{"id":"2710b54f-40c8-4041-922e-38607e4746e3","arxiv_id":"2607.16327","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A localization-infused vision-language fusion method converts textual location cues into multi-scale localization predictions and uses them to guide medical image segmentation, outperforming prior methods on three benchmarks.","lead":"This paper introduces LoG, a framework that makes text reports about lesion location an explicit part of medical image segmentation, using predicted localization maps to guide the model at three levels. It reports top Dice scores on COVID-19 X-ray, CT, and polyp datasets, suggesting that location-aware text guidance can reduce missed or drifting regions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Mixed-protocol comparison: Table I mixes official results and re-implementations with different splits/resolutions and no repeated-seed variance; the 0.16-point QaTa margin over PCSA-Seg is within likely noise, so 'consistently outperforming' is not yet established.","rationale":"The reader's identified weakest assumption is exactly the load-bearing concern: the comparative claim is built on a Table I that mixes published numbers and re-implementations under different protocols, with no statistical uncertainty. I agree with the conditional verdict. The paper's internal evidence—component ablations, sensitivity analysis, and visualizations—is coherent and suggests the localization-infused design is beneficial. Table III shows removing localization costs 3.44 average DSC, which is a substantial effect. However, the headline claim is about consistently beating state-of-the-art methods, and that claim requires a fair, controlled comparison. The absence of code/weights and repeated seeds makes the 0.16-point margin on QaTa-COV19 especially fragile. I also note a minor internal inconsistency: the 'without localization' ablation in Table III (avg 85.52) does not exactly match the baseline in Table II (avg 85.15), even though the text describes them equivalently; this slightly clouds the attribution of gains but is secondary to the comparison-protocol issue. Therefore the verdict should remain CONDITIONAL, pending a unified-protocol re-evaluation or a qualified, uncertainty-aware comparison.","tokens_in":19651,"tokens_out":7171,"duration_ms":69323,"concrete_test":"Run the top-3 text-guided competitors (PCSA-Seg, FMISeg, TeViA) and LoG under a single unified protocol on all three datasets: same 224x224 resize, same data splits (80/20 for QaTa-COV19 per [13]; 2183/273/273 for MosMedData+ per [12]; 8:1:1 for Kvasir-SEG per [48]), same augmentation and 100-epoch training schedule, with at least 3 random seeds each. If LoG's average DSC advantage over the best competitor on QaTa-COV19 falls below ~0.3 or overlaps within ±1 standard deviation, the 'consistently outperforming' claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that LoG 'consistently outperforms state-of-the-art medical image segmentation methods.' This claim rests entirely on Table I, but the protocol described in Section IV-C explicitly mixes 'official results reported in the corresponding papers when available' with re-implementations. Official numbers are not apples-to-apples: different papers use different image sizes, training schedules, validation splits, data augmentation, and post-processing. No repeated seeds or confidence intervals are reported, so even a 0.16 DSC margin over PCSA-Seg on QaTa-COV19 (91.59 vs. 91.43) is within typical run-to-run variability. The gap over FMISeg is 0.38 DSC, still small. On Kvasir-SEG the margin is larger (94.59 vs. 92.24), but the comparison methods may have been evaluated under different splits or resolution settings; without a unified protocol the observed advantage could reflect evaluation conditions rather than architecture. The ablations are internally consistent and suggest the localization branch contributes, but they do not establish superiority over external methods. Since no code or weights are released ('will be made publicly available upon acceptance'), the comparative results cannot be independently verified. This does not invalidate the method's internal logic, but it does mean the headline comparative claim is under-supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LoG, a text-guided medical image segmentation framework that augments a ConvNeXt-Tiny/BiomedBERT encoder pair with multi-scale patch-level localization prediction. Localization predictions are used in three fusion mechanisms: localization-guided feature fusion (LGFF) via cross-attention, localization-gated attention fusion (LGAF) via spatial gating, and localization-constrained loss fusion (LCLF) via a consistency loss. On QaTa-COV19, MosMedData+, and Kvasir-SEG the authors report DSC of 91.59%, 80.71%, and 94.59%, respectively, outperforming compared image-only, pretraining-based, and text-guided methods. Ablations show monotonic gains from adding each component and identify localization prediction as the largest contributor.","tokens_in":19988,"tokens_out":10124,"duration_ms":88100,"significance":"If the empirical claims hold, LoG is a useful architecture-level contribution: it converts holistic textual embeddings into explicit multi-scale localization cues and integrates them at feature, attention, and loss levels. The ablation logic is clean, and the authors correctly supervise localization against patch labels derived from ground-truth masks (Section III-E), preventing the consistency loss from trivial collapse. The main weakness is that the headline comparative claim rests on a mixed-protocol comparison without statistical validation; the architecture could still be sound even if the benchmark comparison is contaminated. Code and weights are not yet available, so independent verification is currently limited.","major_comments":[{"comment":"The headline claim is not established by the reported evidence. Table I mixes official results from other papers with re-implementations, so the compared numbers are not produced under a common protocol (Section IV-B: 224x224, batch 32, 100 epochs). Official numbers can differ in image size, schedule, split, and post-processing. Margins are small on two datasets: 0.16 DSC over PCSA-Seg on QaTa-COV19 and 1.41 DSC over FMISeg on MosMedData+. Moreover, no-text LoG (Table VII, 85.99 avg) already exceeds several text-guided baselines, so part of the apparent advantage may reflect backbone capacity rather than the proposed cross-modal mechanism. Please re-evaluate all comparison methods under one protocol, identify which entries are official vs re-implemented, and report multi-seed mean±std with significance tests, or soften the claim.","section":"Section IV-C, Table I"},{"comment":"Component ablations are internally consistent but are single Dice values without variance. Claimed gains are very small: 0.29% for updating F_T (Table III), 0.41% for multi-scale vs single-scale LGAF (Table IV), and 0.14-0.25% for removing L_con at individual scales (Table V). These differences are within typical run-to-run noise. The claim that each fusion level contributes monotonically needs support from repeated runs. Please provide mean±std over at least three seeds (or paired significance tests) for the main ablations, or state that differences are not statistically significant.","section":"Tables II-V"}],"minor_comments":[{"comment":"The loss weights lambda_loc (0.5/0.4/0.2 per dataset) and lambda_con are chosen empirically on the validation set, but no sensitivity analysis is reported for these choices. A small sensitivity table would help assess robustness.","section":"Section IV-B, Eq. (5)"},{"comment":"The consistency loss denominator can be zero if the predicted segmentation is empty. Add a small epsilon (or handle the empty-prediction case) to avoid numerical instability.","section":"Section III-E, Eq. (4)"},{"comment":"The DSC labels below the qualitative panels appear misaligned (six numeric values are shown for what appears to be seven method columns). Please clarify which value corresponds to each method, and fix the 'Qata-COV19' typo.","section":"Fig. 5"},{"comment":"Please specify the exact fusion operation used in the Table II baseline and whether the same visual encoder/decoder and training budget are used. This will make the baseline comparison more transparent.","section":"Section II-A / Experimental setup"},{"comment":"The paper does not report parameter counts, FLOPs, or inference time for LoG and the compared methods. Given the added cross-attention and localization heads, these numbers would help readers assess the cost of the improvement.","section":"Section IV-B"},{"comment":"The statement that code will be released only upon acceptance limits immediate reproducibility. Consider providing a public repository with the main scripts at submission.","section":"Code Availability"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a competent engineering contribution with a sound internal ablation logic. The main risk is the mixed-protocol comparison and the absence of variance information, which directly affects the headline claim. If the authors can provide a unified re-evaluation and multi-seed statistics, the paper would likely be acceptable. I see no circularity issue: the localization branch is supervised from ground-truth masks, and the consistency loss is protected against collapse by L_loc. No concerns about citation patterns or novelty disclosure beyond the authors' own prior components."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a well-built methods paper for text-guided medical segmentation, and the core architecture is sensible. The claim that it 'consistently outperforms' prior work is not yet supported, because the comparison table mixes official results with re-implementations and reports no variance. The paper is worth a serious referee, but it needs a cleaner comparison before publication.\n\nWhat's new: LoG explicitly predicts multi-scale localization maps from image–text pairs and uses them at three fusion levels — feature fusion, gated attention, and loss-level consistency. That's a real departure from the cited prior work, which mostly fuses holistic text embeddings at the feature level. The ablations are the strongest part: removing the localization task costs 3.44% average DSC, and each fusion level contributes monotonically. The authors also guard against the obvious failure mode of the consistency loss by supervising the localization maps against patch labels, so the loss can't be trivially minimized. That's careful. The sensitivity analysis for patch size and the text truncation experiments are useful.\n\nThe soft spot is Table I. Section IV-C says they used official results when available and re-implemented the rest. That means different resolutions, training schedules, and split protocols are probably in the same table. Without repeated seeds or confidence intervals, the QaTa margin of 0.16 DSC over PCSA-Seg is within run-to-run noise, and even the larger Kvasir margin could be influenced by protocol differences. The ablation tables, by contrast, use a unified protocol, so they're internally valid. It's the external comparison that's under-supported. There's also no code or weights released yet, so the numbers can't be independently checked.\n\nI wouldn't call this circular or sloppy. The authors are clear about their protocol and their limitations. But the headline claim needs to be either backed by a re-run of baselines under a common protocol or qualified more carefully.