{"id":"521bc19f-cb9d-460a-85d2-a3cd565bfebc","arxiv_id":"2508.07028","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"FOCUS-Med reports state-of-the-art polyp segmentation scores by fusing graph, attention, and multi-scale fusion modules, but missing baseline details and an absent appendix undermine the claim.","lead":"This paper introduces FOCUS-Med, a deep learning model that combines graph-based modules and self-attention to outline polyps in colonoscopy images. The authors report top scores on four public benchmarks and propose using GPT-4o to grade segmentation quality, though missing baseline details and an absent appendix weaken the claim.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline comparison is unreliable: PraNet at 69.4% Dice on CVC-ClinicDB vs ~90% published, and non-standard test splits; missing training protocol makes SOTA claim unsupported.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the baseline comparison may be unfair due to undocumented training protocol and inconsistent splits. This is the single most critical issue because the paper's headline claim is empirical superiority over prior methods. If the baselines are under-trained or evaluated on non-standard splits, the reported improvements are not real. Other issues, such as the missing appendix with ablations and the LLM-based evaluation lacking human validation, are secondary: they affect completeness and interpretability but do not independently refute the central SOTA claim. However, the baseline inconsistency alone is sufficient to reject the paper in its current form, as the evidence for the central claim is not trustworthy. The reader's verdict of REJECT is therefore appropriate, and no change is needed.","tokens_in":12454,"tokens_out":3950,"duration_ms":35987,"concrete_test":"Download the official pre-trained PraNet checkpoint from the authors' public repository and run it on the exact 30-image CVC-ClinicDB test subset claimed in this paper, using the paper's evaluation script. If the Dice score is near 90% rather than the reported 69.35%, the re-implementation is not equivalent to the published PraNet. Additionally, release the full training config (epochs, loss, optimizer, input resolution, augmentations, data splits) and re-run all baselines and FOCUS-Med using the standard splits (e.g., CVC-ClinicDB 550/62, Kvasir-SEG 900/100). If the FOCUS-Med margin shrinks below statistical significance, the SOTA claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of state-of-the-art performance rests on Table 1. The baseline numbers are implausible and inconsistent with published results. PraNet is reported at 69.35% Dice on CVC-ClinicDB, while Fan et al. (MICCAI 2020) report roughly 90% for the same dataset. U-Net at 66.79% Dice is also far below typical reported performance (~80%). The paper specifies test-set sizes (120 Kvasir-SEG, 30 CVC-ClinicDB, 200 CVC-ColonDB, 80 EndoScene) that differ from standard splits used in the baseline papers (e.g., CVC-ClinicDB is commonly 550/62). No training hyperparameters, data preprocessing, augmentation, or split details are given. If baselines were trained on smaller or different data, with shorter schedules, or at lower resolution, the reported margins (e.g., +5.5 Dice over DCRNet on CVC-ClinicDB) are artifacts of an unfair comparison. The paper does state that models were implemented from official codes, but without a protocol description, the comparison cannot be reproduced or validated. This directly undermines the SOTA claim, which is the paper's main contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"FOCUS-Med is a deep-learning architecture for endoscopic polyp segmentation that combines a ConvNeXt encoder with (i) a Dual-GCN block consisting of a spatial similarity graph and a Dijkstra shortest-path structural graph, (ii) a location-fused stand-alone self-attention module (LFSA), and (iii) a weighted fast normalized fusion (WFNF). The paper reports quantitative results on Kvasir-SEG, CVC-ClinicDB, EndoScene, and CVC-ColonDB, claims state-of-the-art performance across five metrics, and introduces a GPT-4o-based qualitative evaluation protocol. An ablation study and related work are said to be in an appendix.","tokens_in":12742,"tokens_out":6517,"duration_ms":62917,"significance":"The paper proposes a reasonably well-motivated combination of graph-based structural reasoning, stand-alone self-attention, and learnable fusion for polyp segmentation, and it includes complexity comparisons and a first attempt at LLM-based segmentation assessment. The Dijkstra-derived shortest-path attention is a potentially interesting extension of graph attention to medical images. However, the empirical evidence for the central SOTA claim is not currently convincing. No code is made available, the training protocol is absent, and the reported baseline numbers are inconsistent with published results. The significance of the method therefore cannot be assessed until the evaluation is made reproducible and fair.","major_comments":[{"comment":"The SOTA claim rests on a comparison that is not credible as reported. PraNet is listed at 69.35% Dice on CVC-ClinicDB, whereas the original PraNet paper reports ~90% for that dataset; U-Net at 66.79% is also far below common published values. The stated test sizes (30 CVC-ClinicDB, 200 CVC-ColonDB, 