{"id":"c5c5e2a7-a213-401b-8f57-bb25882a0f92","arxiv_id":"1908.05104","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A hybrid 2D/3D UNet with a dimension-fusion block and a focal-Dice loss achieves a Dice score of 0.5349 on ATLAS chronic stroke lesion segmentation, modestly outperforming 2D and 3D baselines.","lead":"The paper presents D-UNet, a stroke lesion segmentation network that mixes 2D and 3D convolutions in its encoder, and an Enhanced Mixing Loss to speed up training. On the public ATLAS stroke MRI dataset, it reports a Dice score of 0.5349, slightly above a standard 2D UNet, with little added computation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Table 2 comparison confounds architecture and loss; the paper never states which loss the UNet baselines used, and its own Table 4 shows EML is not consistently better than Dice, so the +0.038 DSC cannot be attributed to the dimension-fusion block.","rationale":"The reader's conditional verdict is reasonable, but the sharpest issue is attribution. The paper's title and central claim are about a dimension-fusion architecture, yet the method bundles a new loss. The tables do not separate the two. The internal evidence in Table 4 cuts against the loss being the source of improvement in DSC, which makes the omitted baseline-loss specification critical: if baselines used Dice loss, then D-UNet with Dice loss should be compared against 2D UNet with Dice loss. The small validation set and tuned hyperparameters compound the problem. I do not see fraud or intentional misreporting; the issue is experimental design and reporting. The proposed check would settle whether the architecture contributes independently. Until that is done, the central claim should not be accepted as established, which matches the reader's CONDITIONAL verdict.","tokens_in":14451,"tokens_out":8677,"duration_ms":80079,"concrete_test":"Run a 2x2 factorial comparison on the same 183/46 split: (a) D-UNet with EML, (b) D-UNet with Dice loss, (c) 2D UNet(transform) with EML, (d) 2D UNet(transform) with Dice loss, using identical α/γ/δ, optimizer, epochs, and augmentation. Repeat over at least 5 random splits. If (d) matches (a) within paired bootstrap confidence intervals, or if (b) does not beat (c), then the headline DSC gain is attributable to loss or split, not the D-UNet architecture.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is that Table 2's comparison does not control for the loss function. Sections 3.3 and 4.1 never state which loss was used to train the 2D UNet(original), 2D UNet(transform), or 3D UNet(transform) baselines; only the proposed model is described as using EML, with α=1.1, γ=0.48, δ=1 set 'to fit our randomly selected dataset.' If the baselines used a standard BCE or Dice loss, the reported gains (DSC 0.5349 vs 0.4966, precision 0.6331 vs 0.5511) could stem entirely from the loss rather than the dimension-fusion block. This is not a hypothetical objection: Table 4 shows EML yields lower global DSC than plain DL on D-UNet itself (0.7231 vs 0.7346), so EML is not uniformly better. The architecture's contribution is therefore unmeasured. The single 183/46 split with per-case SD ≈0.28 makes the 0.038 difference additionally vulnerable to chance, and the stated '2%' parameter increase is arithmetically inconsistent (8,640,163 vs 7,771,297 is about +11%).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes D-UNet, a hybrid 2D/3D U-shaped network for chronic stroke lesion segmentation from T1-weighted MRI. The encoder runs parallel 2D and 3D convolutional streams in the early downsampling stages and fuses them through a dimension-transform block that includes squeeze-and-excitation weighting. The paper also proposes an Enhanced Mixing Loss (EML) that combines focal loss and a log Dice loss. The method is evaluated on the public ATLAS dataset with a single random 183/46 train/validation split, reporting DSC = 0.5349±0.2763 and precision = 0.6331±0.2958, which the authors claim is better than 2D and 3D UNet baselines and several other segmentation frameworks while adding few parameters.","tokens_in":14734,"tokens_out":4756,"duration_ms":45269,"significance":"If the central claim were fully established, the contribution would be practically useful: the architecture is comparatively simple, the parameter overhead over a 2D