{"id":"793fa214-509e-40a8-a68e-01ba506214db","arxiv_id":"2504.13597","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"FocusNet, a transformer-based network with local and pooling attention, achieves top segmentation dice scores on four of five modalities in PolypDB, but is slightly worse than PVT-CASCADE on white-light images.","lead":"FocusNet adds three attention-based modules to a standard polyp segmentation network and reports top dice scores on four of five colonoscopy imaging modalities in the new PolypDB dataset. The paper is one of the first to test polyp segmentation across multiple imaging modalities and clinical centers, which matters for real-world colonoscopy screening tools.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-run metrics on 60-146 image test sets cannot support the abstract's 'consistently outperforms' claim, especially since Table I shows FocusNet numerically loses on WLI (0.41% mDSC) and center-wise on BKAI/Simula.","rationale":"The reader's weakest_assumption correctly identifies the most load-bearing concern: single-run metrics on small test sets without confidence intervals. My pass confirms this is central because the reported margins over the second-best method are small (e.g., 0.88% mDSC over RMAMamba-T on FICE; 0.43% on NBI; 0.99% on BLI), and the paper's own text concedes WLI and center-wise deficits that contradict the abstract's overclaim. I found no internal mathematical inconsistency in the proposed modules, and the dataset contribution is real, but the statistical evidence is insufficient to verify the headline claim. A conditional verdict is appropriate: the architecture is plausible, but the claim of consistent SOTA performance requires repeated-seed evaluation with confidence intervals. I do not escalate to REJECT because the design is coherent, the dataset is a genuine contribution, and the failure cases are openly discussed; the weakness is in the evidence base, not the architecture itself.","tokens_in":13710,"tokens_out":2180,"duration_ms":16757,"concrete_test":"Run FocusNet, PVT-CASCADE, and RMAMamba-T for at least 5 random seeds with identical training protocols on PolypDB, and report mean ± std of mDSC/mIoU on each modality-wise test set, especially LCI (n=60) and WLI (n=360). If the 95% confidence intervals for WLI overlap between FocusNet and PVT-CASCADE, then the 'consistently outperforms' claim must be revised, and the unspecified p-values are called into question.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is that FocusNet 'consistently outperforms existing state-of-the-art approaches' across five modalities, yet Table I shows FocusNet is numerically inferior on WLI: mDSC 0.9342 vs PVT-CASCADE 0.9383. The abstract and Section III.C acknowledge this 'slightly inferior' performance, directly undercutting 'consistently outperforms'. More damaging, the entire quantitative comparison rests on single training runs with early stopping (Section III.B) evaluated on very small modality-wise test sets: BLI 70, FICE 70, LCI 60, NBI 146 images. Tables I and II list p-values but never specify the statistical test, the number of runs, or variance structure; no confidence intervals or standard deviations are reported. On a 60-image test set, a few difficult images can shift mean Dice by several points, and the smallest reported gaps are 0.41% (WLI mDSC) and 0.11% (BKAI mDSC). The ablation (Table IV) is likewise single-run on one modality, so the claimed contribution of DEM (+1.06% mDSC over setting #5) is within plausible seed noise. If seed-to-seed variation is comparable to the reported margins, the central claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FocusNet, a Transformer-enhanced polyp segmentation network built on a PVTv2 backbone, with three new modules: a Cross-semantic Interaction Decoder Module (CIDM) in multiplicative and additive variants, a Detail Enhancement Module (DEM) using multi-branch deformable convolutions on shallow features, and a Focus Attention Module (FAM) combining local and pooling attention. The method is evaluated on PolypDB, a multi-center, multi-modality colonoscopy dataset (BLI, FICE, LCI, NBI, WLI), in both modality-wise and center-wise settings, and is compared against seven CNN-, Transformer-, and Mamba-based baselines. The paper reports top mDSC values on four modality-wise test splits (BLI 82.47%, FICE 88.46%, LCI 92.04%, NBI 82.09%) and competitive results on WLI (93.42%), along with an ablation study, computational complexity comparison, qualitative examples, and failure-case analysis. The central claim is that FocusNet consistently outperforms prior state-of-the-art methods across all five modalities and demonstrates strong generalization across centers.","tokens_in":14151,"tokens_out":3595,"duration_ms":33024,"significance":"If the reported results are robust, FocusNet would be a useful contribution to multi-modality polyp segmentation, a clinically relevant direction given the limited modality coverage of most existing datasets and models. The paper's strengths include the use of a newly assembled multi-center, multi-modality dataset, evaluation under both modality-wise and center-wise protocols, inclusion of efficiency metrics, qualitative