{"id":"4cb79359-bc10-4dea-b667-55c381a47937","arxiv_id":"2512.01085","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MedGrounder grounds radiology sentences to zero, one, or multiple scored image regions, outperforming single-box and grounded-report baselines on multi-box and non-groundable phrases.","lead":"The paper redefines medical phrase grounding so each radiology sentence maps to zero, one, or several scored image regions. It introduces MedGrounder, which beats prior single-box systems on multi-region and non-findings phrases with far fewer human box annotations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's 'outperforms ... non-groundable phrases' is contradicted by Table III: MAIRA-2 GR N-Acc 91.0 vs MedGrounder 90.1 on PadChest-GR.","rationale":"The reader's weakest assumption was that the LLM-cleaned Chest ImaGenome labels are faithful, which is a valid concern but is an external validity threat. My concern is an internal inconsistency: the paper's own Table III contradicts the abstract's claim that MedGrounder outperforms GRG baselines on non-groundable phrases. This is more directly load-bearing because it undermines a headline claim as written, without needing to speculate about data quality. I partially agree with the reader because both concerns center on different pillars of the central claim (data quality vs. empirical outperformance). The verdict remains CONDITIONAL: the GMPG formulation and data-efficiency results are interesting, but the paper must either add bootstrap CIs to demonstrate the N-Acc gap is noise, or amend the abstract. The reader already requested error bars and softening, so no verdict change is needed.","tokens_in":15663,"tokens_out":3789,"duration_ms":38701,"concrete_test":"Compute paired bootstrap 95% confidence intervals for the N-Acc difference (MedGrounder fine-tuned M,P minus MAIRA-2 GR) on the PadChest-GR test set (n=874 non-groundable sentences). If the CI includes 0 or is positive, the 'outperforms' claim may be retained; if the CI is negative and excludes 0, revise the abstract to say 'comparable on non-groundable phrases' and report the CI.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on MedGrounder beating REC-style and GRG baselines on non-groundable phrases. In Table III (PadChest-GR test, n=2,112), the fine-tuned MedGrounder (M,P) achieves N-Acc 90.1%, while MAIRA-2 GR achieves 91.0%. This is a direct, within-paper contradiction of the abstract's wording. The gap is small (0.9 points), but without error bars or significance testing, the claim 'outperforms' is unsupported; MAIRA-2 is at least at parity. This is not a peripheral issue—groundability (abstaining from non-groundable phrases) is one of the three core GMPG requirements and a key differentiator from REC methods. The zero-shot MedGrounder N-Acc (86.5%) is also far below MAIRA-2 GR's 91.0%, though MAIRA-2 was trained on PadChest-GR so a zero-shot comparison is unfair; the fine-tuned comparison is the appropriate one. The reader's concern about the unvalidated LLM-based weak labels (89.6% reduction, no manual audit) is plausible, but it is speculative without evidence of systematic bias. The N-Acc contradiction is concrete and can be settled by re-running the evaluation with confidence intervals. If the gap is within noise, the claim may stand; if not, the abstract must be softened to 'comparable' or 'on par' on non-groundable phrases.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reformulates medical phrase grounding as generalised medical phrase grounding (GMPG), in which a phrase is mapped to zero, one, or multiple scored bounding boxes. To solve GMPG the authors propose MedGrounder, an MDETR-style detector pretrained on an LLM-cleaned version of Chest ImaGenome and fine-tuned on MS-CXR and PadChest-GR. They report strong zero-shot transfer, gains over REC-style and GRG baselines on multi-box and non-groundable phrases, and a modular pipeline in which MedGrounder grounds the output of existing report generators. Ablations indicate pretraining is essential, WBF post-processing helps, and a clinical language encoder improves results.","tokens_in":16032,"tokens_out":10397,"duration_ms":84536,"significance":"If the claims hold, GMPG is a well-motivated task reformulation and MedGrounder demonstrates a data-efficient route to multi-box grounding and negative suppression. The paper's strengths are its explicit zero/one/multi formulation, the weak-to-expert training recipe, the inclusion of non-grounding examples, and the planned release of code and