80 EndoScene, 120 Kvasir) do not match the standard splits of the baseline papers (e.g., 62 test images for ClinicDB in PraNet). Since no training protocol is given, the large margins in Table 1 may simply reflect undertrained baselines or different data usage. This must be fixed by re-running all methods under an identical, explicitly described protocol and reporting standard splits.","section":"§4.1, Table 1"},{"comment":"The manuscript does not state the loss function, optimizer, learning rate, batch size, number of epochs, input resolution, data augmentation, or train/validation/test split. Without these details, none of the quantitative results, including the Wilcoxon tests in §4.1, can be reproduced. Please add a full experimental setup section and, ideally, release code to make the comparison verifiable.","section":"§4.1"},{"comment":"The abstract states SOTA 'across five key metrics', but Table 1 itself contradicts this on Kvasir-SEG: FOCUS-Med has MAE 3.20% vs DCRNet 2.90%, Boundary F 79.43% vs 82.05%, and S-measure 91.43% vs 91.49%. Thus the model is not uniformly best on all five metrics on that dataset. Please qualify the claim per dataset/metric.","section":"Abstract and Table 1"},{"comment":"The LLM evaluation is described as 'expert-aligned' but no evidence of alignment with human experts is provided. The prompt, GPT-4o version, temperature, number of test samples, and whether scores were averaged over repeated calls are not reported; no statistical test accompanies Table 2. At best this is an exploratory, LLM-simulated rating; it should not be presented as expert-aligned until calibrated against gastroenterologist judgments.","section":"§5"},{"comment":"Key hyperparameters for Algorithm 1 — maximum path length C, sampling ratio r, number of attention heads K — are never specified. Eq. (6) and (7) introduce trainable/weight parameters ω1, ω2, ω4–ω6 and ϵ without initialization or normalization constraints. Moreover, the paper repeatedly refers to an Appendix (related work, ablation tables, GPT-4o responses) that is not present in the submitted text. This prevents full evaluation of the method's design and the ablation claims.","section":"§3 and missing Appendix"}],"minor_comments":[{"comment":"'In summary, the proposed DSFNet...' — the model name should be FOCUS-Med.","section":"§4.1"},{"comment":"FOCUS-Med* uses an asterisk but no footnote defines it.","section":"Table 1"},{"comment":"Dataset/model naming is inconsistent: 'Endoscene'/'EndoScene', 'UNET'/'U-Net', 'ParNet'/'PraNet'.","section":"Throughout"},{"comment":"These references appear to point to papers on temporal action localization and fashion image retrieval, not the polyp segmentation baselines ACSNet and DCRNet used in Table 1. Verify these citations.","section":"References [10] and [25]"},{"comment":"Wilcoxon tests are reported only for Kvasir-SEG; p-values for the other datasets are missing.","section":"§4.1 Statistical Analysis"},{"comment":"Please specify the GPT-4o model version and API/interface details; the current citation to a general ChatGPT URL is insufficient for reproducibility.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"I agree with the reader's assessment that the baseline comparison is the main weakness of this manuscript. I recommend major revision rather than rejection because the issues are addressable with a full re-run under a standardized protocol, complete reporting of training details, and inclusion of the missing appendix. I also recommend checking the two references that seem to cite unrelated papers, as this is a reproducibility concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the architecture is a reasonable remix of existing modules, but the paper's central SOTA claim rests on a comparison that doesn't survive contact with the published literature. The missing training protocol and implausible baseline numbers make the headline result unverifiable.\n\nWhat's actually new: using Dijkstra shortest paths to define multi-hop attention on a graph built from CNN features is an interesting twist, and the LFSA variant with relative position embeddings is a clean idea. The LLM-based qualitative scoring is a fresh idea, though it is not validated against any human or established metric.\n\nWhat works: the method section is mostly readable; the components are clearly specified (even if Eq. 7 looks like it has six weights for three feature maps, and Eq. 5 mentions beta_c without defining it). The model's efficiency table is useful.\n\nSoft spots, in order of severity:\n1. The comparison is unusable. PraNet is reported at 69.35% Dice on CVC-ClinicDB vs roughly 90% in the original paper; U-Net at 66.79% is also far below typical. The test-set sizes (30, 80, 200) don't match the standard splits in the baseline papers. No optimizer, loss, epochs, resolution, or augmentations are given. This is not a minor omission; it is the load-bearing evidence.\n2. The ablation study is relegated to a missing appendix, so the claimed contribution of each module cannot be checked.\n3. The LLM evaluation is labeled 'expert-aligned' but no human expert was involved; GPT-4o's Likert ratings are used as if they were clinical judgment.\n4. The abstract boasts SOTA 'across five key metrics,' but Table 1 shows FOCUS-Med is not best on Kvasir-SEG for MAE, Boundary F, or S measure.