{"id":"695139c0-0bc0-4753-be88-8ba1b53cfd72","arxiv_id":"2412.08345","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ConDSeg, a contrast-driven medical image segmentation framework, reports higher IoU and Dice scores than prior methods on five public datasets across three imaging modalities.","lead":"This paper presents ConDSeg, a two-stage deep learning framework for medical image segmentation that combines contrastive training, foreground-background feature decoupling, and size-specific decoders. It reports higher benchmark scores than prior methods on five public medical imaging datasets spanning endoscopy, dermatoscopy, and histology.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is unverified without a comparison to the authors' own decoupling-based EPPS and without statistical significance testing.","rationale":"The reader's weakest assumption concerned the Size-Aware Decoder's implicit size-assignment, which is a serious mechanistic gap: without per-size supervision, the three decoders are not demonstrably size-aware, so the claimed avoidance of co-occurrence learning is unsupported. However, that concern weakens the explanation of why ConDSeg works, not the empirical claim that it works. The stronger threat to the paper's central SOTA claim is that the comparison itself is incomplete and non-statistical: the authors cite EPPS, their own decoupling-based polyp segmentation method, but do not compare against it, and no run-to-run variability is reported. If EPPS performs comparably or better, the headline claim fails; if the margins are within noise, the claim is unsupported. The released code makes this directly testable. Since the reader's rationale already notes the omitted EPPS comparison and missing error bars, there is partial agreement, though the reader's stated weakest assumption is a different concern. A conditional verdict remains appropriate: the empirical claim is plausible but must be verified against the missing baseline and with uncertainty quantification before acceptance.","tokens_in":15856,"tokens_out":8129,"duration_ms":89037,"concrete_test":"Using the released ConDSeg code, train ConDSeg and EPPS (official code or independent reimplementation) on the exact splits stated in the paper (Kvasir-SEG 880/120, Kvasir-Sessile 156/20/20, GlaS 85/80, ISIC-2016 900/379, ISIC-2017 2000/150/600) with the same backbone, preprocessing, and at least 5 random seeds. Report mean +/- std mIoU/mDSC with paired significance tests, and verify at least one reported table entry by reproduction. If ConDSeg is not significantly better than EPPS on the polyp datasets, or if the margins over DTAN/EIU-Net collapse under repeated runs, the 'best segmentation performance' claim should be softened to 'competitive' or qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: ConDSeg achieves the best segmentation performance on five datasets. For that claim to hold, the comparison set must be complete and the reported margins reliable. The least secure condition is the completeness and reliability of Tables 1-2. First, the paper cites the authors' own prior work, EPPS (Lei and Wang 2024), which combines edge information with selective feature decoupling and is conceptually the closest predecessor of ConDSeg's SID module, yet EPPS is absent from all comparisons. If EPPS matches or exceeds ConDSeg on Kvasir-SEG or Kvasir-Sessile, the 'best segmentation performance' statement is false. Second, every reported number is a single run with no error bars, random seeds, or significance tests, so the 1-5 point gains over DTAN and EIU-Net cannot be distinguished from training noise. The qualitative co-occurrence evidence (Figure 3) is a single Grad-CAM example and does not quantify false-positive polyps in single-occurrence cases. These gaps do not disprove the method, but they leave the headline claim unsubstantiated as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents ConDSeg, a two-stage medical image segmentation framework. In the first stage, a Consistency Reinforcement (CR) strategy trains the encoder on original and strongly augmented images with a consistency loss. In the second stage, a Semantic Information Decoupling (SID) module splits encoder features into foreground, background, and uncertainty maps; a Contrast-Driven Feature Aggregation (CDFA) module uses the foreground and background maps to guide multi-level feature fusion; and a Size-Aware Decoder (SA-Decoder) assigns small, medium, and large objects to three decoders at different feature levels. The authors report state-of-the-art mIoU, mDSC, Recall, and Precision on Kvasir-SEG, Kvasir-Sessile, GlaS, ISIC-2016, and ISIC-2017, with ablations, convergence curves, and supplementary analyses of thresholds, window sizes, inference speed, backbone choices, and