a V-Net without normalization layers, unlabeled volumes that contribute no loss term have no gradient path to update weights. The +FP rows (e.g., 0.774→0.790 at 4 labels) cannot be produced by the described mechanism. Either the implementation has some other loss or the gain is noise; three seeds without error bars can't distinguish. Third, the text says they outperform UA-MT, SASSNet, and DTC at 4 labels using 76 unlabeled volumes, but Table 1 shows UA-MT at 0.799 and SASSNet at 0.792 versus BoundarySeg's 0.774. That overstates the table.\n\nWho this is for: people working on low-data medical segmentation might find the idea cheap to try, but the paper needs a code release and corrected descriptions before its numbers can be trusted. It still deserves peer review because the underlying method is testable and the low-data regime matters. A competent referee could push for the fixes. I wouldn't cite it as evidence yet.","headline":"A plausible cheap boundary-loss trick for low-data segmentation, undercut by a misdescribed consistency term and an unexplained forward-pass-only gain.","tokens_in":9145,"tokens_out":3172,"would_cite":false,"duration_ms":31515,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"BoundarySeg claims that adding organ-boundary prediction as an auxiliary task with a consistency loss improves 3D medical image segmentation in low-data regimes, matching or beating semi-supervised methods without using unlabeled data.","keywords":["medical image segmentation","morphological operations","annotation-efficient training","few-shot segmentation","multi-task learning","boundary segmentation","semi-supervised learning","low-data regime"],"falsifier":"Re-train BoundarySeg on the same left-atrium split with the ground-truth boundary labels replaced by random binary masks that have the same foreground proportion; if Dice still rises from 0.587 to near 0.774 at four labeled volumes, the improvement does not come from boundary structure, and the central claim is refuted.","tokens_in":8222,"feed_emoji":"🫀","tokens_out":10001,"duration_ms":92422,"temperature":0.7,"pith_summary":"BoundarySeg is a training strategy for medical image segmentation that adds a second output branch, a boundary-prediction head, to an existing segmentation network and trains the two heads together with a consistency loss. The paper argues that this multi-task setup is enough to close most of the gap to semi-supervised methods that rely on large pools of unlabeled volumes, and that when unlabeled data are scarce or absent it outperforms them outright. The practical payoff would be a drop-in modification, one extra convolution layer plus two loss terms, that improves 3D segmentation without any additional annotation or data collection. On left-atrium MRI the method reports Dice of 0.774 versus 0.587 for the single-task baseline with four labeled volumes, and 0.845 versus 0.771 with eight labeled volumes.","feed_headline":"Boundary branch lifts low-data segmentation with no unlabeled data","feed_subtitle":"Adding a boundary Dice loss to a V-Net gains up to 0.19 Dice with four training labels, matching semi-supervised rivals.","key_machinery":"The carrying mechanism is a pair of single-convolution output heads on a shared V-Net encoder, trained with boundary labels produced by morphology: $L_{\\mathrm{Boundary}} = L \\oplus \\Gamma_r(L)$, where $\\Gamma_r$ is erosion with kernel radius $r$ and $\\oplus$ is pixel-wise XOR. The total loss is $L_{\\mathrm{BoundarySeg}} = L_{\\mathrm{Seg}} + \\lambda L_{\\mathrm{Boundary}} + \\lambda_{\\mathrm{Cons}} L_{\\mathrm{Cons}}$, with Dice terms for the organ and boundary heads and an MSE consistency term between the boundary head and the boundary extracted from the predicted organ mask. The boundary head is what carries the argument: it forces shared features to encode edge location, and the consistency term makes the two heads agree, adding supervision at no annotation cost.","core_discovery":"The central claim, stated on the paper's own terms, is that organ boundary prediction is sufficiently coupled to full-organ segmentation that a network can be made significantly more accurate in low-data regimes by predicting both at once and enforcing consistency between them. Boundary labels are obtained for free from existing masks by eroding the mask and XORing it with the original, so no extra annotation is required. With this construction the paper reports that BoundarySeg beats every semi-supervised baseline when all methods are limited to four unlabeled volumes, and that it still holds its own against several semi-supervised methods that consume all 76 or 72 unlabeled volumes. The reported gains are 0.774 Dice with four labeled volumes and 0.845 with eight, against 0.587 and 0.771 for the single-task baseline. The paper also proposes a forward-pass-only extension that feeds unlabeled volumes through the network without any loss term, claiming small additional gains.","pith_inferences":["Editorial inference: The forward-pass-only extension has no obvious mechanism in a V-Net without normalization layers, since unlabeled volumes contribute to no loss and therefore cannot change weights; the reported +FP gains likely come from an unintended path or from optimization noise, so they should not be taken as evidence for the stated semi-supervised mechanism.","Editorial inference: The boundary-consistency loss should transfer to other organs and modalities, since the paper's own ablation shows gains across all kernel sizes; a natural next test is kidney or pancreas segmentation with four and eight labeled volumes.","Editorial inference: The disagreement between the boundary head and the boundary of the predicted organ mask could be read as a per-voxel uncertainty map, giving clinicians a free