data. The ablations are informative: from-scratch training collapses, and WBF and BioClinical ModernBERT give consistent gains. However, the headline non-groundable claim is not supported by the fine-tuned comparison in Table III, and the pretraining signal from the 89.6% filtered weak labels lacks validation. These issues are fixable but require additional experiments or a softened claim.","major_comments":[{"comment":"The abstract states that MedGrounder 'outperforms ... non-groundable phrases', but in the fine-tuned comparison on PadChest-GR, MAIRA-2 GR has N-Acc 91.0 while MedGrounder (M,P) has 90.1. Since N-Acc is the only non-groundable metric, the claimed advantage over GRG baselines on this requirement is unsupported. The 0.9-point gap is reported without confidence intervals or significance tests. N-Acc is also gameable: Table VII shows from-scratch models reach 99.8-99.9 N-Acc with P@F1 near zero. Please add interval estimates or significance tests, report N-Acc together with P@F1, and adjust the abstract to 'comparable' unless superiority is statistically supported.","section":"Abstract; Table III"},{"comment":"The central pretraining step is built on LLaMA-3-70B filtering that discards 89.6% of Chest ImaGenome sentence-box pairs, yet there is no manual audit, agreement study, or comparison against pretraining on the unfiltered dataset. Since Table VII establishes that pretraining is essential, a biased filter could be the source of the zero-shot and fine-tuning gains. I request a sample audit and a raw-vs-cleaned pretraining ablation, or at minimum a per-region error analysis showing the retained labels are not systematically wrong.","section":"§III-D.1; Fig. 3"},{"comment":"The modular-GRG conclusion rests heavily on a 38-case non-overlapping subset. The reported differences (e.g., Spatial F1 53.52 vs 52.03) are within plausible noise for n=38. Please provide confidence intervals or a larger non-overlapping evaluation set before claiming the pipeline matches or exceeds MAIRA-2 GR.","section":"Table VI; §V-B"}],"minor_comments":[{"comment":"Caption typo: 'MobernBERT' should be 'ModernBERT'.","section":"Fig. 2"},{"comment":"The notation 'αtj =0 = 0.1' is confusing; use α_{t_j=0}.","section":"Eq. (3)"},{"comment":"The PadChest-GR from-scratch row shows P@F1=41.3 with CH-F1=1.5; please verify the column alignment.","section":"Table VII"},{"comment":"Several numbers are merged in the text (e.g., '60.775.1', '086.5'); fix the spacing.","section":"Tables III/IV"},{"comment":"CH-F1 and Mask IoU accuracy are described only informally; provide formal definitions or equations.","section":"§III-E"},{"comment":"Figure contains typos: 'sihouette', 'Enlongated', 'efusion'.","section":"Fig. 7"},{"comment":"The 38-case subset is not described; report how it was selected and its disease distribution. Also 'subest' is a typo.","section":"§V-B"}],"recommendation":"major_revision","confidential_remarks":"The paper is not circular; the weak-label concern is a validation gap rather than an internal inconsistency. The main blocking issue is the abstract's overclaim on non-groundable phrases and the lack of uncertainty quantification. If the authors add confidence intervals and a cleaned-vs-raw pretraining comparison, the paper would be publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read this carefully. The core move — reformulate medical phrase grounding as a set prediction over zero-or-more scored boxes — is correct and clinically motivated. That alone is a useful contribution. The two-stage weak-to-expert training works: the ablation shows from-scratch training collapses, and multi-box scores are substantially better than the REC-style baselines. The modular GRG result on the 38-case subset is a reasonable proof of concept, no more.\n\nThe paper does some things well. The LLM-assisted cleanup of Chest ImaGenome is creative, and the disease-specific analysis (Fig 5) is honest about where grounding fails. The comparison tables are thorough. If the code and data actually ship, this becomes a usable benchmark.\n\nNow the soft spots. The abstract's claim of outperforming on non-groundable phrases is not supported by its own numbers. Table III: fine-tuned MedGrounder gets N-Acc 90.1 on PadChest-GR, while MAIRA-2 GR gets 91.0. That is parity, not superiority. On multi-box, MedGrounder is clearly ahead (28.6 vs 14.9 P@F1), so the overall story is mixed but mostly good. Still, the abstract needs rewording.