{"id":"a89b3607-91eb-4bdd-a4b8-38971d647837","arxiv_id":"2412.00888","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A dual-parallel encoder CNN reports slightly higher polyp segmentation Dice and mIoU than several prior networks on Kvasir and CVC-ClinicDB.","lead":"This paper describes DPE-Net, a polyp segmentation network with two parallel encoder branches, one built from paired convolution blocks and one from single residual-style blocks, merged before upsampling. It reports Dice scores of 0.919 on Kvasir and 0.931 on CVC-ClinicDB, claiming gains over several published baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table II contradicts the central claim: Polyp-PVT's reported mDice (0.937) exceeds DPE-Net's (0.931) on CVC-ClinicDB, and Deeplabv3+'s Kvasir row has mIoU > mDice, so the SOTA superiority claim is not supported.","rationale":"The paper's only quantitative support for the central claim is Table II, a comparison of mDice/mIoU values against six baselines on two datasets. Two independent defects make that table unusable. First, in the Kvasir row for Deeplabv3+, mIoU (0.915) exceeds mDice (0.899), which is mathematically impossible for any binary segmentation result because IoU = Dice/(2−Dice) ≤ Dice. Thus the baseline numbers were not produced by the metric definitions given in Sec. III-A. Second, even if the numbers were taken at face value, the central claim that DPE-Net achieves the highest mDice and mIoU is false on CVC-ClinicDB: Polyp-PVT is listed with mDice 0.937, above DPE-Net's 0.931. The abstract, discussion, and conclusion assert DPE-Net 'leads the pack' with highest values, so the paper contradicts itself. Since the reported margins over the closest rival on Kvasir are 0.002 (Dice) and 0.002–0.007 (IoU), without a shared, verified evaluation protocol these differences are within typical run-to-run variation and cannot support the claim of superior performance. The reader's weakest-assumption analysis correctly identifies Table II comparability as the central vulnerability; the metric violation is a concrete, checkable symptom. My stress test therefore leaves the REJECT verdict unchanged.","tokens_in":8177,"tokens_out":5054,"duration_ms":41981,"concrete_test":"Compute IoU from each reported Dice in Table II using the exact relation IoU = Dice/(2−Dice) for binary segmentation. The Deeplabv3+ Kvasir row reports mDice 0.899, which implies mIoU ≈ 0.816, not 0.915; this single violation proves the table mixes incompatible evaluation protocols. If the check fails for any baseline row, the superiority claim against those baselines is unsubstantiated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"DPE-Net's central claim—highest mDice and mIoU on both datasets—fails against the paper's own Table II. On CVC-ClinicDB, Polyp-PVT is reported with mDice 0.937 vs DPE-Net 0.931; the abstract and discussion claim DPE-Net 'leads the pack' with highest mDice, which is false as printed. On Kvasir, Deeplabv3+ is listed with mDice 0.899 and mIoU 0.915. For any binary segmentation, Dice = 2IoU/(1+IoU), so IoU <= Dice always; the reported values cannot both be correct. This shows the baseline numbers are not produced by a single consistent evaluation protocol. Since the claimed improvements over the closest baselines are marginal (e.g., 0.002–0.004 on Kvasir), even a small protocol mismatch erases the advantage. The reader's concern about non-comparable baselines is therefore load-bearing: the evidence table is internally inconsistent and contradicts the paper's own superiority claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DPE-Net, a dual-parallel encoder network for polyp segmentation. One encoder branch uses dual convolution blocks with a residual-style mapping, and the other uses single convolution blocks with identity-style residual connections; the two branches are merged by depth-wise concatenation before transpose-convolution upsampling. Experiments are reported on Kvasir and CVC-ClinicDB, with an ablation study (Table I) and a comparison against UNet, SegNet, DeepLabv3+, PraNet, Polyp-PVT, and MEGANet (Table II). The authors claim the highest mDice and mIoU on both datasets, with 3.4 million parameters and 6-hour training on an RTX 3060.","tokens_in":8550,"tokens_out":2296,"duration_ms":20871,"significance":"If the reported results were reliable, a 3.4M-parameter encoder with parallel dual and single convolution branches would be a useful lightweight alternative for polyp segmentation, and the public-dataset evaluation would be a practical contribution. The paper also includes an ablation study isolating the two encoder branches, which is a positive feature. However, the central quantitative evidence is compromised: Table II contains a mathematically impossible metric pair and directly contradicts the stated superiority claim on CVC-ClinicDB. Because the main contribution is the claimed state-of-the-art performance, these problems are load-bearing and prevent the paper from being accepted in its current form.","major_comments":[{"comment":"The DeepLabv3+ row on Kvasir