results, failure-case discussion, and a public code link. However, the significance of the paper's central claim is weakened by the absence of repeated-run statistics and by the fact that the reported numbers themselves contradict the claimed 'consistent' superiority on several settings. The architectural ideas, especially the combination of local and pooling attention in FAM, are reasonable and potentially impactful, but the current evidence does not fully support the strength of the stated conclusions.","major_comments":[{"comment":"The claim that FocusNet 'consistently outperforms existing state-of-the-art approaches' is not supported by the reported numbers. On WLI, FocusNet has mDSC 0.9342 versus PVT-CASCADE 0.9383 and mIoU 0.8862 versus 0.8927; on BKAI, mDSC is 0.9127 versus PVT-CASCADE 0.9138; on Simula, mDSC is 0.9230 versus PVT-CASCADE 0.9306. Section III.C itself acknowledges that FocusNet is 'slightly inferior' on WLI and on BKAI/Simula. The abstract, contributions list, and conclusion should be revised to state that FocusNet outperforms the compared methods on BLI, FICE, LCI, and NBI, and is competitive on WLI and across centers.","section":"Abstract, Section III.C, Tables I and II"},{"comment":"The manuscript lists P-values in Tables I and II but never specifies the statistical test, the number of runs, or the variance structure; all metrics appear to come from a single training run with early stopping. Test sets are small (BLI 70, FICE 70, LCI 60, NBI 146, Karolinska 30), and the smallest reported gaps are 0.41% mDSC on WLI and 0.11% mDSC on BKAI. To support the claim of statistical significance made in Section IV, the authors should report mean and standard deviation over multiple seeds, provide confidence intervals or bootstrap estimates, and name the test used.","section":"Section III.A and Tables I, II, IV"},{"comment":"The ablation study is performed on a single modality (BLI) with single-run metrics. The DEM contribution is +1.06% mDSC over Setting #5, while Setting #5 itself is 0.33% mDSC below Setting #3, so the incremental benefit of adding DEM and the second CIDM variant is not clearly established. Without repeated runs or an additional validation modality, the claim that each module contributes to the final result is not supported by the present evidence.","section":"Section III.D, Table IV"}],"minor_comments":[{"comment":"Equation (10) has an unbalanced parenthesis: the right-hand side should be P + Conv_{1x1}(Conv_{3x3}(Conv_{3x3}(O_f))) or the intended structure should be made explicit.","section":"II.D, Eq. (10)"},{"comment":"The caption cites reference [34] (Tajbakhsh et al.) for the PolypDB dataset, but the dataset is introduced in reference [27]; the citation should be corrected.","section":"Table II caption"},{"comment":"In the paragraph describing NBI results, the text says 'FocusNet outperforms G-CASCADE by approximately 5.62%, 5.58%, and 4.64%' but no method named G-CASCADE appears elsewhere; this appears to be a typo for PVT-CASCADE.","section":"III.C.1"},{"comment":"The phrase 'Eficient Channel Attention' should be 'Efficient Channel Attention'.","section":"II.C, after Eq. (5)"},{"comment":"In the center-wise description, 'for it’s common in all three centers' should be revised to 'since it is common to all three centers'.","section":"III.A"}],"recommendation":"major_revision","confidential_remarks":"PolypDB appears to be a dataset introduced by the same research group; reference [27] includes several co-authors of the present manuscript. The paper describes the dataset as 'newly introduced' without stating this connection explicitly. This is a transparency issue rather than a correctness issue, but the authors should make the relationship clear and ensure that their use of their own dataset and their own method RMAMamba-T as a baseline is disclosed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi — quick read of FocusNet. Short version: the PolypDB multi-modality evaluation is the most useful thing here, and FocusNet is a reasonable engineering contribution. But the abstract overstates the results, and the statistics as reported aren't enough to support the headline.\n\nWhat's actually new: first multi-modality, multi-center polyp segmentation benchmark on PolypDB, with five modalities and three centers, and a public code release. The architecture combines PVTv2 with a detail-enhancement module using deformable convolutions and a local-plus-pooling attention module. That is incremental—FAM is close to TransNeXt's foveal attention—but the combination is new and it performs well on four of five modalities.\n\nCredit where due: the experimental setup is broad, the comparison includes CNN, transformer, and Mamba baselines, and the failure-case discussion is honest. They also acknowledge in Section III.C that they are slightly behind PVT-CASCADE on WLI, which is more than some papers do.