{"id":"df311de0-87fc-44c0-b5a3-64954a71104e","arxiv_id":"1908.05311","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A lightweight multi-task output head that jointly learns mask, contour, and distance maps improves polyp segmentation across four standard network architectures.","lead":"This paper adds a small multi-task module to standard medical image segmentation networks, asking them to also predict contours and distance maps while segmenting. On one polyp segmentation dataset, the module improved Dice scores for all four tested architectures with little added computation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Empirical case for 'any architecture' rests on a single random split and best-of-three distance-map selection; need multi-dataset/seed evaluation.","rationale":"The stress-test pass agrees with the reader's weakest_assumption. The central claim is not that Conv-MCD works on GIANA, but that it can be added to any state-of-the-art base network and yield significant improvement; that universal claim is being inferred from one dataset, one split, and the best of three distance maps chosen after the fact. The most direct way to break the claim is to show the gains vanish under honest multi-seed, multi-dataset evaluation; the most direct way to support it is to pass that test. I did not find an internal mathematical contradiction that would invalidate the module itself. One possible side issue is that D3 is described as a signed distance transform while the regression head uses a sigmoid output, which cannot represent negative values unless the targets are rescaled; this is not stated, but it affects the auxiliary task rather than the main defensible claim. Because the concern is about evidence strength rather than a known flaw, the CONDITIONAL verdict remains appropriate; a REJECT or UNVERDICTED verdict would require evidence that the gains are actually spurious, which is not established. Verdict unchanged.","tokens_in":5849,"tokens_out":3625,"duration_ms":38273,"concrete_test":"Fix a single distance map (or pre-register all three before evaluation) and rerun Table 2 on GIANA plus an independent dataset such as ORIGA, ACDC, or DRIVE, using 5 random seeds per base architecture; report mean and std of Dice, Jaccard, and HD for base and Conv-MCD. If positive average gains persist across all architectures and datasets and exceed seed noise, the central claim is supported. If gains shrink to within noise or reverse on some architectures, the claim should be narrowed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.4 states: 'The results reported in this paper are from the best performing distance map.' This means the Conv-MCD rows in Table 2 are selected maxima over three distance-map designs, chosen after seeing results. The only in-paper evidence for the central 'any state-of-the-art base network' claim is Table 2, obtained from a single random 70/30 split of GIANA (912 images) with no cross-validation, no error bars, and no significance tests. The ORIGA generalisation check appears only as a footnote pointer to supplementary material, so the main text does not actually demonstrate transfer to a second dataset. Under these conditions, the reported gains (e.g., UNet16 Dice 0.8441 to 0.9124) cannot be separated from selection effects and split luck. The module is plausible and the within-dataset comparison against DCAN/DMTN is informative, but the headline claim of universal improvement is under-supported. This is a correctness risk, not an inconsistency in the module design.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Conv-MCD, a plug-and-play module that appends three parallel convolutional heads to a segmentation network, predicting the segmentation mask, a dilated contour map, and a distance map. The total loss combines NLL losses for mask and contour with MSE loss for the distance map. The authors claim that the module can be added to any state-of-the-art segmentation network with minimal parameter overhead and improves Dice, Jaccard, Hausdorff distance, and boundary F-score. Experiments on the GIANA polyp segmentation dataset compare four base networks (SegNet, UNet, UNet16, LinkNet34) with and without the module, and ablations compare the full module against single-auxiliary-task variants (Conv-MC, Conv-MD).","tokens_in":6088,"tokens_out":3866,"duration_ms":38631,"significance":"If the empirical claims hold, the module is a practically useful contribution: it provides a simple way to incorporate multi-task structural supervision without duplicating decoders, requires only a few thousand extra parameters, and includes code and ablations. The idea of using contour and distance maps as auxiliary tasks is not new, but the module design that decouples these tasks from the base architecture is convenient. However, the evidence for the central 'any architecture' claim is currently weak: the headline results come from a single random split with no error bars or significance tests, and the best distance-map variant was selected after seeing test results. These issues need to be addressed before the broad generalization claim is justified.","major_comments":[{"comment":"The central claim that adding Conv-MCD improves any state-of-the-art network is supported only by results on a single random 70/30 split of the GIANA dataset (912 images), with no