UNet is modest, and the evaluation on a public dataset (ATLAS) supports reproducibility. The paper also provides useful ablations in Table 5 on the placement of the fusion block and the value of SE weighting, and a loss comparison in Table 4/Fig. 5. However, the main quantitative claim is currently not supported by the experiments as reported: the comparison in Table 2 does not control for the loss function, the reported gains are within the per-case standard deviation, the hyperparameters were tuned to the validation split, and the parameter-efficiency claim is arithmetically inconsistent. These issues are fixable with additional controlled experiments and statistical reporting, so the work is a reasonable candidate for major revision rather than rejection.","major_comments":[{"comment":"The comparison confounds architecture and loss. The manuscript never states which loss function was used to train 2D UNet(original), 2D UNet(transform), or 3D UNet(transform), whereas the proposed method is explicitly trained with EML. This matters because Table 4 shows EML is not uniformly better than plain Dice loss on D-UNet itself (global DSC 0.7231 vs. 0.7346). The reported +0.038 DSC gain over the 2D UNet(transform) cannot therefore be attributed to the dimension-fusion block. Please re-run all baselines with the same loss function (e.g., EML or Dice loss) and report those results.","section":"Section 4.1, Table 2"},{"comment":"The statistical basis for the main claim is weak. The evaluation uses a single random split of 183/46 cases with no cross-validation or repeated runs, and no significance tests are reported. The per-case DSC standard deviation is about 0.28, which is much larger than the reported improvement of 0.038. Additionally, the loss parameters α=1.1, γ=0.48, δ=1 are stated in Section 3.3 to be set \"to fit our randomly selected dataset,\" which risks overfitting the validation split. Please report results over multiple splits or seeds, with paired significance tests and confidence intervals on per-case DSC.","section":"Section 3.3, Section 4.1, Table 2"},{"comment":"The claim that the proposed method increases parameters by only \"2%\" is arithmetically incorrect. Comparing 8,640,163 parameters with the 7,771,297 parameters of 2D UNet(transform) gives an increase of roughly 11.2%, not 2%. Please correct this number or adjust the architecture description so the efficiency claim is accurate.","section":"Section 4.1, Table 2"},{"comment":"The loss-function validity experiment does not support the role of EML in the main result. On the proposed architecture, EML achieves lower global DSC than Dice loss (0.7231 vs. 0.7346) and lower recall, with only precision improved. The paper characterizes this as acceptable because the goal is faster convergence, but the convergence claim is supported only by the training-curve figure, not by quantitative convergence-time measurements. Please provide a criterion for the claimed advantage of EML and, if EML is to be credited for the Table 2 gains, demonstrate this in the controlled comparison requested above.","section":"Section 4.3, Table 4"},{"comment":"The internal ablation is informative but still statistically fragile. The best configuration, SEAdd-23, differs from Add-23 by only 0.0101 DSC (0.5349 vs. 0.5248), well within the reported per-case standard deviations, and all results come from the same single split. Please report repeated-run variability or a statistical test to support the conclusion that SE weighting and fusion depth reliably improve performance.","section":"Section 4.4, Table 5"}],"minor_comments":[{"comment":"There is a typographical error in the results: \"DSC = 0.5349+0.2763\" should be \"0.5349±0.2763\", and the final metric \"precision = 0.6331±0.295\" is missing the closing parenthesis or trailing digits.","section":"Abstract"},{"comment":"The word \"infraction\" in the first paragraph should be \"infarction\".","section":"Introduction"},{"comment":"The initialization method \"Hes method\" should be \"He method\" (after He et al. [35]).","section":"Section 3.3"},{"comment":"The text states that the proposed method has the highest precision score of \"0.6631\", but Table 3 reports 0.6331. Please correct the inconsistency.","section":"Section 4.2, text after Table 3"},{"comment":"The sentence \"This is 0.30 higher than the Clusterize method in DSC\" is numerically inconsistent with Table 3: the difference between 0.7231 and 0.23 is 0.4931, not 0.30.","section":"Section 4.2"},{"comment":"The caption lists \"Baseline, DenseUnet, DeepLabv3+, PSPNet, and FCN-8s\" while the text and Table 3 mention SegNet, PSP, and DeepLab v3 plus; the naming should be unified.","section":"Fig. 3 caption vs. Section 4.2"},{"comment":"The sentence \"The results we reported is also consistent with that of [29]\" contains a subject-verb agreement error; please rephrase.