{"id":"a45d6d78-7c38-4fe6-a7bb-15d529037611","arxiv_id":"2412.02443","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MMCC-Net, a multi-scale multi-path cascaded CNN with attention and feature enhancement, reports Dice scores from 77.4 to 94.5 across six polyp datasets, narrowly beating prior models.","lead":"This paper introduces MMCC-Net, a convolutional network for segmenting colorectal polyps in colonoscopy images, and reports small improvements over eight prior models on six public datasets. A generalist might read it to see whether a lightweight CNN can still edge out larger transformer models on a medical segmentation benchmark.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed superiority over the strongest baselines is not statistically established: Table 2 shows overlapping 95% CIs, no paired significance tests, internally inconsistent CI calculations, and a possible video-level split leak in CVC-ClinicDB.","rationale":"The reader's weakest assumption identifies exactly the load-bearing weakness: the paper claims superior performance while reporting overlapping confidence intervals and omitting paired significance tests. My independent reading confirms this is the single most important threat to the central claim. The strongest baseline differences on CVC-ClinicDB and Kvasir are 0.02–0.13 points in Dice, well within the reported uncertainties. I also verified that the CI arithmetic is internally inconsistent, which further weakens the reported statistics rather than merely the interpretation. The CVC-ClinicDB video-sequence leakage concern is a second concrete threat: the paper cites FCB-SwinV2's warning about this dataset but gives no evidence that its own split avoids the problem. These issues are addressable, as the reader noted, by releasing code and weights, reporting paired tests, and clarifying the split protocol. I therefore agree with the CONDITIONAL verdict; no adjustment is needed. The paper does have some independent support: a parameter count of 1.43M, an ablation study, and consistent qualitative figures, but none of these substitute for the missing statistical evidence that would justify the headline claim of superiority.","tokens_in":28625,"tokens_out":2749,"duration_ms":28630,"concrete_test":"Re-run the Kvasir and CVC-ClinicDB experiments using the identical trained models and test folds from Table 2; for each image, compute per-image Dice for MMCC-Net and FCB-SwinV2 (and the other SOTA baselines). Apply a paired Wilcoxon signed-rank test over the 10 runs on the per-image Dice differences; if p > 0.05 for the closest baseline, the claimed superiority is not established. Separately, check whether any CVC-ClinicDB video sequence contributes frames to both train and test/validation under the Table 1 split; if so, retrain MMCC-Net and all baselines with a video-level split and recompute Table 2. Report the CI derivation explicitly (n, t-value, SD) to confirm the intervals are correctly computed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—MMCC-Net outperforms eight SOTA models—rests on tiny mean differences whose uncertainty is not accounted for. On CVC-ClinicDB (Table 2), MMCC-Net Dice is 94.45 ± 0.12 with 95% CI (94.19, 94.71), while FCB-SwinV2 is 94.43 ± 0.13 with CI (94.17, 94.69); the difference is 0.02 points and the intervals overlap almost entirely. On Kvasir, MMCC-Net Dice is 92.65 vs. FCB-SwinV2's 92.52, again with overlapping CIs. No paired significance test (e.g., Wilcoxon signed-rank or paired bootstrap) is reported for any comparison, so the reported 'superior performance' has no demonstrated statistical support. The CI calculations themselves are internally inconsistent: with n=10 runs and t=2.262, MMCC-Net's Kvasir Dice mean 92.65 and SD 0.13 imply CI (92.56, 92.74), not the reported (92.65, 93.25); similar inconsistencies appear in other rows, undermining the credibility of the reported intervals. A second, independent threat is dataset leakage: the paper explicitly flags video-sequence leakage in CVC-ClinicDB when discussing FCB-SwinV2, but does not state whether its own 90/10 split in Table 1 is video-aware. Since CVC-ClinicDB contains 612 images from 29 videos, a random split can place frames from the same video in both training and test sets, inflating MMCC-Net's ClinicDB numbers. If the supposed superiority is not real, the paper's primary contribution—a lightweight CNN that beats transformers—collapses to an incremental architecture study with plausible but unvalidated components.