cross-validation, no multiple random seeds, and no significance tests. The reported gains, such as UNet16 Dice rising from 0.8441 to 0.9124, could be partly due to split luck. Please add results over multiple seeds or cross-validation folds with error bars and, where appropriate, paired significance tests.","section":"Section 3.4, Table 2"},{"comment":"The paper states that the reported results are from the best performing distance map among D1, D2, and D3. Because this choice was made after seeing test-set performance, the Conv-MCD rows in Tables 1 and 2 are selected maxima, which inflates the apparent improvement over the base networks and makes the comparison unfair. Please either pre-specify the distance map, select it on a validation split, or report results for all three distance-map variants in the main text.","section":"Section 3.4, sentence beginning 'The results reported in this paper...'"},{"comment":"The NLL losses are written as positive log-likelihoods, but the total loss in Eq. (1) is minimized by gradient descent. As written, minimizing L_total would maximize the likelihood terms, which is inconsistent with standard NLL loss. Please correct the sign (introduce a negative sign) or clarify that the losses are to be maximized, and ensure the accompanying text uses the corrected formulation.","section":"Equations (2) and (3)"},{"comment":"The claim that the approach generalises beyond GIANA is supported only by a pointer to supplementary material; the main text does not contain any results on the ORIGA dataset. Since the central claim is 'any state-of-the-art base network' and generalisability across datasets, the ORIGA results should either be moved into the main text or the claim should be tempered to reflect that only single-dataset evidence is presented.","section":"Footnote 4"},{"comment":"The ablation study gives mixed evidence for the contribution of the individual auxiliary tasks: Conv-MC yields a small Dice improvement over the base (0.8149 vs 0.8125) but decreases MF (0.6083 vs 0.6144), while Conv-MD improves Dice but lowers MF compared to the base. This does not clearly support the statement that each auxiliary task individually helps, and it complicates the attribution of the full module's gains. Please discuss this and provide repeated-run statistics or at least error bars.","section":"Table 1"}],"minor_comments":[{"comment":"The values of the loss scaling factors λ1, λ2, λ3 are not reported; please state them explicitly for reproducibility.","section":"Section 2.3, Eq. (1)"},{"comment":"The contour dilation disk radius of 5 is said to be empirically optimal for 256x256 images, but no supporting experiment is shown; please clarify whether this was tuned on the training split or chosen post hoc.","section":"Section 2.2"},{"comment":"The distance map D3 is a signed distance transform and can take negative values, but the predicted distance map is passed through a sigmoid activation, which outputs values in [0,1]. Please clarify how negative distances are represented, e.g., whether the signed map is normalized or the sigmoid is used only for unsigned variants.","section":"Section 2.3, Eq. (4)"},{"comment":"The notation '1Enc 1Dec M', '1Enc 2Dec MC', and '1Enc 2Dec MD' is used in the text and Table 1 but is not defined before first use; please define these terms in the table caption or in a short notation paragraph.","section":"Section 3.4, Figures 2a and 3a"},{"comment":"The abstract says 'significant increase in performance' but the paper reports no statistical significance tests; please replace 'significant' with 'consistent' or add actual significance testing.","section":"Abstract and Section 3.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is plausible and the module design is appealing, but the empirical validation is thin for the strength of the claims. The admitted post-hoc selection of the best distance map on the same test set is a serious concern in a conference-style paper and needs to be remedied. The loss sign error in Eqs. (2)-(3) also suggests a need for careful proofreading. If the authors can add multi-seed or cross-validated results, report all distance-map variants, and move the ORIGA generalization results into the main text, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is the module design: three parallel 3x3 convolutional heads on top of a single decoder, producing mask, contour, and distance map simultaneously. That is a legitimately cheap alternative to the two-decoder designs in DCAN and DMTN, and the paper shows it on four architectures with consistent gains on the GIANA polyp dataset. The ablation comparing Conv-MC, Conv-MD, and full Conv-MCD is the right kind of experiment, and the feature-map visualization supports the single-decoder argument qualitatively. Credit is due for the clean experimental setup across architectures and for releasing code.\n\nThat said, the central claim—that the module improves \"any state-of-the-art base network\"—is under-supported. Table 2 is a single random 70/30 split with no error bars, no significance tests, and no cross-validation. Worse, Section 3.4 states the reported results come from \"the best performing distance map,\" meaning the headline numbers are selected maxima over three design choices, chosen after seeing results. That is a real correctness risk, not a nitpick. The ORIGA generalization check is relegated to a footnote and supplementary material, so the main text never actually demonstrates transfer to a second dataset. Loss weights lambda_1 through lambda_3 are also left unspecified, which makes reproducing the exact numbers harder.