a Synapse multi-class extension.","tokens_in":16128,"tokens_out":4794,"duration_ms":47112,"significance":"If the empirical claim is substantiated, ConDSeg would be a broadly applicable, real-time-capable segmentation framework with a released implementation and a generally sound ablation design. The supplementary sensitivity analyses for threshold t, window size K, stage-1 epochs, and encoder choice are useful, and the inference speed measurement is a practical strength. The main significance, however, is incremental: the novelty lies in combining known ideas (contrastive consistency, decoupling, multi-scale decoders) rather than in a new theoretical principle. The headline 'best on all five datasets' is currently under-supported because the comparison omits the authors' own closest predecessor and no statistical significance is reported, and the co-occurrence mechanism is validated only by a single Grad-CAM example.","major_comments":[{"comment":"Tables 1 and 2 omit EPPS (Lei and Wang 2024), which the paper itself cites as a method that combines edge information with selective feature decoupling and is the closest conceptual predecessor of ConDSeg's SID module. Since the authors explicitly claim 'on all five datasets, our method achieves the best segmentation performance,' the comparison set must include this method. If EPPS matches or exceeds ConDSeg on Kvasir-SEG or Kvasir-Sessile, the headline claim is false; the manuscript should report those numbers and discuss the difference.","section":"Experiments, Comparison with Other State-of-the-Art"},{"comment":"All results in Tables 1 and 2 are single values with no error bars, no multiple seeds, and no significance tests. The claimed advantages over the runner-up are as small as 0.5 mIoU on Kvasir-SEG (DTAN: 84.1 vs. 84.6) and 1.3 mIoU on ISIC-2016 (EIU-Net: 85.5 vs. 86.8), which is within typical run-to-run variability for deep segmentation models. The authors should report mean and standard deviation over at least three independent runs or perform paired significance tests before the 'best' claim can be accepted.","section":"Experiments, Comparison with Other State-of-the-Art"},{"comment":"The claim that the Size-Aware Decoder 'avoids erroneous learning of co-occurrence features' rests on the premise that shallow feature maps are suitable for small entities and deep maps for large entities. However, the three decoders are assigned small, medium, and large roles purely by feature level, with no size labels, per-size supervision, or object grouping. The only direct evidence for the co-occurrence benefit is a single Grad-CAM example (Figure 3). To support the mechanism, the authors should quantify false-positive predictions on images containing a single polyp (e.g., precision on a single-occurrence subset) and show that SA-Decoder specifically reduces such errors, rather than only improving aggregate IoU.","section":"Methodology, Size-Aware Decoder; Overcoming Co-occurrence"}],"minor_comments":[{"comment":"The section heading 'Semantic Imformation Decoupling' contains a typo and should read 'Information'.","section":"Methodology, Semantic Information Decoupling"},{"comment":"In the text describing the auxiliary head, 'regions of uncertainty (M_fg)' should be 'M_uc'; the current text uses the same symbol as the foreground mask, which is confusing.","section":"Methodology, Semantic Information Decoupling"},{"comment":"The ablation section says 'we set up 3 groups of ablation experiments on the GlaS dataset' but then lists 'Our 4 experiment setups are: 1) Baseline. 2)... 3)...' The count is inconsistent and should be corrected.","section":"Ablation Study, Proposed Modules"},{"comment":"The reference Xi et al. 2024, 'Evaluating travel behavior resilience across urban and rural areas during the COVID-19 pandemic,' appears unrelated to medical image segmentation and should be replaced with a relevant citation or removed.","section":"References"},{"comment":"In Table 3, the checkmark column header row is ambiguous about which columns correspond to 'Net0', 'CR', and 'Stage 2'; please label the columns explicitly so the five configurations are easy to interpret.","section":"Ablation Study, Training Strategy"},{"comment":"The abstract contains a grammatical error: 'It accurately locate entities' should be 'It accurately locates entities.'","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The reference list contains an entry (Xi et al. 2024) about travel behavior that has no connection to medical imaging, which suggests the bibliography was not carefully checked. I recommend asking the authors to verify all references and to explicitly justify the omission of EPPS from the comparison tables, since it is the closest related work and is cited in the same manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a serious, competently built segmentation framework with a much more thorough experimental appendix than most, but the paper's central claim — state-of-the-art on all five datasets — is not yet earned. The two biggest omissions are the authors' own EPPS baseline and any measure of variance.