{"id":"770521a4-7bbe-46e0-b80c-0cd51dca4764","arxiv_id":"2507.03908","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Optimal Transport alignment between image patches and disease labels, paired with LLM fine-tuning, improves clinical efficacy of generated radiology reports.","lead":"A new framework, OTDRG, uses Optimal Transport to align X-ray image features with disease labels extracted from paired reports, then fine-tunes an LLM with LoRA to generate radiology reports. On MIMIC-CXR and IU X-Ray, it reports competitive language metrics and improved clinical efficacy scores, though the state-of-the-art claim is undercut by a lower CE score than PromptMRG.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported end-to-end clinical gains depend on a disease-label predictor (Sec. 3.4) whose accuracy is never evaluated; because training uses ground-truth labels but inference uses predicted labels, the CE and SOTA claims are unsupported until that module is validated.","rationale":"The reader's weakest assumption is precisely the load-bearing point. The proposed method's novelty is the FIOT alignment, but the reported end-to-end performance cannot be attributed to FIOT unless the disease-label predictor used at inference is reliable. Training conditions on labels extracted from ground-truth reports; testing conditions on labels predicted from images. That train/test mismatch is a real correctness risk, not a matter of research taste. The paper provides independent support for parts of the pipeline: the ablation Table 3 shows monotone CE gains as modules are added, and Figure 4 and Table 4 show sensible sensitivity to OT hyperparameters. But those ablations use oracle labels, so they do not validate the deployed system. The absence of code and the missing Supplementary section that was promised for label details further weaken verification. I did not find a more fundamental flaw: the OT formulation follows the standard Sinkhorn-Knopp update despite a sign typo in Eq. 8, and the comparison tables are internally consistent enough to see what was measured. The abstract overstates SOTA, but that is a calibration issue, not a reason to reject the method; the central empirical claim cannot be settled without the label-predictor evaluation. Hence the verdict remains CONDITIONAL, and the reader's concern is confirmed rather than changed.","tokens_in":16120,"tokens_out":4205,"duration_ms":48405,"concrete_test":"Evaluate the Sec. 3.4 predictor on the MIMIC-CXR test split against CheXbert labels, reporting per-disease and macro F1. Then run the full OTDRG pipeline under three conditions: (a) predicted labels, (b) ground-truth CheXbert labels, and (c) no disease-label conditioning. If condition (a) falls below (b) by more than the reported margin over the strongest baseline, or if (c) matches (a), the label predictor rather than FIOT is carrying the clinical gain. Publish these end-to-end CE numbers alongside the oracle-label ablation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that OTDRG materially improves clinical accuracy depends on the disease label prediction module introduced in Sec. 3.4. At validation and test time, predicted one-hot labels replace the true labels extracted during training, as stated in Sec. 3.1 and Sec. 3.4. No accuracy, precision, recall, or error analysis for this classifier is reported anywhere; the DLE ablation in Table 3 uses labels extracted from ground-truth reports, not predicted labels. The end-to-end numbers in Tables 1 and 4 therefore confound the FIOT/LLM components with this unvalidated classifier. If the predictor is noisy, the LLM is conditioned on inputs that differ from those used in fine-tuning, and the CE improvement from 0.263 to 0.436 in Table 3 may not transfer. The paper even says in Sec. 3.4 that accurate disease status labels are essential and defers details to a Supplementary that is not present. Separately, the abstract's 'state-of-the-art in both NLG and CE' is contradicted by Table 1, where PromptMRG has CE F1 0.476 vs 0.436, and by Table 2, where two models beat the reported ROUGE-L; Sec. 4.3.1 itself concedes being 'second only to PromptMRG' in CE. The decisive missing evidence is a direct evaluation of the inference-time label predictor and its effect on end-to-end report quality.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes OTDRG, a radiology report generation framework that uses Optimal Transport (OT) to align X-ray image features with disease-label features extracted from ground-truth reports, then fine-tunes an LLM with LoRA on the aligned features to generate reports. A separate disease-label prediction module is trained to predict one-hot disease states from images at inference time. The authors evaluate on MIMIC-CXR and IU X-Ray, reporting competitive or state-of-the-art NLG and clinical efficacy (CE) scores, with ablations showing gains from the IRCP contrastive pre-training, disease-label extraction (DLE), and the FIOT alignment module.","tokens_in":16450,"tokens_out":4005,"duration_ms":43186,"significance":"If the results hold, the use of OT for cross-modal alignment in LLM-based report generation is a reasonable and potentially useful idea, and the authors provide a fairly complete pipeline with ablations and qualitative visualizations. However, the claimed state-of-the-art status is not supported by the numbers in Table 1, and the inference-time disease predictor is never evaluated. The central contribution—that FIOT materially improves clinical accuracy—is plausible but not yet convincingly established because the end-to-end evaluation relies on an unvalidated component.","major_comments":[{"comment":"The abstract claims 'state-of-the-art performance in both natural language generation (NLG) and clinical efficacy (CE) metrics,' but Table 1 shows PromptMRG achieving CE F1 0.476 versus OTDRG 0.436, and Section 4.3.1 itself concedes being 'second only to PromptMRG in CE metrics.' Similarly, on IU X-Ray (Table 2), OTDRG does not achieve the best ROUGE-L (R2GenGPT 0.438, MCSAM 0.394, CLR2Gen 0.378 vs. 0.377). The SOTA claim must be revised to reflect these results, or the comparison must be restricted to LLM-based methods and clearly stated as such.","section":"Abstract and Section 4.3.1, Table 1"},{"comment":"The disease-label prediction module is introduced as the method used to provide disease labels during validation and testing, replacing the ground-truth-extracted labels used in training. However, the paper never reports any accuracy, precision, recall, or error analysis for this module. The ablation study in Table 3 uses the DLE module with labels extracted from ground-truth reports, not predicted labels, so the