{"id":"706640e9-33da-40f9-8152-2ccdeb88da22","arxiv_id":"1908.04250","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"An ensemble of three 2D U-Nets trained on axial, sagittal, and coronal MRI slices achieves accuracy close to 3D networks on the BraTS 2018 brain tumour segmentation benchmark.","lead":"The paper describes a brain tumour segmentation method that trains three 2D neural networks on different MRI view planes and averages their predictions, reporting accuracy close to heavier 3D networks on the BraTS 2018 benchmark. It reports Dice scores of 0.79 for enhancing tumour, 0.90 for whole tumour, and 0.82 for tumour core with lower memory and time costs.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The gain of the multi-view ensemble over a single axial model may be an ensembling artifact; without an axial-only ensemble control, the claim that orthogonal views provide 3D context is unsupported.","rationale":"The reader's verdict is CONDITIONAL, and this stress-test identifies an additional control that the revision should include, but does not change the overall disposition. The reported validation scores come from the official BraTS 2018 evaluation tool, which is credible independent evidence, and the method is described in enough detail to reimplement. The load-bearing weakness is internal to the paper's ablation logic: the comparison of one axial model to an ensemble of three multi-view models cannot disentangle the benefit of ensembling from the benefit of orthogonal views. The reader's weakest assumption focused on the cross-method leaderboard comparison to 3D networks; that concern is valid, but the ensembling confound is more directly tied to the paper's stated mechanism and is testable with a same-protocol control experiment. A conditional acceptance with a request for this control experiment, uncertainty quantification, and code release remains appropriate; no change to the conditional verdict is needed.","tokens_in":11274,"tokens_out":4541,"duration_ms":47633,"concrete_test":"Train three identical axial-view baseline networks using the same architecture, weighted Dice loss, 300 epochs, Adam with learning rate 1e-4, L2 regularization 1e-5, and batch size 8, but with different random seeds. Average their softmax outputs at inference and evaluate on the BraTS 2018 validation set with the official online tool. Compare Dice and 95% Hausdorff distance against the multi-view ensemble in Table 2. If the axial-only ensemble matches the multi-view ensemble (e.g., Dice within 0.01 and Hausdorff distance within about 1 mm), then the reported gains are attributable to ensembling rather than to 3D contextual information from orthogonal views; if it is substantially worse, the 3D-context interpretation is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central causal claim, stated in Sections 2.6, 3.2, and 4, is that the ensemble of three 2D networks trained on axial, sagittal, and coronal views improves performance because it captures 3D contextual information. The supporting evidence in Tables 1 and 2 compares a single axial-view model to an ensemble of three models trained on three different views. This comparison conflates two variables: the number of models averaged and the availability of orthogonal views. Ensembling three independent networks is known to reduce variance and outlier errors, and the large improvement in 95% Hausdorff distance (validation: ET 7.29->2.99, WT 10.39->6.28, TC 11.60->5.90) could plausibly come from averaging three axial models alone. No control ensemble of three axial-only networks is reported. Additionally, the text's claim that a single multi-view network cannot capture 3D context is weakened by the tables: in Table 1 the single multi-view model improves all three Hausdorff distances, and in Table 2 it improves TC Hausdorff distance, contradicting the statement that Hausdorff distance increased for all three sub-regions. Thus the least secure part of the argument is not only the comparison against 3D networks, but the attribution of the ensemble gain to orthogonal views rather than to ensembling.