\n\nWhat's new: the exact combination of consistency pretraining, foreground/background/uncertainty decoupling, contrast-driven local aggregation, and parallel size-aware decoders doesn't appear in prior work. Each ingredient is a known family (consistency regularization, feature decoupling, attention fusion, multi-scale decoding) but the conjunction is original. The paper also does a lot of things well: ablations on every major component, threshold sensitivity, window size sensitivity, backbone swap, inference speed, and a 3D extension on Synapse. That is a solid empirical package.\n\nSoft spots, in order of seriousness. First, EPPS (Lei & Wang 2024) is cited as the closest decoupling-based predecessor, yet it never appears in any comparison table. Since the paper claims 'best on all five datasets', omitting your own nearest prior method is a gap the reader can't fill. Second, all numbers are single-run; there are no error bars or significance tests. The Kvasir-SEG margin over DTAN is 0.5 mIoU, which is within training noise. The larger margins on GlaS and ISIC-2017 are more compelling, but with one run per method you can't tell whether the 1–2 point differences are real. Third, the co-occurrence claim rests on a single Grad-CAM example. The SA-Decoder's scale-to-layer assignment is plausible and conventional, but the paper provides no quantitative evidence that it actually suppresses false positives on single-occurrence cases. Minor issues: an unrelated citation (Xi et al. 2024 is about travel resilience, not deep learning) and a typo in a section heading. Also, the uncertainty-reduction framing partly over-reads Eq. 10: that loss enforces mutual exclusion among the three masks by construction, so the observed separation is partly a definitional consequence, not independent evidence of learned uncertainty reduction.\n\nThe math and the architecture hold together: losses are consistent with the stated goals, ablation increments are monotone, and the framework is not a circular derivation. But the external comparison is incomplete and the internal evidence is under-powered for the headline claim. With EPPS added to the tables, repeated runs (or at least a couple of seeds), and a quantitative false-positive analysis on single-occurrence images, this would be a strong paper.\n\nFor a reviewer: yes, send it. It deserves a serious referee, but the referee should insist on those comparisons before acceptance.","headline":"A competent, well-ablated segmentation framework whose 'SOTA on all five datasets' claim is undercut by a missing EPPS baseline and the absence of any error bars.","tokens_in":16583,"tokens_out":4085,"would_cite":false,"duration_ms":39622,"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":"ConDSeg claims that a single framework, built on contrast-driven feature enhancement and size-aware decoding, outperforms every compared method on five medical image segmentation datasets.","keywords":["medical image segmentation","contrastive learning","feature decoupling","co-occurrence","size-aware decoder","polyp segmentation","skin lesion segmentation","gland segmentation"],"falsifier":"Swap the feature-level inputs among the small, medium, and large decoders and retrain: if the size-specialization story is doing the work, performance should drop, while if mean IoU and mean Dice stay roughly the same, the Size-Aware Decoder's stated mechanism is not the source of the gains. A complementary check is to count, per ground-truth object size, which decoder's output dominates the final mask.","tokens_in":15688,"feed_emoji":"🩺","tokens_out":14470,"duration_ms":127071,"temperature":0.7,"pith_summary":"The paper sets out to show that the two recurring failure modes of medical image segmentation—soft, low-contrast boundaries and misleading co-occurrence patterns (statistical associations such as 'polyps often appear in multiples,' which fool the model when an entity appears alone)—can be handled by one general framework rather than by task-specific architectures. ConDSeg trains its encoder first to predict the same mask from an image and from a strongly perturbed copy of it, then decouples the deepest features into foreground, background, and uncertainty maps, using the first two as contrastive guides for multi-level feature fusion. A Size-Aware Decoder splits the final prediction among small, medium, and large branches, on the premise that shallow features carry the detail needed for small entities and deep features