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes MMCC-Net, a lightweight CNN for colorectal polyp segmentation built from multi-scale multi-path cascaded convolutions, dense skip connections, two attention modules, and a feature enhancer, trained with a joint Dice and binary cross-entropy loss. Experiments are conducted on six public datasets (Kvasir, CVC-ClinicDB, CVC-300, ETIS, CVC-ColonDB, EndoCV2020) and compared with eight published models. The paper reports Dice scores from 77.43 to 94.45 and MIoU from 72.71 to 90.16 across datasets, approximately 1.43M parameters, repeated 10-run statistics, 5-fold cross-validation, ablations, loss/optimizer/LR sensitivity, HDD/AUC, and efficiency comparisons. The central claim is that MMCC-Net outperforms all eight SOTA models while being substantially more parameter-efficient.","tokens_in":29018,"tokens_out":6516,"duration_ms":61202,"significance":"The intended contribution is a parameter-efficient CNN that matches or beats transformer-based segmenters; such a model would have practical value for colonoscopy workflows. The paper's strengths are its breadth (six datasets, eight baselines), repeated-run design, ablation of architectural modules, and explicit discussion of failure cases and deployment issues. However, the statistical basis for the headline claim is not established: the largest advantages over the strongest baselines are tiny (0.02-0.13 Dice points), confidence intervals overlap, no paired significance tests are provided, and several reported intervals are arithmetically impossible. The additional risk of video-level leakage in the CVC-ClinicDB split, acknowledged in the paper's own discussion of FCB-SwinV2, could materially inflate the reported results. With corrected statistics and a leakage-free evaluation, the paper could be a solid empirical study, but the current evidence does not support the claimed superiority.","major_comments":[{"comment":"The claim that MMCC-Net \"consistently outperforms\" eight SOTA models is not supported by Table 2. On CVC-ClinicDB, MMCC-Net Dice is 94.45 ± 0.12 with 95% CI (94.19, 94.71) versus FCB-SwinV2 94.43 ± 0.13 (94.17, 94.69); the difference is 0.02 points and the intervals overlap almost entirely. On Kvasir, the corresponding difference is 0.13 Dice points with overlapping intervals. No paired significance test (Wilcoxon signed-rank, paired bootstrap, or corrected resampled t-test) is reported for any comparison, so the reported \"superior performance\" has no demonstrated statistical support. Please add paired tests over the 10 runs and report exact p-values or bootstrap CIs for each SOTA comparison.","section":"Table 2"},{"comment":"The confidence intervals appear internally inconsistent. With n=10 runs and t_{0.025,9}=2.262, the Kvasir proposed Dice mean 92.65 and SD 0.13 imply a 95% CI of approximately (92.56, 92.74), not the reported (92.65, 93.25); the lower bound cannot equal the mean. Similar discrepancies appear in many rows (e.g., several CIs have both endpoints above the mean). Please recompute all CIs from the actual per-run results and state the critical value and formula used.","section":"Tables 2 and 4"},{"comment":"The CVC-ClinicDB split is at risk of video-level data leakage. The dataset consists of 612 images from 29 colonoscopy videos, and Table 1 shows a 490/61/61 train/validation/test split, but the manuscript nowhere states that the split is video-aware. Section 2.2 itself notes that FCB-SwinV2 highlights video-sequence data leakage in CVC-ClinicDB. If frames from the same video appear in both training and test partitions, the reported ClinicDB numbers are inflated. Please either confirm that the split was performed at the video level and describe the procedure, or rerun Experiment 1 with a video-aware split.","section":"Section 4.1 / Table 1"},{"comment":"Table 3 (5-fold cross-validation) contradicts the text's claim that MMCC-Net \"outperforms all other models regarding mDice, MIoU, precision, and recall across both datasets.\" On Kvasir, MMCC-Net's Dice is 92.29 ± 0.22, lower than PVT-CASCADE's 92.49 ± 0.31. This is a direct internal inconsistency in the central comparative claim; please correct the table or the text and discuss the discrepancy.","section":"Table 3"},{"comment":"The loss equations are not usable as written. Eq. (4) defines an L2 Dice loss, Eq. (8) defines Lseg = LDice + LBce, and Eq. (11) introduces α and γ in a placement that is dimensionally inconsistent with Eqs. (8)-(10); the grad-CAM text and Eq. (7) are also garbled. Since the joint loss and its hyperparameters (α=0.22, γ=1.9) are part of the method, please rewrite the loss derivation with consistent notation and verify every equation.","section":"Section 3.2 / Eqs. (4)-(11)"},{"comment":"The manuscript does not state whether α, γ, the learning rate, and other architectural choices were selected using the same test folds on which the final numbers are reported. If these hyperparameters were tuned on the CVC-ClinicDB and Kvasir test partitions, the