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript does not exhibit circularity: the architecture is evaluated on an external public dataset and the authors do not rely on their own prior results for the central claim. The main concerns are experimental rigor and the architecture/loss confound, both of which can be addressed within the scope of a revision. I would not reject on the basis of novelty, but the quantitative claims need to be substantially re-supported before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"If you only read one table, read Table 4: EML is not better than plain Dice loss on global DSC (0.7231 vs 0.7346). That undercuts the loss claim, and it makes the +0.038 DSC gain over 2D UNet in Table 2 uninterpretable because the paper never states which loss the UNet baselines used. If the baselines ran with BCE or Dice while D-UNet used EML, the gain could come entirely from the loss. This is the paper's central weakness.\n\nWhat is actually new: a dimension-fusion block that combines 2D and 3D encoder features with SE weighting, tested on ATLAS. The ablation in Table 5 is genuinely useful—it shows SE helps, and fusing at layers 2–3 beats deeper or shallower fusion. That is real evidence the architecture does something beyond the loss. The paper also presents an honest box plot and includes the loss comparison even though it weakens their story. No circularity; the evaluation is on an external public dataset.\n\nSoft spots, in proportion: first, the confound above is load-bearing. Second, the single random 183/46 split with per-case SD around 0.28 makes the 0.038 difference noise-level; no significance testing, no multi-split validation. Third, the loss hyperparameters were tuned on the same split, explicitly 'to fit our randomly selected dataset'—that invites overfitting. Fourth, the parameter increase is stated as 2% but 8,640,163 vs 7,771,297 is about 11%, a concrete arithmetic error. Fifth, the code link is mentioned but not verified, so reproducibility is uncertain.\n\nThis is not a fatal flaw—the dimension-fusion idea is plausible and the ablation supports it. But the current evidence does not establish the claimed advantage, and the overclaim in the abstract ('best quality performance') is contradicted by their own Table 4.\n\nWho this is for: someone working on chronic stroke lesion segmentation or on lightweight hybrid 2D/3D architectures. It will not change how we think about segmentation, but the fusion block is worth knowing about. The paper deserves a serious referee, not a desk reject. I would send it to review and ask for a controlled loss comparison, multiple splits, and corrected numbers.","headline":"Incremental 2D/3D fusion UNet with a useful ablation, but the headline gain over UNet is confounded by an uncontrolled loss function and a single noisy split.","tokens_in":15303,"tokens_out":2780,"would_cite":false,"duration_ms":28001,"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 2D-3D fusion encoder lifts stroke lesion DSC to 0.535.","keywords":["D-UNet","dimension fusion","2D-3D hybrid CNN","Enhanced Mixing Loss","stroke lesion segmentation","ATLAS dataset","MRI"],"falsifier":"Rerun D-UNet and the tuned 2D UNet on ATLAS across several random train/validation splits with the loss hyperparameters held at their default values; if the mean DSC gap (reported as 0.038) falls within the per-case standard deviation of about 0.28, the architecture's benefit over the 2D baseline would not be established.","tokens_in":14196,"feed_emoji":"🧠","tokens_out":9445,"duration_ms":75788,"temperature":0.7,"pith_summary":"This paper proposes a U-shaped segmentation network, D-UNet, that mixes 2D and 3D convolutions in the encoder so that a mostly 2D model can still exploit the volumetric context of MRI