the semantics needed for large ones. On five public datasets spanning endoscopy, histology, and dermoscopy, the authors report that ConDSeg beats every compared method in mean IoU and mean Dice. If the claim holds, the practical payoff is a single training recipe that lifts segmentation accuracy across modalities while also reducing the model's tendency to hallucinate co-occurring objects.","feed_headline":"ConDSeg outperforms 13 rival networks on five medical datasets","feed_subtitle":"ConDSeg lifts mIoU and Dice on polyp, gland, and skin-lesion benchmarks.","key_machinery":"The load-bearing machinery is a two-stage pipeline made of four named components. Consistency Reinforcement (CR) isolates the ResNet-50 encoder with a simple prediction head and trains it so that the predicted mask from an original image matches the predicted mask from a strongly augmented copy (brightness, contrast, saturation, hue, grayscale, Gaussian blur); the matching loss alternately binarizes one prediction and applies binary cross-entropy to the other, which the paper argues is simpler and numerically stabler than KL or JS divergence. Semantic Information Decoupling (SID) splits the deepest feature map into foreground, background, and uncertainty channels, and penalizes both per-channel errors and pairwise overlap, so the uncertainty channel shrinks as training proceeds. Contrast-Driven Feature Aggregation (CDFA) uses the foreground and background maps to compute two softmax attention matrices over a $3\\times 3$ window and applies them to the fused feature values, making foreground/background contrast the guide for multi-level feature fusion. The Size-Aware Decoder (SA-Decoder) runs three parallel decoders—small, medium, and large—that receive features from adjacent CDFA outputs, then fuses their predictions, on the premise that shallow levels should handle small entities and deep levels large ones.","core_discovery":"ConDSeg's central claim is that the same mechanisms can solve both ambiguous-boundary and co-occurrence problems, and that this generality pays off empirically: the authors report the best mean IoU and mean Dice on Kvasir-SEG, Kvasir-Sessile, GlaS, ISIC-2016, and ISIC-2017. The decisive design step is to make foreground-versus-background contrast an explicit feature, not just a training signal: the Semantic Information Decoupling module converts the encoder's deepest map into separate foreground, background, and uncertainty channels, and the Contrast-Driven Feature Aggregation module turns those channels into attention weights that steer multi-level feature fusion. The Consistency Reinforcement stage hardens the encoder against illumination and contrast shifts before the full network is fine-tuned. The Size-Aware Decoder then assigns small, medium, and large objects to separate branches fed by different feature levels, which the paper argues prevents the network from learning spurious size correlations such as one polyp implying another nearby. Together these components are claimed to produce a network that segments accurately even when an entity appears alone or under poor imaging conditions.","pith_inferences":["The paper leaves untested whether the shallow-to-large mapping in the Size-Aware Decoder actually holds: a probe that records which decoder fires on ground-truth objects of each size would show whether size specialization is real or merely an implicit bias.","Because CDFA uses foreground and background maps as attention generators rather than as extra fused features, it could be inserted into other encoder-decoder backbones as a contrastive fusion layer; the paper does not report such transfer experiments.","The complementarity loss that drives the uncertainty channel toward zero resembles a general strategy for ambiguous-boundary labeling, so one could test it on pseudo-label refinement or open-set segmentation where an explicit unknown region is useful.","The threshold used in the consistency loss is fixed at 0.5, and the supplementary sweep shows sensitivity near the extremes, so a learned or per-dataset threshold might push the reported numbers further, though the paper does not explore this."],"forward_implications":["If ConDSeg's reported numbers are correct, a single architecture can replace method families specialized to polyp, gland, and skin-lesion segmentation, because the contrast, uncertainty, and size mechanisms are not modality-specific.","The Consistency Reinforcement stage offers a cheap robustness gain: it trains only the encoder with a simple head, and the ablations show that the two-stage scheme plus CR beats one-stage training of the full network.","Because the Size-Aware Decoder deliberately keeps small and large entities in separate branches, models trained under co-occurrence-heavy distributions