reported confidence intervals are selection-conditional and would be optimistic. Please describe the model selection protocol used for each experiment.","section":"Section 4.3.1 / Table 8"}],"minor_comments":[{"comment":"The text states that the study uses eight evaluation measures but then enumerates nine: Dice, accuracy, sensitivity, precision, specificity, IOU, AUC, CI, and HDD.","section":"Section 4.2"},{"comment":"The architecture diagram is difficult to read at the resolution provided; please supply a higher-resolution version with all modules and pathways clearly labeled.","section":"Figure 3"},{"comment":"The Hausdorff distance formula is incorrect as printed: the second term should be max over b in B of min over a in A of ||b - a||, not another maximization over a in A.","section":"Eq. (17)"},{"comment":"The citations \"FCB-Former [27]\" and \"FCB-SwinV2 [32]\" appear to be wrong; these should likely refer to references [47] and [52], respectively.","section":"Section 4.3.2"},{"comment":"Tables 5 and 6 report CVC-ClinicDB Dice of 94.41 and Kvasir Dice of 92.40, while Table 2 reports 94.45 and 92.65 for the same model; please clarify whether these are the same 10-run averages or different runs.","section":"Tables 2, 5, and 6"},{"comment":"The training-time description is inconsistent: the text says 80 epochs took about 8 hours with no improvement beyond 60 epochs (about 6 hours), while Table 10 reports a training time of 6.0 hours; please align these statements.","section":"Section 4.3"},{"comment":"Notation is inconsistent between \"IOU\" in Eq. (16) and \"MIoU\" in the tables; please unify the terminology and define HDD before first use in Section 4.2.","section":"Equations (12)-(18)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: MMCC-Net is a sensible lightweight CNN for polyp segmentation, but the paper does not establish that it beats the strongest baselines. Against FCB-SwinV2 the Dice differences are 0.02–0.13 points with overlapping 95% CIs, and no paired significance test is reported. The CI arithmetic is internally inconsistent in several rows (e.g., Kvasir MMCC-Net Dice 92.65±0.13 with n=10 cannot give CI (92.65, 93.25)), and the CVC-ClinicDB split is not shown to be video-aware even though the paper itself flags exactly that leakage risk when discussing FCB-SwinV2. The headline claim of superior performance is therefore not backed by the evidence as presented.\n\nWhat is genuinely new: the specific multi-scale multi-path cascade with dual attention, dense skip paths, and a feature enhancer is a new assembly of known components, and the parameter count (1.43M) is genuinely small compared to the transformer baselines. The paper also does a full experimental run: six datasets, eight baselines, ablations of the modules, loss-function comparisons, optimizer and learning-rate sweeps, HDD/AUC tables, and efficiency figures. That is more than the usual incremental architecture study. The Grad-CAM visuals and the failure-case section are also a plus.\n\nWhere it falls short: the statistical case is the load-bearing weakness. Beyond the missing paired tests and the CI inconsistencies, the paper does not report per-run results, so a referee cannot check the intervals. The possible video-level leakage in CVC-ClinicDB is a real concern; the authors need to state explicitly how the split was made. The loss equations (Eq. 4–11) are garbled and internally inconsistent, and no code or weights are released, so the implementation cannot be verified as is. These issues are addressable, but they are not minor: they undercut the central claim.\n\nAlso, the hyperparameters alpha and gamma were tuned on the same datasets used for the final numbers, so the reported results are optimistic for those datasets. The parameter-efficiency property is real, though, even if the performance edge over transformers is not.\n\nWho this is for: readers interested in lightweight CNNs for low-resource clinical polyp segmentation will find the architecture and ablations informative. This is not a breakthrough paper, but it is a legitimate engineering study with a useful efficiency angle. It deserves a serious referee: the topic is relevant, the experiments are extensive, and the claims are falsifiable if the authors supply code, paired significance tests, and a video-aware split. I would send it to review, expecting major revision on statistical evidence and reproducibility.","headline":"A sensible lightweight CNN with thorough experiments, but the claimed statistical superiority over strong baselines is not supported by the reported confidence intervals and split protocol.","tokens_in":29577,"tokens_out":2864,"would_cite":false,"duration_ms":29385,"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 