quality-control signal that the paper does not explore.","Editorial inference: Because the consistency term compares two predictions derived from the same network, a trivial solution would be for the boundary head to copy the organ head's boundary; ablating the boundary Dice term while keeping the consistency term would separate the two contributions and is not reported."],"forward_implications":["With only four or eight labeled volumes, a lab can add this boundary head to its existing segmentation model and improve Dice by roughly 0.19 or 0.07 without collecting unlabeled scans.","Because boundary labels are derived from existing masks by erosion and XOR, the method adds no annotation burden and no post-processing.","In privacy-constrained settings where unannotated data cannot be shared, the method offers a fully supervised alternative that the paper reports as comparable to or better than semi-supervised methods.","The consistency loss can be layered onto any segmentation backbone, making the improvement a plug-and-play training modification rather than a new architecture.","If the reported comparisons hold, semi-supervised methods' dependence on large unlabeled pools is not a necessary cost: a purely supervised auxiliary task can reach similar accuracy when unlabeled data are scarce."],"supporting_citations":[{"why":"Provides the left-atrium MRI benchmark and fixed 80/20 split on which all reported comparisons are made.","marker":"[26]"},{"why":"Defines the V-Net architecture that serves as the segmentation backbone for BoundarySeg and all baselines.","marker":"[1]"},{"why":"Peri-Loss is the boundary-loss-only baseline that the multi-task framework is compared against and outperforms, isolating the value of the consistency mechanism.","marker":"[9]"},{"why":"BCP is the strongest semi-supervised baseline with full unlabeled data and the one showing the largest drop when unlabeled data is limited to four volumes.","marker":"[2]"},{"why":"SASSNet is a shape-aware semi-supervised baseline using signed distance field prediction; BoundarySeg is compared against it in both full and limited unlabeled settings.","marker":"[11]"},{"why":"DTC is a dual-task consistency baseline using level-set prediction, providing the closest architectural relative for the auxiliary-task idea.","marker":"[12]"},{"why":"SS-Net is the best semi-supervised baseline at four and eight labels in the limited-unlabeled setting, setting the bar BoundarySeg claims to beat.","marker":"[25]"},{"why":"UA-MT is a student-teacher pseudo-label semi-supervised baseline with the smallest drop under limited unlabeled data, used as another comparison point.","marker":"[29]"}],"fun_headline_variants":["BoundarySeg: boundary task boosts low-data segmentation","No unlabeled data needed: boundary aux task lifts segmentation","Simple boundary trick improves segmentation with few labels","Boundary prediction as auxiliary task beats semi-supervised in low data","Auxiliary boundary loss: free supervision for low-data segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that unlabeled volumes that are only passed through the network, with no loss and thus no gradient, can still improve the model; in the V-Net used here, which has no normalization layers, such volumes have no route to change the weights, so the forward-pass gains currently lack a mechanism.","fun_headline_variants_meta":{"raw":{"variants":["BoundarySeg: boundary task boosts low-data segmentation","No unlabeled data needed: boundary aux task lifts segmentation","Simple boundary trick improves segmentation with few labels","Boundary prediction as auxiliary task beats semi-supervised in low data","Auxiliary boundary loss: free supervision for low-data segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000595,"raw_usage":{"total_tokens":2778,"prompt_tokens":928,"completion_tokens":1850,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":1770}},"tokens_in":544,"tokens_out":1850,"duration_ms":11444,"temperature":1.0,"reasoning_tokens":1770,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:23:23.222036+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-train BoundarySeg on the same left-atrium split with the ground-truth boundary labels replaced by random binary masks that have the same foreground proportion; if Dice still rises from 0.587 to near 0.774 at four labeled volumes, the improvement does not come from boundary structure, and the central claim is refuted.","supporting_citations":[{"cited_title":"Medical image analysis67, 101832 (2021)","cited_arxiv_id":null,"evidence_quote":"Provides the left-atrium MRI benchmark and fixed 80/20 split on which all reported comparisons are made."},{"cited_title":"Ieee Access8, 179424–179436 (2020)","cited_arxiv_id":null,"evidence_quote":"Defines the V-Net architecture that serves as the segmentation backbone for BoundarySeg and all baselines."},{"cited_title":"In: Medical Imaging with Deep Learning","cited_arxiv_id":null,"evidence_quote":"Peri-Loss is the boundary-loss-only baseline that the multi-task framework is compared against and outperforms, isolating the value of the consistency mechanism."},{"cited_title":"In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"BCP is the strongest semi-supervised baseline with full unlabeled data and the one showing the largest drop when unlabeled data is limited to four volumes."},{"cited_title":"In: Medical Image Comput- ingandComputerAssistedIntervention–MICCAI2019:22ndInternationalConfer- BoundarySeg11 ence, Shenzhen, China, October 13–17, 2019, Proceedings, Part II 22","cited_arxiv_id":null,"evidence_quote":"UA-MT is a student-teacher pseudo-label semi-supervised baseline with the smallest drop under limited unlabeled data, used as another comparison point."}],"review_version":1}