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 2D fully residual U-Net with a weighted Dice loss for brain tumour segmentation on the BraTS 2018 dataset. The authors report two methods for incorporating 3D context: a single network trained on patches from axial, sagittal, and coronal views, and an ensemble of three networks each trained on one orthogonal view with averaged softmax outputs. They report validation Dice scores of 0.79 (ET), 0.90 (WT), and 0.82 (TC) and Hausdorff distances of 2.99, 6.28, and 5.90, and argue that the ensemble of three 2D networks can capture 3D context and deliver performance comparable to 3D CNN methods. They also report that increasing U-Net depth from 3 to 4 decreases performance and that a single multi-view 2D network cannot capture 3D context.","tokens_in":11463,"tokens_out":5849,"duration_ms":56444,"significance":"If the central claims were fully supported, this would be a practically useful result: a simple 2D ensembling strategy that matches 3D CNNs on BraTS with lower memory and time costs would be of interest to the medical imaging community. The paper's strengths include the use of the official BraTS 2018 validation tool for measuring Dice and Hausdorff distances, a clear description of the architecture and training procedure, and a comparison against several published 3D methods. The weighted Dice loss is standard but appropriately motivated by the extreme class imbalance. However, the experimental design does not currently isolate the causal factors behind the ensemble's improvement, and the comparison to 3D methods is based on uncontrolled leaderboard scores. The central claims are therefore plausible but not established as stated.","major_comments":[{"comment":"The comparison between \"Baseline\" (a single axial-view model) and \"Ensemble of three models\" changes two variables simultaneously: the number of networks averaged and the diversity of views. The large improvements in 95% Hausdorff distance on the validation set (ET 7.29 to 2.99, WT 10.39 to 6.28, TC 11.60 to 5.90) could plausibly be an ensembling effect alone, because averaging independent models is known to reduce variance and outlier errors. To support the claim that orthogonal views provide 3D contextual information beyond simple ensembling, the authors should add a control ensemble of three axial-only networks. Without this control, the central attribution of the ensemble gain to multi-view context is not established.","section":"Sections 2.6 and 3.2, Tables 1 and 2"},{"comment":"The statement that training a single model on three orthogonal views \"resulted in little improvement in Dice score for the enhancing tumour but the Hausdorff distance increased for all the three tumour sub-regions\" is contradicted by the paper's own tables. In Table 1, all three Hausdorff distances decrease compared with the baseline (5.71 to 4.96, 11.94 to 8.31, 9.71 to 7.29), and in Table 2 the TC Hausdorff distance decreases from 11.60 to 7.22. The abstract and Section 4 repeat the stronger claim that a single 2D network \"cannot capture 3D contextual information,\" but the reported results are mixed. The text should be corrected and the conclusion re-framed to match the actual data.","section":"Section 3.2, text versus Tables 1 and 2"},{"comment":"The claim that a 2D network \"can deliver comparable performance to a 3D network\" rests on a comparison with leaderboard scores from other teams' 3D architectures. These methods differ in architecture, training data usage, ensembling, post-processing, and compute budget, and the paper does not train a 3D version of its own network. The score gaps in Table 4 (e.g., ET Dice 0.79 versus 0.80-0.82, TC Dice 0.82 versus 0.84-0.86) are small, but without controlling for the confounds, the results do not demonstrate that the 2D-versus-3D design choice, rather than other factors, explains the comparable performance. Please either add an internal 3D baseline or explicitly present Table 4 as an uncontrolled benchmark reference, and soften the causal wording.","section":"Section 3.2, Table 4"},{"comment":"Every configuration appears to be evaluated from a single training run, with no variance estimates or statistical tests. This is particularly problematic for the paper's secondary findings: the depth conclusion is based on a single increment (depth 3 to depth 4), and the differences between the single multi-view model and the baseline are mixed (some metrics improve, some worsen). The authors should report multiple-run statistics or clearly acknowledge that the observed differences may not be significant, especially when making claims such as \"we must always use an optimum depth.