compact CNN, MMCC-Net, is claimed to outperform eight state-of-the-art models on pixel-level colorectal polyp segmentation across six public datasets using about 1.43 million parameters.","keywords":["colorectal polyp segmentation","semantic segmentation","multi-scale CNN","cascaded convolution","attention modules","feature enhancer","dice loss","colonoscopy"],"falsifier":"Run a paired per-image significance test, such as a Wilcoxon signed-rank or bootstrap test on per-image Dice, between MMCC-Net and FCB-SwinV2 and PVT-CASCADE on a CVC-ClinicDB split that keeps frames from the same colonoscopy video entirely in either training or testing; if the difference is not significant at p < 0.05, the paper's central outperformance claim fails.","tokens_in":28441,"feed_emoji":"🩺","tokens_out":5436,"duration_ms":52640,"temperature":0.7,"pith_summary":"This paper introduces MMCC-Net, a convolutional network for pixel-level segmentation of colorectal polyps in colonoscopy images. The authors claim that by routing features through multiple parallel paths at several scales and fusing them with dense skip connections, attention modules, and a feature enhancer, the network captures both fine edges and global context better than existing designs. On six public polyp datasets, they report Dice scores from 77.43 to 94.45 and mean IoU from 72.71 to 90.16, with point estimates above eight state-of-the-art baselines. The network uses about 1.43 million parameters, far fewer than most comparators, so the claim matters for clinical tools that need speed and low compute. If correct, accurate polyp localization could run cheaply and in near real time during colonoscopy.","feed_headline":"1.4M-parameter CNN beats eight polyp segmentation models","feed_subtitle":"MMCC-Net posts the highest Dice and IoU on six colonoscopy datasets with far fewer parameters.","key_machinery":"The load-bearing mechanism is the dense multi-scale feature aggregation written as $DFA = Fa \\otimes F1 \\otimes F2 \\otimes F4$, $DFB = AFi \\otimes E1 \\otimes E2 \\otimes E4 \\otimes AFB$, and $DFC = DFB \\otimes AFA$, where $\\otimes$ is depth-wise concatenation. Three parallel routes produce features at different dilation and stride factors, a mid-block adds an attention-filtered path, and a feature enhancer preserves low-level spatial cues. This cascade lets the decoder combine local edges, small-polyp details, and broad context, while the joint loss $L_{seg} = L_{Dice} + L_{BCE}$ with L2 smoothing on the Dice term addresses class imbalance.","core_discovery":"The paper's central claim is that a carefully balanced convolutional network can outperform transformer-based and hybrid competitors on polyp segmentation without large parameter counts. MMCC-Net couples multi-scale and multi-path cascaded convolutions with three routes for feature fusion, two attention modules, and a feature enhancer, and it is trained with a joint Dice-plus-cross-entropy loss with L2 smoothing on the Dice term. Across Kvasir, CVC-ClinicDB, CVC-300, ETIS, CVC-ColonDB, and EndoCV2020, the authors report that MMCC-Net achieves the highest point estimates for Dice, MIoU, precision, recall, accuracy, and the lowest Hausdorff distance among the eight compared methods, with narrow confidence intervals. They interpret this as evidence that strong global context can come from cascaded convolutions and dense feature aggregation rather than from transformers.","pith_inferences":["Because the reported margins over the strongest baselines are often a few tenths of a percentage point and some confidence intervals overlap, a reader should treat the superiority claim as provisional until paired significance tests are reported.","A natural next check is to evaluate MMCC-Net on a CVC-ClinicDB split that keeps frames from the same colonoscopy video entirely in either training or testing, following the data-leakage caveat the paper itself cites for FCB-SwinV2.","The ablation pattern suggests the feature enhancer, rather than skip connections alone, drives much of the improvement; isolating that component on harder small-polyp datasets would be a direct stress test.","The combination of Grad-CAM heatmaps and a compact architecture points toward a screening assistant role, though real clinical deployment would require testing on diverse imaging equipment and lighting conditions."],"forward_implications":["At roughly 1.43 million parameters and 20.85 G FLOPs, the model could run on modest hardware or edge devices in colonoscopy suites.","A joint Dice-plus-BCE loss with L2 smoothing can handle polyp/background imbalance without explicit class weighting, a recipe transferable to other lesion segmentation tasks.","Multi-scale dense fusion with few filters per layer may generalize to small or