\n5. Some references don't match the methods they are compared against (ACSNet and DCRNet are not polyp segmentation methods in the cited papers).\n\nThe summary sentence even calls the model DSFNet, which suggests a last-minute name change.\n\nMy take: this is a desk-reject or at most a 'revise before review' situation. The underlying modules might be worth a paper, but this manuscript does not provide a reliable empirical test of them. If the authors return with full training details, standard splits, corrected baselines, and the appendix, I'd be happy to see it again.","headline":"A plausible architecture, but the empirical claims are unsupported: missing training protocol and inconsistent baselines make the SOTA assertion unverifiable; the LLM evaluation is unvalidated.","tokens_in":13261,"tokens_out":3649,"would_cite":false,"duration_ms":36009,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FOCUS-Med claims state-of-the-art polyp segmentation on four colonoscopy benchmarks by fusing spatial and shortest-path structural graphs with location-aware self-attention.","keywords":["polyp segmentation","endoscopic imaging","graph neural network","Dual-GCN","self-attention","Dijkstra shortest path","medical image segmentation","LLM evaluation"],"falsifier":"Retrain every baseline and FOCUS-Med from a single shared codebase with identical data splits, input resolution, optimizer, and epoch count, then compare Dice and IoU on all four datasets. The state-of-the-art claim stands only if the reported margins persist.","tokens_in":12342,"feed_emoji":"🩺","tokens_out":5943,"duration_ms":51597,"temperature":0.7,"pith_summary":"This paper argues that precise polyp segmentation in endoscopic images is best served by fusing two complementary graph views of the image—a spatial graph that captures local pixel relationships and a structural graph built from shortest paths that captures long-range topological dependencies—and then blending the result with a location-aware self-attention decoder. The proposed model, FOCUS-Med, adds trainable weighted fusion to combine multi-scale encoder–decoder features, and uses a large language model to provide clinically framed qualitative scoring of segmentation quality. On four public colonoscopy datasets, the paper reports state-of-the-art Dice, IoU, MAE, boundary F, and structure measure scores, and claims the first use of an LLM for evaluating such segmentations. If the results hold under a fair comparison, the architecture offers a practical way to combine graph-based structure reasoning with attention-based context modeling for a clinically important task.","feed_headline":"FOCUS-Med reports best Dice on 4 colonoscopy datasets","feed_subtitle":"Spatial and structural graphs plus location-aware attention lift polyp segmentation across five metrics.","key_machinery":"The load-bearing mechanism is the Dual-GCN feature enhancement block: a spatial graph constructed from pairwise pixel similarity and a structural graph whose edges are weighted by attention coefficients and whose messages follow cheapest paths (retained up to a maximum length). The path-level and cross-length attention updates in this block let distant but semantically related regions influence each other's features. Around this block sit LFSA, which injects relative position into self-attention keys, and WFNF, which adaptively fuses decoder, skip, and bottleneck features.","core_discovery":"FOCUS-Med is a ConvNeXt-based encoder-decoder whose bottleneck contains a Dual-GCN block: one graph models fine-grained spatial similarity among pixels, the other builds a structural graph whose edge costs come from averaged attention coefficients and whose node updates aggregate features along shortest paths, allowing each node to attend to high-order semantically relevant nodes. The decoder inserts a location-fused stand-alone self-attention (LFSA) that adds row and column offsets into the attention key, and a weighted fast normalized fusion (WFNF) combines three feature sources with trainable weights. The paper reports that this combination outperforms existing polyp segmentation models o","pith_inferences":["The reported baseline scores (for example, PraNet at 69.35% Dice on CVC-ClinicDB) are far below the numbers published in the original baseline papers; a reader who re-runs all models under identical training settings may find the margins shrink, which would weaken the state-of-the-art claim.","The shortest-path structural graph has a natural interpretation as a message-passing scheme that lets features travel through chains of similar regions, a mechanism that could transfer to other medical segmentation tasks with elongated or branching structures such as vessels or nerves.","The LLM scoring protocol treats the language model as a proxy for clinical judgment, but the paper does not calibrate it against human raters; a direct comparison with gastroenterologist scores would test whether the LLM adds signal beyond Dice and IoU."],"forward_implications":["On the benchmarks tested, FOCUS-Med reports the highest Dice and IoU, and either best or near-best MAE, boundary F, and S-measure, across all four datasets.","The five-metric gains are statistically significant on Kvasir-SEG (Wilcoxon p-values all below 0.001), so the improvement is not attributed to a single lucky split.","The