\n\nWho should read it: anyone working on language-guided medical segmentation. The localization-oriented extraction idea is likely to be picked up. I'd send it to review, with the request that the authors unify the comparison and report variance. If those numbers hold, it's a solid contribution.","headline":"A well-designed localization-infused fusion method whose internal ablations are consistent, but the 'consistently outperforms' claim outruns the mixed-protocol comparison in Table I.","tokens_in":20439,"tokens_out":3177,"would_cite":true,"duration_ms":43683,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LoG turns clinical text location cues into multi-scale localization maps that guide medical image segmentation, beating prior text-guided methods on three datasets.","keywords":["medical image segmentation","text-guided segmentation","vision-language fusion","localization","multi-scale prediction","clinical reports","Dice score","deep learning"],"falsifier":"Run every baseline with the same training code, image resolution, augmentation, and epoch count as LoG; if LoG still leads on all three datasets, the architectural contribution is confirmed.","tokens_in":19563,"feed_emoji":"🩻","tokens_out":5792,"duration_ms":51210,"temperature":0.7,"pith_summary":"The paper argues that text-guided medical image segmentation improves when the model is forced to explicitly predict where the target is, at multiple scales, rather than just blending a holistic text embedding into visual features. It proposes LoG, which jointly predicts multi-scale localization maps from image and report, then injects them into segmentation at three levels: feature fusion, attention gating, and loss supervision. On chest X-ray, CT, and colonoscopy datasets with paired clinical reports, LoG reports Dice scores of 91.59%, 80.71%, and 94.59%, consistently ahead of existing methods. The ablation evidence pins the gain to the localization branch: removing it costs 3.44% average Dice. The authors note the method assumes paired structured reports and was tested only on 2D images.","feed_headline":"Localization cues from clinical text lift medical segmentation","feed_subtitle":"LoG fuses report location cues at feature, attention, and loss levels to beat existing text-guided segmenters.","key_machinery":"The central object is a multi-scale localization-oriented semantic extraction module that converts an image-report pair into feature-space semantic tokens and explicit patch-wise localization maps at four scales. The maps are produced by a lightweight head on top of token-level cross-attention between a visual encoder and a medical text encoder. The maps are then recycled into the segmentation pipeline in three fusion mechanisms: LGFF (cross-attention-based feature infusion), LGAF (spatial gating with a sigmoid mask built from all four localization maps), and LCLF (a localization-consistency loss term). This design turns location language into spatial supervision.","core_discovery":"LoG's central discovery is that target localization should be an explicit intermediate representation for vision-language fusion in medical segmentation. By performing patch-wise localization prediction at four image scales with bidirectional cross-attention between visual and textual features, the model extracts what the authors call localization-oriented semantics, embedded both as implicit feature-space tokens and explicit localization maps. These semantics are then used in three complementary ways: Localization-Guided Feature Fusion cross-attends visual tokens to localized semantic tokens; Localization-Gated Attention Fusion multiplies visual features by a spatial gating mask derived fro","pith_inferences":["Because the localization branch helps even with no text at all, the same localization-infused design could be applied to image-only segmentation pipelines as a form of self-supervision.","The localization-consistency loss should transfer to 3D volumetric segmentation, where anatomical location language is more informative, if patch-wise labels are replaced with voxel-grid labels.","The authors' stated limitation to structured reports suggests a natural stress test: feeding the model free-text clinical notes with paraphrased or contradictory location phrases and measuring whether localization maps and Dice scores degrade."],"forward_implications":["Text-guided segmentation systems should treat explicit localization prediction as a core component: removing the localization branch causes the largest single performance drop (3.44% average DSC).","Multi-level fusion is additive: each of the three levels contributes a small but consistent gain (about 2.17%, 0.95%, and 0.69% average DSC, respectively).","The localization branch remains useful even without text input: the no-text variant still beats many text-guided baselines, suggesting spatial self-guidance is part of the gain.","Truncating reports to 4 tokens drops average DSC by 2.35%, while truncating to 12 tokens slightly improves it, implying most location cues sit in a small central span of the report.","Localization patch size is not a critical hyperparameter; performance is stable across the tested granularities."],"fun_headline_variants":["Report location cues sharpen text-guided medical segmentation","LoG fuses report location into vision-language segmentation","Explicit location cues from reports help segment medical images","Text-guided segmentation gains from localization-aware fusion","Medical segmentation improves when clinical text location is fused"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The head-to-head comparison numbers come from a mix of officially reported results and re-implementations, so the claimed 'consistently outperforms' rests on the assumption that differences in training protocol do not account for the margin.","fun_headline_variants_meta":{"raw":{"variants":["Report location cues sharpen text-guided medical segmentation","LoG fuses report location into vision-language segmentation","Explicit location cues from reports help segment medical images","Text-guided segmentation gains from localization-aware fusion","Medical segmentation improves when clinical text location is fused"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000863,"raw_usage":{"total_tokens":3599,"prompt_tokens":786,"completion_tokens":2813,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":2742}},"tokens_in":530,"tokens_out":2813,"duration_ms":20961,"temperature":1.0,"reasoning_tokens":2742,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T03:28:19.024234+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run every baseline with the same training code, image resolution, augmentation, and epoch count as LoG; if LoG still leads on all three datasets, the architectural contribution is confirmed.","supporting_citations":[],"review_version":1}