irregular polyps because low-level edge cues are preserved through the feature enhancer.","The same architecture could be adapted to video polyp segmentation or other lumen and tissue segmentation tasks that require boundary precision.","The low parameter count and short training time make repeated retraining on hospital-specific data practical."],"supporting_citations":[{"why":"PraNet is a primary attention-based baseline and also provides the train/test split protocol the paper follows.","marker":"[42]"},{"why":"ColonFormer is a transformer baseline and a second source of the partitioning guidelines used in the experiments.","marker":"[48]"},{"why":"FCB-SwinV2 is the strongest baseline and the source of the paper's warning about video-sequence data leakage in CVC-ClinicDB.","marker":"[52]"},{"why":"PVT-CASCADE is a transformer-CNN baseline that the paper must beat on every dataset.","marker":"[56]"},{"why":"U-Net is the foundational encoder-decoder baseline against which the largest gains are reported.","marker":"[65]"},{"why":"HarDNet-MSEG is a compact CNN baseline representing the multi-branch feature-interaction approach MMCC-Net extends.","marker":"[29]"},{"why":"Polyp-PVT is a hybrid CNN-transformer baseline used for comparison on all six datasets.","marker":"[49]"},{"why":"FCB-Former is a transformer-CNN fusion baseline compared directly on Kvasir and CVC-ClinicDB.","marker":"[47]"},{"why":"CVC-ClinicDB supplies the primary training and test images for the colonoscopy-video-based comparisons.","marker":"[61]"},{"why":"Kvasir supplies the other primary training and test set used in the two main experiments.","marker":"[62]"}],"fun_headline_variants":["Multi-path CNN beats eight SOTA polyp models","Efficient cascaded CNN tops polyp segmentation benchmarks","Small convolutional net beats transformers on polyp segmentation","MMCC-Net: a 1.4M-parameter polyp segmentation win"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's results stand on the assumption that the differences between MMCC-Net and the strongest baselines are real and not artifacts of random training variation, split choice, or data leakage between video frames; the reported confidence intervals overlap for some key metrics and no paired significance tests are shown.","fun_headline_variants_meta":{"raw":{"variants":["Multi-path CNN beats eight SOTA polyp models","Efficient cascaded CNN tops polyp segmentation benchmarks","Small convolutional net beats transformers on polyp segmentation","MMCC-Net: a 1.4M-parameter polyp segmentation win"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000796,"raw_usage":{"total_tokens":3507,"prompt_tokens":954,"completion_tokens":2553,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":2488}},"tokens_in":570,"tokens_out":2553,"duration_ms":18665,"temperature":1.0,"reasoning_tokens":2488,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:26:12.343678+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a paired per-image significance test, such as a Wilcoxon signed-rank or bootstrap test on per-image Dice, between MMCC-Net and FCB-SwinV2 and PVT-CASCADE on a CVC-ClinicDB split that keeps frames from the same colonoscopy video entirely in either training or testing; if the difference is not significant at p < 0.05, the paper's central outperformance claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PraNet is a primary attention-based baseline and also provides the train/test split protocol the paper follows."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ColonFormer is a transformer baseline and a second source of the partitioning guidelines used in the experiments."},{"cited_title":"Oanh, N.T","cited_arxiv_id":null,"evidence_quote":"FCB-SwinV2 is the strongest baseline and the source of the paper's warning about video-sequence data leakage in CVC-ClinicDB."},{"cited_title":"Sánchez, G","cited_arxiv_id":null,"evidence_quote":"U-Net is the foundational encoder-decoder baseline against which the largest gains are reported."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"HarDNet-MSEG is a compact CNN baseline representing the multi-branch feature-interaction approach MMCC-Net extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Polyp-PVT is a hybrid CNN-transformer baseline used for comparison on all six datasets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FCB-Former is a transformer-CNN fusion baseline compared directly on Kvasir and CVC-ClinicDB."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CVC-ClinicDB supplies the primary training and test images for the colonoscopy-video-based comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Kvasir supplies the other primary training and test set used in the two main experiments."}],"review_version":1}