\n\nThe problems. First, the abstract and conclusion say 'consistently outperforms,' but Table I shows FocusNet loses to PVT-CASCADE on WLI by 0.41% mDSC, and Table II shows losses on BKAI and Simula centers. So the central claim is not consistent. That should be fixed. Second, all results are single runs, early stopping, and test sets as small as 60 images. The p-values in the tables don't state which test, how many runs, or any variance. A 0.41% gap on a 60-image set is easily seed noise. The ablation is single-run on BLI, so the claimed DEM gain of 1.06% mDSC is also within plausible noise. Third, there is a self-comparison with RMAMamba-T, a method from their own group, which is fine but should be disclosed more clearly.\n\nNone of this makes the paper worthless. The benchmark is useful, the code is public, and the model is competitive. But the claims need to be scaled back and the statistics need to be redone with multiple seeds and specified tests.\n\nFor peer review: yes, send it. A serious referee can get this into good shape, but as it stands the headline is not supported.","headline":"Useful multi-modality benchmark and a decent architecture, but the 'consistently outperforms' claim is contradicted by the paper's own Table I, and the single-run statistics can't support the marginal gains.","tokens_in":14524,"tokens_out":1937,"would_cite":true,"duration_ms":17042,"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":"FocusNet, a Transformer-attention network, segments colon polyps across five imaging modalities and outperforms prior methods on four of them.","keywords":["polyp segmentation","colonoscopy","multi-modality imaging","transformer","attention mechanism","FocusNet","PolypDB","medical image segmentation"],"falsifier":"Retrain FocusNet and the strongest baselines several times with different random seeds on the same PolypDB splits; if the seed-to-seed spread in dice exceeds 0.59 percentage points, the smallest reported winning margin, the claim that FocusNet consistently outperforms is not established.","tokens_in":13559,"feed_emoji":"🩺","tokens_out":8816,"duration_ms":74568,"temperature":0.7,"pith_summary":"The paper tries to establish that one network, FocusNet, can segment colorectal polyps accurately across the five endoscopic imaging modalities used in colonoscopy—white light (WLI), blue laser (BLI), flexible spectral imaging (FICE), linked color (LCI), and narrow band (NBI)—and across data from multiple clinical centers. If that holds, a single model could serve in real screening environments instead of being retrained for each scope setting. FocusNet builds a pyramid vision transformer encoder with three added modules: two cross-semantic decoders that fuse deep features into coarse maps, a detail-enhancement module that sharpens shallow edges, and a focus attention module that mixes local and pooled global context. On the PolypDB benchmark the paper reports dice scores of 82.47% (BLI), 88.46% (FICE), 92.04% (LCI), 82.09% (NBI), and 93.42% (WLI), with FocusNet leading on all modality-wise tests except WLI, where it trails PVT-CASCADE by 0.41 points.","feed_headline":"FocusNet tops four of five colonoscopy modalities in polyp tests","feed_subtitle":"The model posts 82–93 percent dice across five imaging modes, easing the path to real-world colonoscopy AI.","key_machinery":"The load-bearing mechanism is the Focus Attention Module (FAM), a dual-path attention block that balances local detail with global context: it builds a shared query from the detail-enhanced shallow features, derives separate key/value pairs from local windows and from adaptive average pooling, concatenates the two similarity scores before a single softmax, and adds the local and pooled attention outputs before multiplying with a channel-attention-refined feature. Two Cross-semantic Interaction Decoder Modules (CIDM-M and CIDM-A) supply the coarse segmentation maps and fused features by combining deep encoder levels with element-wise multiplication or addition plus channel/spatial attention. The Detail Enhancement Module (DEM) processes shallow features through four parallel branches of 1×3 and 3×1 deformable convolutions followed by efficient channel attention. The network's final prediction is the element-wise sum of four segmentation maps, two from the decoders and two from the shared FAM.","core_discovery":"FocusNet's central claim is that combining coarse semantic maps with boundary-focused shallow features, then reweighting them through a focus attention module, yields a polyp segmenter that generalizes across modalities and centers better than existing CNN-, transformer-, and mamba-based models. The architecture uses a pyramid vision transformer backbone; two cross-semantic interaction decoder variants (multiplicative CIDM-M and additive CIDM-A) assemble deep feature maps into coarse predictions, while the detail enhancement module passes the first-stage features through four branches of deformable convolutions and channel attention. The focus attention module then computes a shared query and separate local and pooling keys/values, concatenates the two similarity scores before softmax, and combines the attended values with channel-attended input. The paper reports this design produces the highest dice coefficient on BLI, FICE, LCI, and NBI in the modality-wise protocol, and on WLI it is competitive but slightly below PVT-CASCADE; the same pattern appears in center-wise tests, which the paper attributes to pooling attention sacrificing some fine detail.","pith_inferences":["A further step, not in the