{"id":"1134e911-ed7e-4917-88ec-bc8efd517ec1","arxiv_id":"2505.03401","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A longitudinal radiology report generator that aligns prior image-text features and fuses difference-aware features at every encoder stage to improve clinical efficacy scores.","lead":"DDaTR is a neural network that writes a radiology report by comparing a patient's current chest X-ray with the prior X-ray and prior report. It is designed to capture how findings change between visits, which could help radiologists track disease progression automatically.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The learned gates in Eq. (8) and Eq. (20) are the only defense against noisy or irrelevant prior exams, yet the paper never measures whether they actually suppress such priors; the reported CE gains could partly reflect copying stable findings from the prior report rather than difference-aware…","rationale":"The reader's weakest assumption identifies the same load-bearing point: the learned gate and the binary alpha are the only mechanisms protecting against noisy priors, and the paper does not measure how often the gate suppresses prior information. I agree with that diagnosis. The paper has genuine independent support: a public code repository, extensive comparisons across three datasets and multiple backbones, ablations showing consistent improvements, and an honest limitations section acknowledging hallucinations and misalignments. The central result may well hold, but the mechanism-level claim that the dynamic gates adaptively ignore irrelevant priors is not directly evidenced. The proposed concrete test would settle whether the concern lands: if the gates do suppress mismatched or label-disjoint priors, the stratified CE numbers should show it; if they do not, the reported gains may be inflated by prior-report copying on stable findings, and the difference-aware interpretation of the architecture would be unsupported. This does not change the reader's CONDITIONAL verdict, because the issue is empirical and testable rather than a demonstrated internal contradiction; it does mean the paper should add gate-behavior analysis and stratified results before the SOTA claim is relied upon.","tokens_in":21951,"tokens_out":5488,"duration_ms":65504,"concrete_test":"Stratify the MIMIC-CXR test set by (i) view mismatch between prior and current studies (frontal-to-lateral or lateral-to-frontal versus matched views) and (ii) CheXbert label overlap between the prior and current reference reports (Jaccard = 0 versus Jaccard > 0). On each stratum, compare DDaTR's CE F1 against the no-prior PromptMRG* baseline and against a control in which the prior report is replaced by a randomly selected other patient's report. Also record the mean and fraction of near-zero Fdy gate norms in each stratum. If DDaTR does not beat the no-prior baseline on the Jaccard = 0 or view-mismatched strata, or if the gate norms are not substantially smaller there than on matched-prior strata, the net-positive-prior assumption fails and the claimed SOTA improvement cannot be attributed to difference-aware temporal encoding.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DDaTR's longitudinal encoding improves CE over single-period and prior LRRG baselines. That claim depends on the assumption that the most recent prior exam, after DFAM alignment and DDAM gating, is net-positive for every current report where a prior exists. In the architecture, Eq. (8) always adds the DFAM output to the prior stream, and Eq. (20) always injects alpha*(Fdy⊙Fda) into the current stream whenever a prior exam is present; alpha is a binary indicator, not a learned selector. The only learned protection is the gate Fdy, but nothing in the training objective forces Fdy toward zero, and the paper reports no statistics on gate activations, no ablation that removes the prior report text, and no evaluation stratified by prior relevance or view mismatch. Because MIMIC-CXR prior reports often contain findings that persist into the current report, a model could inflate CE by copying stable prior-report statements while contributing little genuine difference awareness. Conversely, when the prior is a different view or contains unrelated findings, the gated features can degrade the current report; the paper's Fig. 2 case A discusses this risk only qualitatively. The reported p-value of 0.0415 is also computed by an unspecified procedure and cannot, as described, establish that the improvement is robust across the relevant subsets. Thus the load-bearing gap is not the architecture's novelty but the absence of evidence that the gate actually behaves as the safety mechanism the method requires.