should generalize better to single-occurrence cases, the failure mode visualized with TGANet.","At roughly 59.6 frames per second on an RTX 4090, the complete ConDSeg pipeline is fast enough for real-time use, so the accuracy gains would not come at the cost of deployability."],"supporting_citations":[{"why":"Supplies the Kvasir-SEG polyp dataset, the primary benchmark whose best mean IoU and mean Dice ConDSeg claims.","marker":"(Jha et al. 2020)"},{"why":"Supplies the Kvasir-Sessile small-polyp subset used to test co-occurrence and size handling.","marker":"(Jha et al. 2021)"},{"why":"Supplies the GlaS histology gland dataset, establishing the cross-modality claim.","marker":"(Sirinukunwattana et al. 2017)"},{"why":"Supplies the ISIC-2016 dermoscopy skin-lesion dataset used in the comparison.","marker":"(Gutman et al. 2016)"},{"why":"Supplies the ISIC-2017 dermoscopy skin-lesion dataset used in the comparison.","marker":"(Codella et al. 2018)"},{"why":"U-Net is the baseline architecture and comparison method that ConDSeg builds from and must beat.","marker":"(Ronneberger, Fischer, and Brox 2015)"},{"why":"PraNet is a polyp-segmentation baseline whose parallel reverse-attention design is compared against.","marker":"(Fan et al. 2020)"},{"why":"TGANet is the text-guided baseline used in the Grad-CAM co-occurrence comparison.","marker":"(Tomar et al. 2022)"},{"why":"XBoundFormer is the transformer boundary-modeling baseline in the comparison tables.","marker":"(Wang et al. 2023)"},{"why":"DTAN is the diffusion-based baseline in the comparison tables that ConDSeg claims to surpass.","marker":"(Zhao et al. 2024)"}],"fun_headline_variants":["ConDSeg's contrast-driven features beat 13 rivals on 5 medical sets","Contrast-driven framework ConDSeg tops five medical segmentation datasets","ConDSeg beats 13 rivals by making contrast an explicit feature","Explicit contrast helps ConDSeg outperform 13 rivals on 5 medical datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that shallow feature maps suit small entities and deeper maps suit large ones: the three decoders are assigned those roles by feature level alone, with no size labels, object grouping, or per-size supervision to confirm the mapping.","fun_headline_variants_meta":{"raw":{"variants":["ConDSeg's contrast-driven features beat 13 rivals on 5 medical sets","Contrast-driven framework ConDSeg tops five medical segmentation datasets","ConDSeg beats 13 rivals by making contrast an explicit feature","Explicit contrast helps ConDSeg outperform 13 rivals on 5 medical datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001219,"raw_usage":{"total_tokens":5077,"prompt_tokens":1069,"completion_tokens":4008,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":685,"completion_tokens_details":{"reasoning_tokens":3928}},"tokens_in":685,"tokens_out":4008,"duration_ms":27688,"temperature":1.0,"reasoning_tokens":3928,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:54:39.619050+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Swap the feature-level inputs among the small, medium, and large decoders and retrain: if the size-specialization story is doing the work, performance should drop, while if mean IoU and mean Dice stay roughly the same, the Size-Aware Decoder's stated mechanism is not the source of the gains. A complementary check is to count, per ground-truth object size, which decoder's output dominates the final mask.","supporting_citations":[{"cited_title":"H.; Riegler, M","cited_arxiv_id":null,"evidence_quote":"Supplies the Kvasir-SEG polyp dataset, the primary benchmark whose best mean IoU and mean Dice ConDSeg claims."},{"cited_title":"H.; Johansen, D.; De Lange, T.; Johansen, H","cited_arxiv_id":null,"evidence_quote":"Supplies the Kvasir-Sessile small-polyp subset used to test co-occurrence and size handling."},{"cited_title":"P.; Chen, H.; Qi, X.; Heng, P.-A.; Guo, Y","cited_arxiv_id":null,"evidence_quote":"Supplies the GlaS histology gland dataset, establishing the cross-modality claim."},{"cited_title":"C.; Gutman, D.; Celebi, M","cited_arxiv_id":null,"evidence_quote":"Supplies the ISIC-2017 dermoscopy skin-lesion dataset used in the comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PraNet is a polyp-segmentation baseline whose parallel reverse-attention design is compared against."},{"cited_title":"K.; Jha, D.; Bagci, U.; and Ali, S","cited_arxiv_id":null,"evidence_quote":"TGANet is the text-guided baseline used in the Grad-CAM co-occurrence comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"XBoundFormer is the transformer boundary-modeling baseline in the comparison tables."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DTAN is the diffusion-based baseline in the comparison tables that ConDSeg claims to surpass."}],"review_version":1}