\n\nThe within-dataset comparison against DCAN and DMTN is useful, and the UNet16 gain is striking, but I would not put much trust in the magnitude until I see seeds and multiple datasets. The module is plausible and the overhead is tiny, so it may well be a practical addition—but the evidence as presented cannot separate real gains from split luck and selection effects.\n\nIf I were editing, I would send this to peer review rather than desk-reject it. The idea is sound, the experiments are mostly well designed, and the flaws are fixable with more rigorous evaluation. The authors need to add multi-seed runs, report standard deviations, pre-specify or at least cross-validate the distance-map choice, and put the ORIGA results in the main text. The paper is for people working on medical image segmentation who might want a cheap multi-task boost; readers outside that niche will not gain much. I would not cite it in my own work yet, but I would point to it as an example of a plausible module that needs stronger validation.","headline":"A plausible, low-cost multi-task module with honest ablations, but the headline 'any architecture' claim rests on a single split and post-hoc distance-map selection.","tokens_in":6615,"tokens_out":1485,"would_cite":false,"duration_ms":16231,"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 plug-in module that learns mask, contour, and distance together improves medical segmentation across four base networks.","keywords":["Multi-task learning","Medical image segmentation","Contour extraction","Distance map","Plug-and-play module","Convolutional neural network","Dice score","Polyp segmentation"],"falsifier":"Run the module on several independent medical segmentation datasets with a fixed, pre-registered distance-map type and multiple random splits; if the Dice gain over each base network is not consistently positive or is within split-to-split noise, the generic plug-and-play claim is falsified.","tokens_in":5700,"feed_emoji":"🩺","tokens_out":5278,"duration_ms":48417,"temperature":0.7,"pith_summary":"This paper claims that a small plug-and-play module, Conv-MCD, improves any medical image segmentation network by making it learn three related tasks at once: mask prediction, contour extraction, and distance-map estimation. The module attaches to the end of an existing encoder-decoder network and adds only a few thousand parameters, yet the authors report higher Dice and Jaccard scores and lower Hausdorff distances on polyp segmentation for UNet, SegNet, LinkNet, and UNet16. The point of the claim is that structural information already present in ground-truth masks, namely contours and distance transforms, can regularize segmentation at no extra annotation cost, replacing heavier dual-decoder multi-task designs. If the claim holds, it gives a cheap, architecture-agnostic way to tighten boundaries, suppress outliers, and reduce overfitting in medical segmentation.","feed_headline":"One plug-in module lifts medical segmentation scores","feed_subtitle":"Learning mask, contour, and distance together improves Dice and boundary metrics for little extra cost.","key_machinery":"The load-bearing object is the Conv-MCD module itself: a set of three parallel 3x3 convolutional filters placed at the end of any segmentation network, with two filters for the mask and contour classification outputs and one for the distance-map regression output. The contour target comes from dilating ground-truth boundaries by a disk of radius 5, and the distance targets come from Euclidean or signed distance transforms of the mask or contour; the paper tests three variants and reports the best. The total loss is $\\lambda_1 L_{mask} + \\lambda_2 L_{contour} + \\lambda_3 L_{distance}$ with negative log-likelihood for the two classification tasks and mean squared error for the regression task. This machinery forces the shared encoder and decoder to represent structural sub-features that a vanilla cross-entropy-trained network does not, which is what the ablations and feature-map visualizations are meant to show.","core_discovery":"The central claim is that the Conv-MCD module, three parallel convolutional filters that output a mask, a contour map, and a distance map from the same feature representation, yields better segmentation than the base network alone and than existing dual-decoder multi-task networks, while keeping parameter count and inference time close to the plain base network. On the GIANA polyp dataset, adding Conv-MCD to UNet16 raises Dice from 0.8441 to 0.9124 and Jaccard from 0.7676 to 0.8559, and similar improvements appear for SegNet, UNet, and LinkNet34. The paper attributes the gain to joint learning of a classification task for the mask, another classification task for the contour, and a regression task for the distance map, which imposes shape and smoothness priors, handles foreground-background class imbalance, and removes small outliers that a distance-only dual-decoder design tends to discard in multi-instance cases.","pith_inferences":["Beyond what the paper tests, the 'any architecture' claim