\n\nSecond, there are no error bars or significance tests anywhere. A 0.9-point gap with no uncertainty is a coin flip. Given that N-Acc is one of the three GMPG pillars, the authors need to either run multiple seeds and report CIs or soften the language.\n\nThird, the weak-label cleaning discards 89.6% of the sentence-box pairs, and there is no manual audit, agreement study, or probe showing the retained labels are unbiased. The pretraining ablation shows pretraining matters, but it does not show that the cleaned labels are faithful. That is a legitimate concern, not a fatal one, but worth addressing.\n\nThe GRG experiment on 38 cases is fine as a preliminary but should not be the basis for a headline claim.\n\nFor a medical imaging venue, this deserves a serious referee. The task definition and the data pipeline are worth publishing even if the strongest claims need damping. I would send it out with instructions to focus on the N-Acc contradiction and the validation of the weak labels.","headline":"GMPG is a real step forward, but the abstract overstates the non-groundable advantage; MAIRA-2 GR is at parity on N-Acc.","tokens_in":16538,"tokens_out":4139,"would_cite":true,"duration_ms":36786,"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":"Medical phrase grounding should return a set of zero, one, or multiple scored image regions per sentence, and the new MedGrounder model shows this is achievable with far fewer human box annotations than end-to-end grounded report generators","keywords":["medical phrase grounding","generalised referring expression comprehension","chest X-ray","radiology report generation","DETR","weak supervision","zero-shot transfer","grounded report generation"],"falsifier":"A human audit of a random sample of GMPG-ImaGenome would settle it: if annotators often find that the retained region is not entailed by the sentence, or that a sensible region was discarded, the pretraining signal is biased. A cheaper quantitative check is an ablation training MedGrounder identically on raw Chest ImaGenome versus the cleaned version; if raw-pretrained performance matches or beats the cleaned version on multi-box and diffuse findings, the LLM filter is not the source of the reported gains.","tokens_in":15551,"feed_emoji":"🩻","tokens_out":5949,"duration_ms":54093,"temperature":0.7,"pith_summary":"Real radiology reports routinely violate the one-box-per-phrase assumption that most medical phrase grounding systems inherit from referring-expression comprehension: a sentence can describe findings in several regions, or be a negation or non-diagnostic statement that should ground to nothing. The paper redefines the task as generalised medical phrase grounding (GMPG), in which each sentence maps to a set of zero, one, or more scored bounding boxes, and introduces MedGrounder, a detection-transformer model trained in two stages. It first pretrains on a large-scale, weakly labelled sentence–anatomy alignment dataset (Chest ImaGenome, cleaned with an LLM that keeps only the most specific regions and removes spurious boxes) and then fine-tunes on small expert-annotated sets. The paper's central claim is that this recipe yields strong zero-shot transfer and state-of-the-art results on multi-region and non-groundable phrases, beating REC-style and grounded-report-generation baselines while using roughly one tenth of the human-annotated boxes of the strongest competitor. A sympathetic reader cares because, if correct, GMPG is a more faithful and data-efficient formulation of grounding, and it lets existing report generators be turned into grounded, verifiable systems without retraining them.","feed_headline":"MedGrounder maps each finding to zero, one, or many boxes","feed_subtitle":"Multi-region and negated phrases get grounded using one tenth of the human boxes.","key_machinery":"The load-bearing object is the GMPG task definition itself: a variable-size set of scored boxes, including the empty set, replaces the fixed single-box regression of REC. MedGrounder realises it with a DETR-style architecture—ResNet-101 and BioClinical ModernBERT encoders, a cross-modal transformer encoder-decoder, bipartite matching between predictions and ground-truth sets, and a confidence head whose threshold permits abstention. The two-stage training regime is the other piece of machinery: an LLM-filtered version of Chest ImaGenome supplies 426,749 weak sentence–box pairs that teach the model which anatomical regions a sentence is about, and a short fine-tuning stage on human boxes refi","core_discovery":"On PadChest-GR and MS-CXR, the pretrained-only MedGrounder outperforms