\"","section":"Section 3.2, Tables 1 and 2"}],"minor_comments":[{"comment":"The 80/20 split of extracted patches appears to be performed at the slice level rather than the patient level, so patches from the same patient can appear in both the training and internal validation sets. Please state whether the split was patient-independent or discuss the potential impact on parameter tuning.","section":"Section 2.1"},{"comment":"The paper claims that the 2D method \"consumes very less time and memory\" compared with 3D methods, but no quantitative timing or memory measurements are reported. Please provide actual inference time and GPU memory usage if this advantage is to be considered a contribution.","section":"Section 3.2"},{"comment":"The caption reads \"On the top of each box the number channels are denoted\"; this should be \"the number of channels are denoted\" or similar.","section":"Figure 1 caption"},{"comment":"The text describing Figure 4 says \"green colour represents the ET,\" but the caption correctly says green represents ED. This appears to be a typo and should be corrected.","section":"Section 3.2, final paragraph"},{"comment":"References [13] and [14] contain malformed parentheses: \"ISLES-MICCAI. (, 2015\" and \"(IEEE, 2015\" respectively. Please fix these formatting errors.","section":"References"},{"comment":"It is unclear whether the single multi-view model is trained by concatenating all patches from all views into one training file or by alternating views per epoch. Please clarify the exact training protocol.","section":"Section 2.6"}],"recommendation":"major_revision","confidential_remarks":"The paper is a straightforward BraTS method description with a modest novelty claim centered on the multi-view ensemble. The main experimental deficiency is the absence of an axial-only ensemble control, which is essential to distinguish the effect of ensembling from the effect of orthogonal views. The comparison to 3D methods is uncontrolled and should be framed more cautiously. The reported scores are from the official validation tool, which is a strength, but the single-run nature of all experiments limits the strength of the conclusions. I would recommend major revision rather than rejection, as the issues are addressable with additional experiments or substantial rewording of the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a straightforward BraTS 2018 write-up with credible official-tool scores, but the headline interpretation—that orthogonal views provide 3D context beyond simple ensembling—is not supported by the experiments as reported, and one of the paper's own tables contradicts the text on the single-multiview baseline.\n\nWhat's genuinely useful: the paper records a clean empirical comparison between a single 2D U-Net trained on all three views and an ensemble of view-specific 2D U-Nets, on an external held-out validation set. The ensemble numbers (Dice 0.79/0.90/0.82, HD 2.99/6.28/5.90) are plausibly competitive, and the practical motivation—2D nets are cheap and simple—is legitimate. The depth experiment, though only one increment, honestly reports a negative result that other practitioners may find informative.\n\nThe soft spots are real. The central causal claim conflates number of models with view diversity: there is no axial-only ensemble control. Averaging three models alone could easily explain the large Hausdorff improvements. Second, Section 3.2 states that the single multi-view model increased Hausdorff distance for all three sub-regions, but Table 1 (training) shows HD decreasing for all three, and Table 2 (validation) shows TC improving from 11.60 to 7.22. That's a direct internal contradiction, and it undermines the conclusion that a single 2D network cannot capture 3D context. Third, the comparison to 3D methods is against other teams' leaderboard entries, not a same-architecture 3D baseline, so the 'comparable to 3D' claim is not isolated from training budget, ensembling, or tuning. There are no error bars anywhere.\n\nThis is a modest, honest paper with addressable flaws. A serious editor should send it to review, but the reviewers should ask for the ensemble control, a corrected reading of the tables, and ideally uncertainty quantification or code. It's not a wasted read for anyone working on efficient segmentation or challenge evaluation design.","headline":"A credible, useful BraTS 2018 write-up whose central interpretive claim—multi-view ensembling captures 3D context—is confounded with ensembling itself and contradicted by one of its own tables.","tokens_in":12070,"tokens_out":3881,"would_cite":false,"duration_ms":37331,"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":"A 2D-only ensemble of three view-specific U-Nets, averaged at inference, matches 3D networks on brain tumour segmentation while using far less memory and time.","keywords":["brain tumour segmentation","glioma","2D U-Net","fully residual convolutional network","orthogonal views ensemble","weighted Dice loss","BraTS 2018","Hausdorff distance"],"falsifier":"Train a 3D version of this exact residual U-Net on the same BraTS 2018 training and validation splits with the same loss, augmentation, and training schedule, and compare it with the three-view 2D ensemble; if the 