would be stronger with evidence on more modalities; the benefit may be largest for blob-like objects with fuzzy boundaries similar to polyps.","The reported sensitivity to distance-map type suggests that distance transform choice could be tuned per organ or imaging modality, a cheap validation-set hyperparameter worth testing.","A natural testable extension is semi-supervised use, where predicted masks on unlabeled images generate contour and distance targets for consistency training."],"forward_implications":["Adding Conv-MCD to a segmentation network should consistently improve Dice, Jaccard, Hausdorff distance, and boundary F-score over the same network without the module, as observed for all four tested architectures.","Because the auxiliary targets are derived from ground-truth masks, the improvement comes at no additional annotation cost and with only a few thousand extra parameters.","The module should reduce outliers and improve boundary smoothness, including in multi-instance cases where a distance-map-only dual-decoder network removes small objects.","Learning mask, contour, and distance in parallel acts as a regularizer, so validation loss should plateau higher and overfitting should appear later than in base-network training."],"supporting_citations":[{"why":"Supplies the UNet baseline and the single-decoder mask-only network to which Conv-MCD is compared.","marker":"[7]"},{"why":"Provides the dual-decoder contour-aware DCAN baseline whose auxiliary contour task Conv-MCD reproduces without a second decoder.","marker":"[4]"},{"why":"Provides the dual-decoder distance-map DMTN baseline that Conv-MCD improves on in multi-instance cases.","marker":"[9]"},{"why":"Supplies the multi-task learning rationale that related tasks act as an inductive bias and regularizer.","marker":"[2]"},{"why":"Supplies the GIANA polyp segmentation dataset used for all reported experiments.","marker":"[10]"},{"why":"Supplies the SegNet architecture tested with and without the module.","marker":"[1]"},{"why":"Supplies the LinkNet34 architecture tested with and without the module.","marker":"[3]"},{"why":"Supplies the VGG16 pretrained encoder used in the UNet16 variant that shows the largest gain.","marker":"[8]"}],"fun_headline_variants":["Multi-task plug-in raises segmentation Dice across four nets","One add-on module: mask, contour, distance for better cuts","Conv-MCD: plug-in that sharpens medical image segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim of universal improvement rests on results from one randomly split polyp dataset, evaluated without cross-validation or significance tests, with the best of three distance-map variants chosen after seeing the test numbers.","fun_headline_variants_meta":{"raw":{"variants":["Multi-task plug-in raises segmentation Dice across four nets","One add-on module: mask, contour, distance for better cuts","Conv-MCD: plug-in that sharpens medical image segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000364,"raw_usage":{"total_tokens":1957,"prompt_tokens":935,"completion_tokens":1022,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":967}},"tokens_in":551,"tokens_out":1022,"duration_ms":9484,"temperature":1.0,"reasoning_tokens":967,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:16:59.119873+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the module on several independent medical segmentation datasets with a fixed, pre-registered distance-map type and multiple random splits; if the Dice gain over each base network is not consistently positive or is within split-to-split noise, the generic plug-and-play claim is falsified.","supporting_citations":[{"cited_title":"In: Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2015","cited_arxiv_id":null,"evidence_quote":"Supplies the UNet baseline and the single-decoder mask-only network to which Conv-MCD is compared."},{"cited_title":"In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)","cited_arxiv_id":null,"evidence_quote":"Provides the dual-decoder contour-aware DCAN baseline whose auxiliary contour task Conv-MCD reproduces without a second decoder."},{"cited_title":"In: 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018)","cited_arxiv_id":null,"evidence_quote":"Provides the dual-decoder distance-map DMTN baseline that Conv-MCD improves on in multi-instance cases."},{"cited_title":"In: ICML (1993)","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-task learning rationale that related tasks act as an inductive bias and regularizer."},{"cited_title":"Journal of healthcare engineering 2017 (2017)","cited_arxiv_id":null,"evidence_quote":"Supplies the GIANA polyp segmentation dataset used for all reported experiments."},{"cited_title":"IEEE Transactions on Pattern Analysis and Machine Intelligence 39, 2481--2495 (2016)","cited_arxiv_id":null,"evidence_quote":"Supplies the SegNet architecture tested with and without the module."},{"cited_title":"2017 IEEE Visual Communications and Image Processing (VCIP) pp","cited_arxiv_id":null,"evidence_quote":"Supplies the LinkNet34 architecture tested with and without the module."},{"cited_title":"In: 3rd International Conference on Learning Representations ICLR (2015)","cited_arxiv_id":null,"evidence_quote":"Supplies the VGG16 pretrained encoder used in the UNet16 variant that shows the largest gain."}],"review_version":1}