paper: if the multi-modality result replicates, PolypDB-style multi-modality benchmarks should become the default for polyp segmentation, because single-modality WLI leaderboards cannot detect whether a model will survive a modality shift in the clinic.","The FAM's local and pooling window sizes are set to 3 and the paper says they are adjustable; a natural next experiment is to sweep window sizes and pooling strides to see whether the small WLI and center-wise gap to PVT-CASCADE closes or widens.","A fair reading of the reported p-values against PVT-CASCADE on WLI and the Simula center (0.34 and 0.21) is that those differences may be within noise; a multi-seed evaluation with confidence intervals would settle whether FocusNet truly matches or trails on those splits."],"forward_implications":["A single FocusNet model trained on WLI can be applied to BLI, FICE, LCI, and NBI images without modality-specific fine-tuning, the precondition for realistic multi-modality colonoscopy deployment.","On the modality-wise benchmark, FocusNet's largest gains come on the hardest, smallest test sets (BLI and NBI), where it beats the prior best by 4.2 and 5.6 percentage points of dice.","The WLI and center-wise results show FocusNet does not dominate every setting: PVT-CASCADE retains the top WLI dice by 0.41 points and leads on all three center-wise tests, so the paper's 'consistently outperforms' phrasing holds only for four of five modality tests.","Ablations on BLI show the three modules are additive: the full model reaches 0.8247 dice versus 0.8170 for the backbone plus CIDM-M alone, so each component earns its place."],"supporting_citations":[{"why":"Supplies the pyramid vision transformer backbone; FocusNet adopts its multi-scale features f1-f4 as the basis for all downstream modules.","marker":"[22]"},{"why":"Introduces PolypDB, the multi-modality, multi-center dataset whose modality-wise and center-wise splits carry the reported results.","marker":"[27]"},{"why":"PVT-CASCADE is the strongest comparative baseline; it posts the top WLI and center-wise scores and defines the gap FocusNet must address.","marker":"[31]"},{"why":"RMAMamba-T is the strongest baseline on FICE and LCI; FocusNet's claimed gains over it support the multi-modality contribution.","marker":"[33]"},{"why":"PraNet, an attention-based CNN method, serves as the reference point for the introduction's largest per-modality gains (BLI and NBI).","marker":"[29]"},{"why":"ECA is the channel-attention component inside the Detail Enhancement Module; the ablation study credits the DEM with the final performance increment.","marker":"[25]"}],"fun_headline_variants":["FocusNet tops 4 of 5 polyp imaging modes","FocusNet wins on four of five endoscopy modes","FocusNet sets best dice on 4 colonoscopy modalities","FocusNet beats rivals on most polyp imaging modes","FocusNet: four modality wins, one close second"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported dice differences come from the models' architectures rather than from luck in a single training run, because the paper reports one run per model with early stopping and no confidence intervals.","fun_headline_variants_meta":{"raw":{"variants":["FocusNet tops 4 of 5 polyp imaging modes","FocusNet wins on four of five endoscopy modes","FocusNet sets best dice on 4 colonoscopy modalities","FocusNet beats rivals on most polyp imaging modes","FocusNet: four modality wins, one close second"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000821,"raw_usage":{"total_tokens":3624,"prompt_tokens":1004,"completion_tokens":2620,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":2557}},"tokens_in":620,"tokens_out":2620,"duration_ms":17853,"temperature":1.0,"reasoning_tokens":2557,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:03:42.925118+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain FocusNet and the strongest baselines several times with different random seeds on the same PolypDB splits; if the seed-to-seed spread in dice exceeds 0.59 percentage points, the smallest reported winning margin, the claim that FocusNet consistently outperforms is not established.","supporting_citations":[{"cited_title":"Pvt v2: Improved baselines with pyramid vision transformer,","cited_arxiv_id":null,"evidence_quote":"Supplies the pyramid vision transformer backbone; FocusNet adopts its multi-scale features f1-f4 as the basis for all downstream modules."},{"cited_title":"A Reverse Mamba Attention Network for Pathological Liver Segmentation","cited_arxiv_id":"2502.18232","evidence_quote":"RMAMamba-T is the strongest baseline on FICE and LCI; FocusNet's claimed gains over it support the multi-modality contribution."},{"cited_title":"Pranet: Parallel reverse attention network for polyp segmenta- tion,","cited_arxiv_id":null,"evidence_quote":"PraNet, an attention-based CNN method, serves as the reference point for the introduction's largest per-modality gains (BLI and NBI)."},{"cited_title":"Eca- net: Efficient channel attention for deep convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"ECA is the channel-attention component inside the Detail Enhancement Module; the ablation study credits the DEM with the final performance increment."}],"review_version":1}