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"DDaTR is a longitudinal encoder for chest X-ray report generation. The paper builds on PromptMRG's diagnosis-prompt framework and inserts two modules at every stage of the visual encoder: DFAM aligns prior-image features with the prior report's frozen-BERT text features, and DDAM computes gated pixel-level differences between prior and current features. Residual connections pass prior information unidirectionally into the current stream; when no prior exists, alpha in Eq. (20) is zero and the model falls back to single-period encoding. Experiments on MIMIC-CXR, Longitudinal-MIMIC, and IU-Xray report higher Clinical Efficacy F1 than the baselines (e.g., 0.441 vs 0.389 for PromptMRG* on MIMIC-CXR, 0.527 vs 0.492 on Longitudinal-MIMIC), with three visual backbones, ablations, ReXrank comparisons, and Grad-CAM case studies. The paper releases code and explicitly acknowledges residual hallucinations and omissions in Section V.","tokens_in":22280,"tokens_out":12935,"duration_ms":129220,"significance":"The contribution is potentially valuable because the reported gains are on Clinical Efficacy, the metric most relevant to radiology practice, and they are consistent across MIMIC-CXR, Longitudinal-MIMIC, and three backbones; the code release and the explicit experimental protocol are strengths. The ablations in Table III and the controlled comparison against PromptMRG* in Table I support the broad claim that longitudinal encoding with DFAM/DDAM improves CE. I agree with the reader that there is no circularity: test labels are derived by CheXbert from ground-truth reports and prior inputs are real historical exams. The authors' own limitation statement in Section V correctly acknowledges that hallucinations and omissions remain, and I do not treat those as fatal. The main gaps are statistical rigor (one underspecified p-value, single runs) and mechanism evidence (no gate-level or progression-stratified analysis), both of which are addressable in revision.","major_comments":[{"comment":"The sentence reporting PF1=0.0415 < 0.05 as evidence of statistical significance is not verifiable as written because no test procedure is described, no unit of analysis is specified, and the MIMIC-CXR test set contains multiple studies from the same patient. Please specify the test (e.g., paired bootstrap or permutation at the patient level), report the effect size and confidence interval, and state whether any correction for multiple comparisons was applied. Since all metrics in Tables I-III come from single runs, please also report results over multiple seeds or otherwise justify that the headline F1 differences exceed run-to-run variability.","section":"Section IV-B (statistical significance)"},{"comment":"The only mechanisms that can down-weight an irrelevant or misleading prior exam are the learned gates in DFAM and DDAM, while alpha in Eq. (20) is a binary switch that always injects prior information when a prior exists. The paper does not report gate-activation statistics, does not ablate the prior report input, and does not stratify CE results by whether the prior exam is stable, progressive, or mismatched in view or laterality. Because stable findings are legitimately repeated in longitudinal reports, the observed CE gains could partly reflect copying stable prior-report statements rather than difference-aware reasoning. Please add quantitative evidence on gate suppression and a progression-aware breakdown of CE to substantiate the central 'dynamic difference-aware' claim.","section":"Section III-D, Eqs. (16) and (20)"},{"comment":"The ablation nomenclature is internally inconsistent: Table III says 'w/o DAM' removes the difference-aware module, while the text says 'DAM and FAM represent the configurations where dynamic fusion is removed from the DDAM and DFAM modules.' These are different operations, and the interpretation of the ablation (e.g., 'DAM plays a more critical role') depends on which one was actually removed. Please rename the rows to state precisely which component (DDAM, dynamic fusion, or both) is ablated in each experiment.","section":"Section III-C/D and Table III"},{"comment":"The 'Vanilla' baseline in Table II appears to concatenate visual and text features without the diagnosis-aware classification and prompt branch used in the PromptMRG framework and in DDaTR itself. As a result, the gains over Vanilla are not an ablation of DFAM/DDAM but combine the effect of the whole PromptMRG-style framework with the proposed longitudinal modules. Please state explicitly whether Vanilla includes the disease-classification branch and, if it does not, add a controlled comparison (e.g., PromptMRG with each backbone) so that the backbone generalization