all baselines in zero-shot transfer, and after fine-tuning it sets the best results on multi-box phrases—for example 28.6% P@F1=1 on PadChest-GR multi-box cases, nearly double the strongest competitor MAIRA-2, while abstaining correctly on non-groundable sentences. When composed with two different report generators (MAIRA-2 in finding-generation mode and CXRMate-RRG24), MedGrounder produces grounded reports whose RadFact spatial F1 is on par with or better than the end-to-end MAIRA-2 GR, despite being trained on one tenth of the human annotations. The authors attribute the transfer to the anatomical inductive bias learned","pith_inferences":["If the weak-label assumption holds, the same LLM-filtering recipe could be ported to other weakly aligned medical corpora (CT, MRI, pathology) to create pretraining data for grounding in those modalities.","The performance gap between anatomically fixed findings (cardiomegaly) and diffuse or spatially variable ones (lung opacity, rib fractures) suggests an intrinsic ceiling to what fixed anatomical-region priors can teach; further gains will require finer supervision such as segmentation masks or disease-specific box sets.","The modular generation-plus-grounding result implies that progress in text-only report generation can be directly converted into progress in verifiable grounded reporting, decoupling the two improvement curves.","A concrete test of the cleaning pipeline's value would be to compare MedGrounder pretrained on the raw versus the LLM-cleaned Chest ImaGenome; if raw pretraining matches or beats it on diffuse findings, the 89.6% discard is removing signal, not just noise."],"forward_implications":["GMPG becomes the natural evaluation setting for medical phrase grounding; future datasets and models should report metrics that separate non-groundable, single-box, and multi-box cases, not only mIoU.","Weakly supervised pretraining on sentence–anatomy alignments can substitute for a large fraction of human box annotation, lowering the annotation cost of building grounding systems.","Grounded report generation becomes modular: any existing report generator can be upgraded to a grounded generator by appending MedGrounder, avoiding expensive image–report–box training triplets.","Confidence scores and explicit abstention make AI-generated reports more checkable by radiologists, since every grounded sentence carries spatial evidence.","The released GMPG-ImaGenome cleanup provides a reusable weak-supervision resource for pretraining future grounding models."],"fun_headline_variants":["MedGrounder grounds any finding count with 10× fewer labels","Zero-shot MedGrounder beats tuned baselines on multi-box phrases","One tenth of the boxes, same grounded reports: MedGrounder","MedGrounder: multi-region grounding without per-phrase labels"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole pretraining benefit rests on the assumption that the LLM-cleaned Chest ImaGenome annotations are a faithful 'most specific regions' decomposition of each sentence; the paper reports no manual audit or agreement study of this cleaned data, and if the cleanup removes legitimate regions or keeps hallucinated ones, the model's inductive bias—and the zero-shot and fine-tuning gains built on it—are unvalidated.","fun_headline_variants_meta":{"raw":{"variants":["MedGrounder grounds any finding count with 10× fewer labels","Zero-shot MedGrounder beats tuned baselines on multi-box phrases","One tenth of the boxes, same grounded reports: MedGrounder","MedGrounder: multi-region grounding without per-phrase labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000562,"raw_usage":{"total_tokens":2512,"prompt_tokens":760,"completion_tokens":1752,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":1677}},"tokens_in":504,"tokens_out":1752,"duration_ms":14013,"temperature":1.0,"reasoning_tokens":1677,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T19:17:44.263106+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A human audit of a random sample of GMPG-ImaGenome would settle it: if annotators often find that the retained region is not entailed by the sentence, or that a sensible region was discarded, the pretraining signal is biased. A cheaper quantitative check is an ablation training MedGrounder identically on raw Chest ImaGenome versus the cleaned version; if raw-pretrained performance matches or beats the cleaned version on multi-box and diffuse findings, the LLM filter is not the source of the reported gains.","supporting_citations":[],"review_version":1}