3D model clearly beats the ensemble on Dice or Hausdorff distance, the central claim would be undermined. A smaller test is to train a single 2D network on mixed-view patches with more epochs and see whether it can match the ensemble, which would challenge the claim that view-specific ensembles are necessary.","tokens_in":10977,"feed_emoji":"🧠","tokens_out":14308,"duration_ms":132419,"temperature":0.7,"pith_summary":"The paper's thesis is that 3D context, normally the reason to use 3D convolutional networks, can be supplied to a much cheaper 2D pipeline by ensembling three view-specific networks. The author trains the same fully residual U-Net on axial, sagittal, and coronal slices and averages their softmax outputs at inference, reaching Dice scores of 0.79, 0.90, and 0.82 for enhancing tumour, whole tumour, and tumour core on the BraTS 2018 validation set. These numbers sit beside recent 3D results in the paper's comparison table, which supports the claim that a 2D network with multi-view context can be comparable to a 3D network. The practical payoff is that automated tumour segmentation could run with much lower memory and time costs. The paper also reports that a single 2D network trained on mixed-view patches fails to capture the same context, and that adding a fourth level to the U-Net hurts performance.","feed_headline":"Three 2D U-Nets match 3D networks on brain tumour segmentation","feed_subtitle":"Averaging axial, sagittal, and coronal views yields Dice 0.90 whole tumour on BraTS 2018 validation.","key_machinery":"The central object is a 2D fully residual U-Net. Its encoder and decoder are built from residual blocks, each containing two 3×3 convolutions with batch normalisation and ReLU; downsampling uses stride-2 convolutions, upsampling uses bilinear interpolation, and skip connections concatenate encoder activations to the corresponding decoder level. Around this backbone sits the decisive mechanism: an ensemble of three such networks trained separately on axial, sagittal, and coronal slices, with predictions obtained by averaging the three softmax outputs. The loss is a weighted Dice loss, $w_c = 1/\\sum_i g_{ci}$, which gives more weight to small tumour classes so that a single network can handle the severe class imbalance. The ensemble is what lets a 2D model see volume context without 3D convolutions.","core_discovery":"On the paper's own account, the discovery is that 3D contextual information cannot be captured by one 2D network trained on patches from multiple views, but can be captured by an ensemble of three 2D networks each trained on a single orthogonal view. The ensemble improves both Dice scores and the 95% Hausdorff distances, which drop from 7.29, 10.39, and 11.60 mm in the axial baseline to 2.99, 6.28, and 5.90 mm for enhancing tumour, whole tumour, and tumour core respectively. The author reads this as evidence that a 2D network combining axial, sagittal, and coronal context can deliver improvements comparable to 3D networks while remaining simpler and less memory-intensive.","pith_inferences":["If the benefit comes from averaging view-consistent predictions, the same ensemble design should transfer to other 2D segmentation architectures and to other imaging settings where slice spacing is anisotropic.","Because the ensemble mainly improves Hausdorff distance rather than Dice, the mechanism probably cleans boundary errors; an ablation that measures border error separately could confirm this reading.","A direct runtime and peak-memory comparison of the 2D ensemble against a 3D U-Net on the same GPU would turn the paper's cost advantage into a quantitative claim, since the paper reports accuracy but leaves runtime unmeasured.","The optimum-depth finding suggests that depth should be tuned per view rather than once globally, a variation the paper leaves untested."],"forward_implications":["A 2D-only pipeline can reach Dice scores comparable to published 3D networks on the BraTS 2018 validation set, so 3D convolutions are not strictly necessary for competitive glioma segmentation.","Mixing axial, sagittal, and coronal slices into one network's training set is not enough to capture 3D context; the improvement requires separate view-specific models whose outputs are averaged.","The weighted Dice loss with inverse-frequency weights lets a single model handle the extreme class imbalance, where roughly 98.88% of all voxels are background.","A four-level U-Net underperforms a three-level version, so an optimal U-depth exists and must be chosen rather than maximised.","The three-view ensemble gives 95% Hausdorff distances of 2.99, 6.28, and 5.90 mm for enhancing tumour, whole tumour, and tumour