results isolate the contribution of the proposed modules.","section":"Section IV-A and Table II"}],"minor_comments":[{"comment":"The text says the decoder is conditioned on prompts derived from Lp, but Eq. (12) uses Lc; please clarify which labels are used and how they are converted into prompt tokens.","section":"Section III-B, Eq. (12)"},{"comment":"The CE metric is macro-averaged on MIMIC-CXR and IU-Xray and micro-averaged on Longitudinal-MIMIC; please state this explicitly wherever cross-dataset comparisons are discussed so readers do not over-interpret the magnitudes in Table I.","section":"Section IV-A"},{"comment":"For PromptMRG*, please report the visual backbone and evaluation split used so that the comparison is fully reproducible.","section":"Table I"},{"comment":"There is a typo 'an an increase' in the FAM ablation paragraph.","section":"Section IV-C"},{"comment":"The sub-figure labels include two 'G' entries; please re-letter them.","section":"Figure 9"},{"comment":"The phrase 'SOTA in both RRG and LRRG' should be qualified to Clinical Efficacy metrics, since the reported NLG scores are not state-of-the-art (e.g., Table I BLEU-4 on MIMIC-CXR is 0.113 for DDaTR versus 0.125 for RECAP).","section":"Abstract and Section I"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the paper makes a credible and useful empirical contribution, and the main comparison against PromptMRG* appears fair and reproducible in spirit. The concerns I raise are about overclaimed statistical significance and insufficient evidence for the mechanism; both are addressable with additional analysis and careful rewriting. I do not see a circularity problem or a fundamental flaw in the method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuine contribution to longitudinal radiology report generation. The architecture is new—DFAM and DDAM inserted at every stage of the visual encoder, with unidirectional residual flow—and the ablations back up that the modules matter. The three-benchmark, three-backbone evaluation is extensive, and the zero-shot IU-Xray result plus ReXrank comparison give a fair sense of where it stands.\n\nWhat it does well: the experiments are honest. NLG scores are basically flat, which is credible for a method aimed at clinical efficacy. The comparison against vanilla encoders and BioViL-T is useful, and the ablation study isolates each component's contribution. The analysis of hallucinations and misalignments in Fig. 8 is candid.\n\nThe soft spots are real but manageable. The biggest is the gating story. The stress-test concern lands: Fdy in Eq. (20) is the only learned defense against noisy or irrelevant priors, and the paper never measures whether it actually suppresses them. Alpha is a binary flag, not a learned selector. The w/o DF ablation removes dynamic fusion entirely, but doesn't isolate gate behavior on noisy priors. No gate-activation statistics, no stratification by view mismatch or prior relevance. Since MIMIC-CXR priors often share stable findings, the CE gains could partly reflect copying stable report statements rather than genuine difference awareness. That doesn't refute the central claim, but a referee should push on it.\n\nSecond, the significance test is thin. p=0.0415 with no procedural detail and no patient-level clustering. Single-run results are the norm here, but a few seeds would have made things sturdier.\n\nMinor: code is public but no commit hash; the parameter count is higher than PromptMRG, though the EfficientNetV2 variant is close. The 'CE matters more than NLG' argument is defensible but slightly self-serving, since CE comes from CheXbert, an imperfect labeler.\n\nOverall, the central argument holds up. The architecture is new, the experiments are broad, and limitations are discussed openly. The missing gate analysis is a fixable omission, not a fatal flaw. I'd send this to peer review and ask reviewers to request a proper significance analysis and a gate-behavior experiment. The paper deserves a serious referee's time.","headline":"The paper is a genuine, well-tested architecture contribution to longitudinal RRG; the main weaknesses are the unverified gating safety claim and a thin significance test, both fixable.","tokens_in":22805,"tokens_out":2960,"would_cite":true,"duration_ms":26996,"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":"DDaTR claims that longitudinal report generation improves when prior text and images are aligned and difference signals are fused into current features at every encoder stage, raising reported clinical-efficacy scores on three benchmarks.","keywords":["Radiology report generation","Longitudinal radiology report generation","Dynamic difference-aware module","Feature alignment","Temporal