reports mDice = 0.899 and mIoU = 0.915. For any binary segmentation, Dice = 2*IoU/(1+IoU), so IoU is always less than or equal to Dice. These two reported values cannot both be correct for the same set of predictions, indicating that the baseline numbers were not produced under a single consistent evaluation protocol. This invalidates the comparison basis for the entire table.","section":"Section III-B, Table II"},{"comment":"The paper's central claim of 'superior performance' and 'highest mDice and mIoU values' is contradicted by its own Table II: on CVC-ClinicDB, Polyp-PVT is reported with mDice = 0.937, which exceeds DPE-Net's mDice = 0.931. The claimed advantage over Polyp-PVT on Kvasir is only 0.002 in mDice and 0.002 in mIoU, which is within run-to-run variation, yet no error bars, standard deviations, or statistical tests are reported anywhere in the manuscript.","section":"Abstract and Section III-B, Table II"},{"comment":"The manuscript does not specify whether the baseline models were retrained on the same 80/10/10 split with the same preprocessing and image resolution (384x288) or whether their numbers were copied from prior publications. The inconsistent metric pair in Table II suggests that at least some baseline values come from incompatible protocols. Without a standardized experimental setup applied to all methods, the claimed superiority of DPE-Net is not established.","section":"Section III-B"},{"comment":"The ablation study reports that the proposed DPE-Net achieves accuracy 0.971 on Kvasir and 0.971 on CVC-ClinicDB, while Network 3 (both parallel blocks with LR 10^-3) achieves accuracy 0.921 and 0.918. The only stated difference between Network 3 and DPE-Net appears to be the learning rate (10^-3 vs 1e-4, described in Section III), but this is not explained in the ablation discussion, leaving the source of the substantial accuracy gain unclear.","section":"Section III-C, Table I"}],"minor_comments":[{"comment":"Equations (3) and (4) are garbled in the text; the definitions of Dice and IoU should be written with clear notation for true positives, false positives, and false negatives. The current rendering is not readable.","section":"Section III-A, Eqs. (3)-(4)"},{"comment":"The introduction cites 'MSRF-Net [28]' and 'Polyp-PVT [28]' with the same reference number; reference [28] is listed as Polyp-PVT only, so the MSRF-Net citation appears to be an error.","section":"Introduction"},{"comment":"The paper refers to visual comparisons in Figures 3 and 4, but the figures as provided are largely unreadable in the manuscript text, with missing or broken image placeholders. The visual evidence cannot be assessed.","section":"Section III-B, Figures 3-4"},{"comment":"The architectural description is incomplete: the number of layers, kernel sizes for each stage, channel widths, the exact placement of batch normalization and max-pooling, and the decoder structure after the transpose convolution are not specified with sufficient detail for reproduction.","section":"Section II"},{"comment":"The paper states 'the datasets were split using an 80-10 ratio' and then says 'the remaining 10% was used for validation purposes,' which is ambiguous (80/10/10 presumably). The sentence should be clarified.","section":"Section III"}],"recommendation":"reject","confidential_remarks":"The central evidence table (Table II) is internally inconsistent and contradicts the abstract's performance claims. This is not a presentation issue that can be fixed locally; the comparison would need to be re-run under a unified protocol with statistical reporting, and the claims would need to be rewritten. Given the journal context, I do not see a path to acceptance without a major new experimental study. I also note that the reference list contains several self-citations and near-duplicate references to the same group's work, which is worth checking for relevance during any future resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new bit here is the specific encoder configuration: two parallel branches, one built from dual convolution blocks and one from single-convolution residual blocks, merged by depth-wise concatenation before the transpose layers. That exact combination is not in the prior work I know, and the model is pleasantly lightweight at 3.4M parameters. The visual results on Kvasir and CVC-ClinicDB look plausible, and the architecture description, though terse, is understandable. The soft spots are not minor, and they are load-bearing. Table II, which is the entire quantitative case, contains rows that violate the metric formulas the paper itself gives. Deeplabv3+ on Kvasir is listed with mIoU 0.915 and mDice 0.899; for any segmentation mask, IoU cannot exceed Dice. SegNet on CVC-ClinicDB has mDice 0.654 and mIoU 0.767, which is the same impossible direction. These are not rounding artifacts; they indicate the baseline numbers were not produced under one consistent protocol. More damaging, the claim of leading the pack on both metrics is false as printed: on CVC-ClinicDB, Polyp-PVT is listed with mDice 0.937, above