scans. On the ATLAS chronic stroke dataset, the authors report a Dice similarity coefficient of 0.5349±0.2763 and a precision of 0.6331±0.2958, outperforming a tuned 2D UNet (DSC 0.4966) and a 3D UNet (DSC 0.4710) while adding only about 2% more parameters than the 2D network. The paper also introduces an Enhanced Mixing Loss that combines focal loss with a log-transformed Dice loss, which they report converges faster than either loss alone. If the results hold, the practical message is that volumetric context does not require a full 3D network; a light 2D-plus-early-3D fusion can capture enough 3D information for small, irregular lesions at low computational cost.","feed_headline":"Fusing 2D and 3D convolutions pushes stroke lesion DSC to 0.535","feed_subtitle":"It adds only 2% more parameters than a 2D U-Net while beating it and a pure 3D U-Net.","key_machinery":"The dimension-fusion block is the load-bearing component: a shallow 3D branch runs 3D convolutions over a stack of four consecutive slices in the early encoder, and its output is channel-compressed with a $1\\times1\\times1$ convolution, spatially squeezed, and projected with a 2D $3\\times3$ convolution to match the 2D branch; both branches are then reweighted by squeeze-and-excitation blocks and added. Fusion blocks are placed after the second and third downsampling stages, and the paper finds that adding a third fusion stage degrades performance. The second component, Enhanced Mixing Loss, combines a voxel-averaged focal loss with the negative logarithm of the Dice loss, $EML = \\frac{1}{N}FL - \\log(DL)$, which the paper reports accelerates and smooths convergence.","core_discovery":"The central claim is that fusing 3D feature maps into the early downsampling stages of a 2D U-Net, through a dimension-fusion block, improves chronic stroke lesion segmentation relative to both pure 2D and pure 3D U-Nets on the ATLAS dataset. The D-UNet reaches a per-case DSC of 0.5349±0.2763 and precision of 0.6331±0.2958, beating the tuned 2D UNet (DSC 0.4966) and the 3D UNet (DSC 0.4710) with only about 2% more parameters than the 2D baseline. The authors further claim that the Enhanced Mixing Loss, $EML(p,g)=\\frac{1}{N}FL(p,g)-\\log(DL(p,g))$, produces faster and smoother training convergence than focal loss or Dice loss alone, while slightly improving precision at a small cost in DSC.","pith_inferences":["The design suggests a general recipe for other volumetric segmentation tasks: capture local 3D context in a shallow, cheap branch, then let a 2D network refine boundaries; this is a testable extension to datasets beyond ATLAS.","The drop in DSC when fusion is added at every downsampling stage hints that deep 2D/3D fusion may create gradient conflicts; ablating fusion depth on other datasets could determine whether the optimum is dataset-specific or general.","Because the loss hyperparameters ($\\alpha=1.1$, $\\gamma=0.48$, $\\delta=1$) were chosen to fit the single random split, the reported convergence advantage of EML over Dice loss may depend on that tuning; fixing these parameters across multiple splits is a straightforward robustness check."],"forward_implications":["On the ATLAS dataset, D-UNet improves per-case DSC by about 3.8 percentage points over a tuned 2D UNet, while increasing total parameters by only about 2%.","The hybrid encoder reaches a higher per-case DSC and precision than a pure 3D UNet while avoiding the 3D network's heavy memory and training-time burden.","Fusion after the second and third downsampling stages (SEAdd-23) gives the best DSC; fusing at all three stages (SEAdd-123) lowers DSC, indicating the fusion depth has an optimum.","The Enhanced Mixing Loss converges faster on the training curve than focal loss or Dice loss alone, and it achieves the highest precision among the three losses tested."],"supporting_citations":[{"why":"Supplies the ATLAS dataset of 229 chronic stroke T1-weighted MRIs with manual lesion segmentations used for training and validation.","marker":"[7]"},{"why":"Provides the 2D U-Net architecture that D-UNet starts from and that serves as its main baseline.","marker":"[33]"},{"why":"Defines the focal loss whose voxel-averaged form enters the proposed Enhanced Mixing Loss.","marker":"[31]"},{"why":"Defines the Dice loss whose negative logarithm is combined with focal loss