core, showing that the gain is not only overlap but boundary shape."],"supporting_citations":[{"why":"This reference defines the BraTS benchmark, the three tumour sub-regions, and the evaluation metrics used throughout the paper.","marker":"[6]"},{"why":"This reference provides the U-Net encoder-decoder with skip connections that the proposed fully residual 2D network adapts.","marker":"[15]"},{"why":"This reference introduces the idea of training on three orthogonal views and averaging their outputs, which the paper tests with 2D networks.","marker":"[16]"},{"why":"This reference supplies the residual block structure used in every encoder and decoder block of the proposed network.","marker":"[18]"},{"why":"This reference reports the strongest 3D-network validation results that anchor the comparison in Table 4.","marker":"[19]"},{"why":"This reference reports a 3D U-Net variant's validation scores used as another comparison row in Table 4.","marker":"[20]"},{"why":"This reference reports the closest whole-tumour Dice comparison among the table's 3D methods.","marker":"[22]"},{"why":"This reference reports a cascade 3D method's validation scores in the comparison table.","marker":"[24]"},{"why":"This reference is the official BraTS 2018 validation leaderboard that supplies the 3D scores compared against the proposed ensemble.","marker":"[37]"}],"fun_headline_variants":["Ensemble of 3 2D U-Nets matches 3D brain tumour segmentation","Three 2D U-Nets capture 3D context for tumour segmentation","Multi-view 2D U-Net ensemble mimics 3D performance on BraTS","Three 2D U-Nets outperform single-view on brain tumour segmentation","Averaging three 2D views gives 3D-like tumour segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that differences between the proposed method and the published 3D methods on the BraTS 2018 validation leaderboard can be attributed to the 2D-versus-3D design, even though those 3D scores were produced by other teams with different architectures, ensembling strategies, training budgets, and tuning, and the paper does not train a 3D version of its own network.","fun_headline_variants_meta":{"raw":{"variants":["Ensemble of 3 2D U-Nets matches 3D brain tumour segmentation","Three 2D U-Nets capture 3D context for tumour segmentation","Multi-view 2D U-Net ensemble mimics 3D performance on BraTS","Three 2D U-Nets outperform single-view on brain tumour segmentation","Averaging three 2D views gives 3D-like tumour segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001226,"raw_usage":{"total_tokens":5029,"prompt_tokens":925,"completion_tokens":4104,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":3999}},"tokens_in":541,"tokens_out":4104,"duration_ms":28017,"temperature":1.0,"reasoning_tokens":3999,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:46:12.670305+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a 3D version of this exact residual U-Net on the same BraTS 2018 training and validation splits with the same loss, augmentation, and training schedule, and compare it with the three-view 2D ensemble; if the 3D model clearly beats the ensemble on Dice or Hausdorff distance, the central claim would be undermined. A smaller test is to train a single 2D network on mixed-view patches with more epochs and see whether it can match the ensemble, which would challenge the claim that view-specific ensembles are necessary.","supporting_citations":[{"cited_title":"1993--2024","cited_arxiv_id":null,"evidence_quote":"This reference defines the BraTS benchmark, the three tumour sub-regions, and the evaluation metrics used throughout the paper."},{"cited_title":"` Automatic Brain Tumor Segmentation Using Cascaded Anisotropic Convolutional Neural Networks '","cited_arxiv_id":null,"evidence_quote":"This reference introduces the idea of training on three orthogonal views and averaging their outputs, which the paper tests with 2D networks."},{"cited_title":"` Ensembles of Densely-Connected CNNs with Label-Uncertainty for Brain Tumor Segmentation '","cited_arxiv_id":null,"evidence_quote":"This reference reports the closest whole-tumour Dice comparison among the table's 3D methods."},{"cited_title":"` Learning Contextual and Attentive Information for Brain Tumor Segmentation '","cited_arxiv_id":null,"evidence_quote":"This reference reports a cascade 3D method's validation scores in the comparison table."},{"cited_title":"Available from: https://www.cbica.upenn.edu/BraTS18/lboardValidation.html, accessed July 2019","cited_arxiv_id":null,"evidence_quote":"This reference is the official BraTS 2018 validation leaderboard that supplies the 3D scores compared against the proposed ensemble."}],"review_version":1}