model runs at about 10.6 FPS with 43.6 million parameters, meaning the accuracy gains do not require prohibitive compute and could plausibly fit in a clinical workflow.","The LLM-based evaluation shows more consistent high scores on boundary accuracy and clinical usefulness, suggesting the model avoids low-quality tail cases.","Ablation results reported in the appendix attribute the gains to the Dual-GCN block, LFSA, and WFNF, with each component contributing."],"supporting_citations":[{"why":"Supplies the ConvNeXt backbone that produces the feature hierarchy FOCUS-Med refines.","marker":"[12]"},{"why":"Introduces shortest-path graph attention, the mechanism on which the structural graph construction is based.","marker":"[15]"},{"why":"Provides Dijkstra's algorithm used to compute the cheapest paths that define the structural graph edges.","marker":"[17]"},{"why":"Establishes the self-attention formulation that LFSA extends with relative location offsets.","marker":"[18]"},{"why":"Supplies the Kvasir-SEG benchmark dataset used for one of the four evaluation comparisons.","marker":"[19]"},{"why":"Supplies the CVC-ClinicDB benchmark dataset used for evaluation.","marker":"[20]"},{"why":"Supplies the EndoScene benchmark dataset used for evaluation.","marker":"[21]"},{"why":"Supplies the CVC-ColonDB benchmark dataset used for evaluation, including the LLM qualitative comparison.","marker":"[22]"},{"why":"A baseline (DCR-Net) that FOCUS-Med must beat and the comparison model in the LLM evaluation.","marker":"[25]"},{"why":"Defines the structure measure (S-measure) metric used as one of the five evaluation criteria.","marker":"[27]"}],"fun_headline_variants":["FOCUS-Med sets new standard for polyp segmentation","Graph attention network boosts colonoscopy imaging","Spatial and structural cues refine polyp boundaries","FOCUS-Med outperforms prior polyp segmentation models","LLM-assisted evaluation validates FOCUS-Med's top scores"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The evaluation depends on all baseline models being trained and tested under the same protocol as FOCUS-Med, including data split, loss, optimizer, and input size; the paper does not disclose this protocol, and some reported baseline numbers differ widely from their published values.","fun_headline_variants_meta":{"raw":{"variants":["FOCUS-Med sets new standard for polyp segmentation","Graph attention network boosts colonoscopy imaging","Spatial and structural cues refine polyp boundaries","FOCUS-Med outperforms prior polyp segmentation models","LLM-assisted evaluation validates FOCUS-Med's top scores"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000565,"raw_usage":{"total_tokens":2520,"prompt_tokens":756,"completion_tokens":1764,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":1690}},"tokens_in":500,"tokens_out":1764,"duration_ms":13358,"temperature":1.0,"reasoning_tokens":1690,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:22:36.236312+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain every baseline and FOCUS-Med from a single shared codebase with identical data splits, input resolution, optimizer, and epoch count, then compare Dice and IoU on all four datasets. The state-of-the-art claim stands only if the reported margins persist.","supporting_citations":[{"cited_title":"A convnet for the 2020s","cited_arxiv_id":null,"evidence_quote":"Supplies the ConvNeXt backbone that produces the feature hierarchy FOCUS-Med refines."},{"cited_title":"SPAGAN: Shortest Path Graph Attention Network","cited_arxiv_id":"2101.03464","evidence_quote":"Introduces shortest-path graph attention, the mechanism on which the structural graph construction is based."},{"cited_title":"A note on two problems in connexion with graphs","cited_arxiv_id":null,"evidence_quote":"Provides Dijkstra's algorithm used to compute the cheapest paths that define the structural graph edges."},{"cited_title":"Attention is all you need","cited_arxiv_id":null,"evidence_quote":"Establishes the self-attention formulation that LFSA extends with relative location offsets."},{"cited_title":"Kvasir-seg: A segmented polyp dataset","cited_arxiv_id":null,"evidence_quote":"Supplies the Kvasir-SEG benchmark dataset used for one of the four evaluation comparisons."},{"cited_title":"Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs","cited_arxiv_id":null,"evidence_quote":"Supplies the CVC-ClinicDB benchmark dataset used for evaluation."},{"cited_title":"A benchmark for endoluminal scene segmentation of colonoscopy images","cited_arxiv_id":null,"evidence_quote":"Supplies the EndoScene benchmark dataset used for evaluation."},{"cited_title":"Automated polyp detection in colonoscopy videos using shape and context information","cited_arxiv_id":null,"evidence_quote":"Supplies the CVC-ColonDB benchmark dataset used for evaluation, including the LLM qualitative comparison."},{"cited_title":"Dcr-net: Dilated convolutional residual network for fashion image retrieval","cited_arxiv_id":null,"evidence_quote":"A baseline (DCR-Net) that FOCUS-Med must beat and the comparison model in the LLM evaluation."},{"cited_title":"Structure-measure: A new way to eval- uate foreground maps","cited_arxiv_id":null,"evidence_quote":"Defines the structure measure (S-measure) metric used as one of the five evaluation criteria."}],"review_version":1}