residual network","Chest X-ray","Clinical efficacy","Multimodal encoder"],"falsifier":"Take the longitudinal MIMIC-CXR test set, keep the prior image fixed, and substitute a randomly selected prior report from another patient; if the CE F1 does not drop, or if the gate output $F_{\\mathrm{dy}}$ almost never has near-zero norm on such mismatched inputs, then the prior-report alignment is not the source of the claimed gain.","tokens_in":21778,"feed_emoji":"🩻","tokens_out":8790,"duration_ms":80241,"temperature":0.7,"pith_summary":"The paper sets out to show that longitudinal radiology report generation—producing the current report while referring to a patient's prior chest X-ray and prior report—should not treat the prior as a late-stage add-on. It proposes DDaTR, an encoder that at every visual stage aligns the prior image with the prior report text and then fuses the difference between prior and current image features into the current representation, with temporal information flowing only from prior to current. The authors report that this design reaches a macro clinical-efficacy F1 of 0.441 on MIMIC-CXR versus 0.389 for the PromptMRG baseline, a micro F1 of 0.527 on Longitudinal-MIMIC versus 0.492, and a zero-shot F1 of 0.262 on IU-Xray versus 0.246. If the claim holds, automated reporting can track real changes—resolved effusions, growing opacities, stable cardiomegaly—rather than merely describing a single snapshot.","feed_headline":"Stage-by-stage prior comparison lifts chest-report accuracy","feed_subtitle":"DDaTR aligns prior reports and images at every encoder stage, lifting clinical-efficacy F1 to 0.441 on MIMIC-CXR.","key_machinery":"The load-bearing object is the DDaTR encoder: a multi-stage vision backbone with two lightweight modules inserted after every stage. The Dynamic Feature Alignment Module (DFAM) uses the frozen pretrained text encoder's features of the prior report as keys and values in cross-attention with the prior image features as queries, then gates the aligned feature with a tanh-activated 1x1 convolution, so clinical semantics are injected at pixel level. The Dynamic Difference-aware Module (DDAM) applies learnable descriptive convolutions to both prior and current features, forms a pixel-difference map $F_{\\mathrm{pd}}=\\mathrm{Sigmoid}(\\mathrm{AvgPool}(F_{\\mathrm{enc}}-F_{\\mathrm{enp}}))$, amplifies both features by it, concatenates them, and fuses the result through the same gating structure. The final update $\\hat{F}_{\\mathrm{cur}}=F_{\\mathrm{cur}}+\\alpha(F_{\\mathrm{dy}}\\odot F_{\\mathrm{da}})$ with $\\alpha\\in\\{0,1\\}$ makes the prior contribution optional and controlled, and a classifier on the prior features acts as a regularizer to keep the two streams semantically consistent.","core_discovery":"DDaTR's central claim is that the temporal and spatial correlations needed for longitudinal reporting should be modeled inside the visual encoder, stage by stage, rather than by late fusion of independently encoded images. The paper argues that prior reports carry the clinician's intent and should be injected into the prior image features at every stage through the Dynamic Feature Alignment Module, and that the difference between the prior and current image features, after texture enhancement, should be computed, gated, and added to the current stream through the Dynamic Difference-aware Module. A residual path transmits this longitudinal information only forward, from prior to current. The authors report that this design achieves macro CE F1 of 0.441 on MIMIC-CXR versus 0.389 for the PromptMRG baseline, micro CE F1 of 0.527 on Longitudinal-MIMIC versus 0.492, and zero-shot CE F1 of 0.262 on IU-Xray versus 0.246, with the MIMIC-CXR improvement having a reported p-value of 0.0415.","pith_inferences":["The gating weights $F_{\\mathrm{dy}}$ could be read as an empirical measure of prior relevance; correlating them with time interval, view mismatch, or actual change would test whether the network learns a clinically meaningful 'when to ignore the prior' policy.","The zero-shot IU-Xray gain suggests that training on longitudinal pairs may teach a transferable change-detection representation that helps even single-image reporting, a benefit the paper does not explicitly claim.","A natural next experiment is to feed more than one prior study or to use structured anatomical knowledge instead of free-text prior reports; the paper tests neither, so whether the mechanism scales to longer timelines is open."],"forward_implications":["On MIMIC-CXR, the reported macro CE F1 of 0.441 (against 0.389 for the PromptMRG baseline, with $p=0.0415$) implies that adding gated, per-stage prior information can improve diagnostic accuracy without