in the Enhanced Mixing Loss.","marker":"[32]"},{"why":"Supplies the squeeze-and-excitation blocks used to reweight 2D and 3D feature channels before fusion.","marker":"[34]"},{"why":"A hybrid 2D/3D segmentation network whose finding that 3D structures can underperform 2D on this dataset motivates the lightweight fusion design.","marker":"[29]"},{"why":"Supplies the traditional lesion-segmentation baselines (Clusterize, ALI, LINDA, and others) that D-UNet is compared against.","marker":"[37]"},{"why":"A strong segmentation baseline (DeepLab v3 plus) that D-UNet outperforms on the same train/validation split.","marker":"[40]"}],"fun_headline_variants":["D-UNet fuses 2D-3D for faster, sharper stroke lesion segmentation","Hybrid 2D-3D U-Net tops both pure 2D and 3D on stroke lesions","Dimension-fusion U-Net cuts compute, beats 2D on stroke segmentation","New loss function helps fused U-Net reach 0.535 DSC on stroke","Efficient 2D-3D fusion nets top stroke lesion accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported advantage of D-UNet over the 2D UNet is attributed to the architecture itself, rather than to the specific random train/validation split or the loss parameters tuned on that split.","fun_headline_variants_meta":{"raw":{"variants":["D-UNet fuses 2D-3D for faster, sharper stroke lesion segmentation","Hybrid 2D-3D U-Net tops both pure 2D and 3D on stroke lesions","Dimension-fusion U-Net cuts compute, beats 2D on stroke segmentation","New loss function helps fused U-Net reach 0.535 DSC on stroke","Efficient 2D-3D fusion nets top stroke lesion accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000512,"raw_usage":{"total_tokens":2517,"prompt_tokens":998,"completion_tokens":1519,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":1418}},"tokens_in":614,"tokens_out":1519,"duration_ms":11455,"temperature":1.0,"reasoning_tokens":1418,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:22:49.251924+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun D-UNet and the tuned 2D UNet on ATLAS across several random train/validation splits with the loss hyperparameters held at their default values; if the mean DSC gap (reported as 0.038) falls within the per-case standard deviation of about 0.28, the architecture's benefit over the 2D baseline would not be established.","supporting_citations":[{"cited_title":"A large, open source dataset of stroke anatomical brain images and manual lesion segmentations[J]","cited_arxiv_id":null,"evidence_quote":"Supplies the ATLAS dataset of 229 chronic stroke T1-weighted MRIs with manual lesion segmentations used for training and validation."},{"cited_title":"U-net: Convolutional networks for biomedical image segmentation[C]","cited_arxiv_id":null,"evidence_quote":"Provides the 2D U-Net architecture that D-UNet starts from and that serves as its main baseline."},{"cited_title":"Focal loss for dense object detection[C]","cited_arxiv_id":null,"evidence_quote":"Defines the focal loss whose voxel-averaged form enters the proposed Enhanced Mixing Loss."},{"cited_title":"V-net: Fully convolutional neural networks for volumetric medical image segmentation[C]","cited_arxiv_id":null,"evidence_quote":"Defines the Dice loss whose negative logarithm is combined with focal loss in the Enhanced Mixing Loss."},{"cited_title":"Squeeze-and-excitation networks[C]","cited_arxiv_id":null,"evidence_quote":"Supplies the squeeze-and-excitation blocks used to reweight 2D and 3D feature channels before fusion."},{"cited_title":"H-DenseUNet: hybrid densely connected UNet for liver and tumor segmentation from CT volumes[J]","cited_arxiv_id":null,"evidence_quote":"A hybrid 2D/3D segmentation network whose finding that 3D structures can underperform 2D on this dataset motivates the lightweight fusion design."},{"cited_title":"A comparison of automated lesion segmentation approaches for chronic stroke T1-weighted MRI data[J]","cited_arxiv_id":null,"evidence_quote":"Supplies the traditional lesion-segmentation baselines (Clusterize, ALI, LINDA, and others) that D-UNet is compared against."},{"cited_title":"Encoder-decoder with atrous separable convolution for semantic image segmentation[C]","cited_arxiv_id":null,"evidence_quote":"A strong segmentation baseline (DeepLab v3 plus) that D-UNet outperforms on the same train/validation split."}],"review_version":1}