DPE-Net's 0.931. So the abstract and discussion overstate the result even relative to the paper's own table. The evaluation also lacks error bars, seeds, and statistical tests, and no code is provided. The baseline numbers appear to be copied from earlier publications rather than retrained on the same 80/10/10 split, and the ablation study is confusing: Network 3 already has both blocks and an LR of 1e-3, yet DPE-Net is reported separately with no explanation of what actually changed. There is also a citation slip: MSRF-Net is mentioned in the introduction but shares a reference number with Polyp-PVT. What the paper does well is present a lightweight architecture idea that someone might want to try or adapt. But the evidence that it beats established methods does not survive contact with the table. A corrected version with code, retrained baselines, and a clean metric table might be worth a workshop-level submission; as it stands, I would not spend referee time on it.","headline":"A plausible lightweight dual-branch encoder idea, but the paper's own Table II contains impossible metric values and contradicts the claimed superiority, so the empirical case collapses.","tokens_in":643,"tokens_out":941,"would_cite":false,"duration_ms":29629,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that DPE-Net, a two-branch encoder network combining dual and single convolution blocks with depth-wise concatenation, outperforms six established deep-learning models for polyp segmentation on Kvasir and CVC-ClinicDB…","keywords":["polyp segmentation","dual parallel encoder","semantic segmentation","colonoscopy images","deep learning","Kvasir-SEG","CVC-ClinicDB","residual learning"],"falsifier":"Retrain every baseline on the exact same 80/10/10 splits of Kvasir and CVC-ClinicDB with the same loss, input size, and metric code, and check that each reported pair obeys $\\mathrm{IoU} \\le \\mathrm{Dice}$; if Polyp-PVT or MEGANet then matches or exceeds DPE-Net's mDice/mIoU, or if the Deeplabv3+ Kvasir row cannot be reproduced, the central claim of superiority would be contradicted.","tokens_in":7955,"feed_emoji":"🩺","tokens_out":8789,"duration_ms":57990,"temperature":0.7,"pith_summary":"DPE-Net is a polyp segmentation architecture built from two encoder branches running in parallel: one uses dual convolution blocks to carry feature information through deeper layers, and the other uses single convolution blocks with a residual identity connection to preserve the input-like character of features. The branch outputs are merged by depth-wise concatenation before each transpose-convolution up-sampling stage, followed by a sigmoid pixel classifier. The paper's central claim is that this two-branch design reaches higher mean Dice and mean IoU than six established deep-learning models on Kvasir and CVC-ClinicDB, with mDice/mIoU of 0.919/0.866 on Kvasir and 0.931/0.891 on CVC-ClinicDB. It also claims the gain comes from combining the two branches, since an ablation shows either branch alone scores far lower. A sympathetic reader would care because the model uses only 3.4 million parameters and trains in about six hours on consumer hardware, which suggests a lightweight alternative for a clinically relevant segmentation task.","feed_headline":"Dual-branch encoder reports best polyp segmentation scores","feed_subtitle":"With 3.4M parameters, it posts the top Dice and IoU on Kvasir and CVC-ClinicDB.","key_machinery":"The central mechanism is the dual-branch encoder. Branch A is the dual convolution block: 1x1 and 3x3 convolutions with ReLU and batch normalization, producing a residual update $I_{k+1} = M + S$, where $M$ is the mapped feature input and $S$ is a 1x1-convolution/batch-norm projection that aligns channel counts. Branch B is the single convolution identity block: a 3x3 convolution with ReLU and batch normalization plus an element-wise skip, written $I_{k+1} = F_{I_1} + I_1$, where $I_1$ is the input feature map and $F_{I_1}$ its convolved version. The two branches are fused with depth-wise concatenation before each transpose-convolution layer, and the final segmentation comes from a sigmoid classifier. The pair of update rules is what the paper says maintains feature information over depth while adding diversity, and the ablation results are the evidence that the combination, not either branch alone, carries the performance.","core_discovery":"On the paper's own terms, DPE-Net establishes that running a dual-convolution encoder branch and a single-convolution residual branch in parallel, then merging them depth-wise before up-sampling, yields the best quantitative segmentation performance among the methods compared. The reported numbers are mDice 0.919 and mIoU 0.866 on Kvasir, and mDice 0.931 and mIoU 0.891 on CVC-ClinicDB, ahead of UNet, SegNet, Deeplabv3+, PraNet, Polyp-PVT, and MEGANet. The ablation study backs the design story: the dual-convolution branch alone gives mDice 0.640 on Kvasir and 0.552 on CVC-ClinicDB, the single-convolution branch alone gives 0.632 and 0.475, and the combined network with learning rate $10^{-3}$ jumps to 0.901 and 0.898 before the final configuration reaches 0.919 and 0.931. The authors read this as evidence that balancing depth-rich feature extraction with identity-preserving feature extraction is what drives the improvement.","pith_inferences":["An implication the authors leave implicit: the dual and single branches may be learning complementary feature types, so inspecting where each branch fires on polyp boundaries versus interior could motivate fusion rules beyond depth-wise concatenation.","A testable extension the paper does not run: pairing the same encoder with attention or transformer decoders, which are standard options in current polyp segmentation work, to see whether the encoder's gains persist.","The reported 3.4M-parameter size and six-hour training time point toward real-time clinical use, but the paper does not report inference latency, so measuring frames-per-second would settle that implicit promise."],"forward_implications":["If the reported comparison holds, DPE-Net is a lighter alternative to the six compared models, using about 3.4 million parameters while posting the highest mDice and mIoU on both datasets.","The ablation indicates that deleting either branch is costly, so the parallel combination itself is the source of the gain; designs that keep both streams should generalize better than choosing one block type.","The same dual/single encoder pattern can be dropped into other encoder-decoder architectures at the pre-transpose fusion point, since the blocks are not specific to polyp datasets.","A 40-epoch SGDM schedule, 384x288 input size, and mixed-precision training give a reproducible six-hour training budget on an RTX 3060, which is practical for clinical data pipelines."],"supporting_citations":[{"why":"Supplies the CVC-ClinicDB colonoscopy dataset and its 288x384 native image size used for training and evaluation.","marker":"[37]"},{"why":"Supplies the Kvasir dataset of 1,000 polyp images, the second benchmark for the comparison.","marker":"[38]"},{"why":"Defines the UNet baseline whose mDice and mIoU appear in Table II.","marker":"[39]"},{"why":"Defines the SegNet/VGG16 baseline used in the quantitative comparison.","marker":"[40]"},{"why":"Defines the Deeplabv3+ with ResNet-50 baseline used in the comparison.","marker":"[41]"},{"why":"Defines the PraNet baseline, a close competitor in the reported results.","marker":"[27]"},{"why":"Defines the Polyp-PVT baseline, the nearest reported competitor on Kvasir and CVC-ClinicDB.","marker":"[28]"},{"why":"Defines the MEGANet baseline, another recent method the paper compares against.","marker":"[29]"}],"fun_headline_variants":["Dual-parallel encoder lifts polyp segmentation scores","Twin-branch encoder outperforms single-path nets on polyps","DPE-Net: parallel encoders beat UNet and PraNet on polyps","Parallel branches boost polyp segmentation to top Dice","Two-branch encoder achieves best polyp segmentation yet"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the baseline scores in Table II were produced under the same data split and evaluation protocol as DPE-Net; the paper does not state that the baselines were retrained on the same 80/10/10 split, so if those numbers came from different protocols the reported superiority over them would not be established.","fun_headline_variants_meta":{"raw":{"variants":["Dual-parallel encoder lifts polyp segmentation scores","Twin-branch encoder outperforms single-path nets on polyps","DPE-Net: parallel encoders beat UNet and PraNet on polyps","Parallel branches boost polyp segmentation to top Dice","Two-branch encoder achieves best polyp segmentation yet"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000798,"raw_usage":{"total_tokens":3522,"prompt_tokens":965,"completion_tokens":2557,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":2476}},"tokens_in":581,"tokens_out":2557,"duration_ms":19004,"temperature":1.0,"reasoning_tokens":2476,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:52:56.521012+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain every baseline on the exact same 80/10/10 splits of Kvasir and CVC-ClinicDB with the same loss, input size, and metric code, and check that each reported pair obeys $\\mathrm{IoU} \\le \\mathrm{Dice}$; if Polyp-PVT or MEGANet then matches or exceeds DPE-Net's mDice/mIoU, or if the Deeplabv3+ Kvasir row cannot be reproduced, the central claim of superiority would be contradicted.","supporting_citations":[{"cited_title":"Kendall, and R","cited_arxiv_id":null,"evidence_quote":"Defines the SegNet/VGG16 baseline used in the quantitative comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Deeplabv3+ with ResNet-50 baseline used in the comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the PraNet baseline, a close competitor in the reported results."},{"cited_title":"MEGANet: Multi-Scale Edge-Guided Attention Network for Weak Boundary Polyp Segmentation","cited_arxiv_id":"2309.03329","evidence_quote":"Defines the MEGANet baseline, another recent method the paper compares against."}],"review_version":1}