changing the decoder.","The consistent gains across SwinT-B, ResNet101, and EfficientNetV2 imply the DFAM and DDAM modules are backbone-agnostic and can be plugged into any multi-scale visual encoder.","The larger CE gain relative to NLG gain implies the prior report contributes clinical knowledge rather than linguistic style, so evaluations of longitudinal methods should weight clinical-efficacy metrics heavily.","The zero-shot result on IU-Xray implies the model does not need longitudinal training data to benefit from the learned change-aware representation in single-image settings."],"supporting_citations":[{"why":"Supplies the MIMIC-CXR dataset and official split used for training, validation, and longitudinal re-organization.","marker":"[23]"},{"why":"Defines the LRRG setting and the macro CE evaluation protocol for MIMIC-CXR that DDaTR adopts and reports against.","marker":"[8]"},{"why":"Provides the PromptMRG baseline framework, including the disease classification branch and diagnosis-driven prompts that DDaTR extends.","marker":"[21]"},{"why":"Supplies the Longitudinal-MIMIC dataset and the micro CE scoring used for the longitudinal benchmark.","marker":"[3]"},{"why":"CheXbert converts reports into the 14 disease-label attributes that drive both the CE metrics and the classification branch.","marker":"[32]"},{"why":"BioViL-T serves as the specialist LRRG visual encoder compared in the experiments with varying proportions of longitudinal cases.","marker":"[6]"}],"fun_headline_variants":["Stage-wise prior alignment and difference gating lift chest-report accuracy","Align prior at each encoder stage, then gate the difference for reports","Per-stage prior injection and dynamic difference gating improve RRG","DDaTR: Dynamic difference-aware residual net beats late fusion on MIMIC","Stage-level prior fusion and residual gating boost longitudinal RRG"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole design assumes that the most recent prior study, once aligned by the report text and passed through the learned gate, adds useful information for every current report; if the gate cannot suppress an irrelevant or misleading prior, the added features can actively hurt the report.","fun_headline_variants_meta":{"raw":{"variants":["Stage-wise prior alignment and difference gating lift chest-report accuracy","Align prior at each encoder stage, then gate the difference for reports","Per-stage prior injection and dynamic difference gating improve RRG","DDaTR: Dynamic difference-aware residual net beats late fusion on MIMIC","Stage-level prior fusion and residual gating boost longitudinal RRG"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000361,"raw_usage":{"total_tokens":1989,"prompt_tokens":1022,"completion_tokens":967,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":876}},"tokens_in":638,"tokens_out":967,"duration_ms":9407,"temperature":1.0,"reasoning_tokens":876,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:52:26.828426+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the longitudinal MIMIC-CXR test set, keep the prior image fixed, and substitute a randomly selected prior report from another patient; if the CE F1 does not drop, or if the gate output $F_{\\mathrm{dy}}$ almost never has near-zero norm on such mismatched inputs, then the prior-report alignment is not the source of the claimed gain.","supporting_citations":[{"cited_title":"Recap: Towards precise radiology report generation via dynamic disease progression reasoning,","cited_arxiv_id":null,"evidence_quote":"Defines the LRRG setting and the macro CE evaluation protocol for MIMIC-CXR that DDaTR adopts and reports against."},{"cited_title":"Promptmrg: Diagnosis-driven prompts for medical report generation,","cited_arxiv_id":null,"evidence_quote":"Provides the PromptMRG baseline framework, including the disease classification branch and diagnosis-driven prompts that DDaTR extends."},{"cited_title":"Utilizing longitudinal chest x-rays and reports to pre-fill radiology reports,","cited_arxiv_id":null,"evidence_quote":"Supplies the Longitudinal-MIMIC dataset and the micro CE scoring used for the longitudinal benchmark."},{"cited_title":"Combining automatic labelers and expert annotations for accurate radiology report labeling using bert,","cited_arxiv_id":null,"evidence_quote":"CheXbert converts reports into the 14 disease-label attributes that drive both the CE metrics and the classification branch."},{"cited_title":"Learning to exploit temporal structure for biomedical vision-language processing,","cited_arxiv_id":null,"evidence_quote":"BioViL-T serves as the specialist LRRG visual encoder compared in the experiments with varying proportions of longitudinal cases."}],"review_version":1}