end-to-end results in Tables 1 and 4 confound the FIOT and LLM contributions with the unvalidated predictor. The statement in Section 3.4 that 'accurate extraction of disease status labels is essential (more details can be found in the Supplementary)' is not verifiable because the Supplementary is not provided. The authors should either evaluate the predictor directly or report end-to-end results using predicted labels.","section":"Section 3.4"},{"comment":"The ablation shows that adding FIOT to IRCP+DLE increases CE F1 from 0.349 to 0.436 but decreases ROUGE-L from 0.301 to 0.295. This regression is not discussed in the text. While the CE gain is the central motivation, the unexplained NLG degradation should be addressed, especially since the paper claims improvements in both aspects. A short explanation or additional analysis (e.g., whether the FIOT module trades off language diversity for clinical precision) would strengthen the ablation.","section":"Table 3, last two rows"},{"comment":"All baseline numbers are taken from the original papers, with no error bars, no statistical significance tests, and no indication of whether the same preprocessing, split, or evaluation code was used. Given that the paper's central claim is state-of-the-art performance, at least the most important comparisons (e.g., PromptMRG on MIMIC-CXR) should be re-run under a unified protocol, or the claim should be weakened to 'competitive' rather than 'state-of-the-art.'","section":"Tables 1 and 2, experimental setup"}],"minor_comments":[{"comment":"The text defines sim(·,·) as a similarity function and gives Euclidean distance as an example. Euclidean distance is a dissimilarity measure; if used directly in the InfoNCE loss it would push positive pairs apart. Please clarify whether the implementation uses negative Euclidean distance or cosine similarity.","section":"Equation (2), Section 3.2.2"},{"comment":"The sign convention in the entropic OT objective is confusing: Ω(T) is defined as Σ T log T, which is the negative entropy, and the objective is written as minimize ⟨T,C⟩ − εΩ(T). This is mathematically equivalent to entropy regularization, but the notation should be aligned with the Sinkhorn-Knopp derivation. In Algorithm 1, the initialization of α and β, the use of softmax on α and β, and the final 'return dOT = minimize <T,C> − h(T*)' are unclear and should be rewritten.","section":"Equation (8) and Algorithm 1"},{"comment":"The sentence 'more details can be found in the Supplementary' is unverifiable because the arXiv submission does not appear to include a supplementary document. Either include the supplementary material or remove the reference.","section":"Section 3.4"},{"comment":"The conclusion contains a typo: 'MIME-CXR' should be 'MIMIC-CXR.'","section":"Section 5 (Conclusion)"},{"comment":"The list of compared methods in the text includes '[46]' and 'Clinical-Bert [45]', but the reference list numbering does not cleanly map to the cited works (e.g., [46] is given as 'Radiology report generation with a learned knowledge base' while the text also cites [47] for the same topic). Please check that all inline citations match the reference list.","section":"Section 4.3.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's main selling point—state-of-the-art clinical efficacy—is undermined by (a) the comparison table where PromptMRG has higher CE F1, and (b) the lack of any evaluation of the inference-time disease predictor. These are not merely presentational issues; they affect the central claim. The paper may be salvageable with a major revision that either validates the predictor or restricts claims to the LLM-based setting, and that repositions the contribution as 'competitive with current methods' rather than 'SOTA.' I would recommend against acceptance in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is a reasonable engineering contribution: use Optimal Transport to align image patch features with disease-label embeddings before fine-tuning an LLM for report generation. That combination is new for this task, and the ablation study gives some evidence that each added component helps clinical efficacy metrics. The paper is also honest in one place (Sec. 4.3.1) that its CE performance is second only to PromptMRG, and the visualization of transport-plan heatmaps is a nice touch.\n\nThe soft spots are real and the stress-test note lands. The biggest issue is the disease prediction module in Sec. 3.4. Training uses ground-truth labels extracted from reports, but inference uses predicted labels from the X-ray. The accuracy of that classifier is never reported, so the end-to-end CE and NLG numbers in Tables 1 and 4 confound the FIOT/LLM components with an unvalidated predictor. If the predictor is noisy, the LLM is conditioned on inputs it never saw during fine-tuning, and the reported gains may not transfer. This is a load-bearing gap, not a cosmetic omission.\n\nThe SOTA claim in the abstract is also overstated, as Table 1 shows PromptMRG has higher CE F1 (0.476 vs. 0.436). The paper even concedes being second in the comparison section, so the abstract should be corrected. Comparisons also lack error bars and are taken from original papers, which is standard but weakens any strong SOTA claim. There is a small ROUGE-L regression when FIOT is added (0.301 to 0.295); it is minor but worth acknowledging. Missing experimental details (hyperparameters, code, the referenced Supplementary) would need to be supplied.\n\nThat said, the paper does not have a circular or incoherent argument. The OT loss is supervised by the ground-truth report and labels; the disease-label extraction is a standard auxiliary task, not a hidden fit to the metric. The central flaw is missing evidence, not bad reasoning.\n\nFor whom: researchers working on LLM-based radiology report generation who want to see another alignment strategy and an ablation methodology. It is not a breakthrough, but it is a legitimate incremental approach that a serious referee could help bring into shape. My recommendation: send it to peer review, but require an evaluation of the disease predictor and a corrected SOTA statement before acceptance.","headline":"Plausible incremental OT-based alignment for radiology report generation, but the unvalidated disease-label predictor and overstated SOTA claims make the main result only conditionally supported.","tokens_in":16977,"tokens_out":1177,"would_cite":false,"duration_ms":15885,"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":"The paper argues that aligning chest X-ray image features with disease labels through optimal transport, then fine-tuning an LLM with LoRA, produces radiology reports that are both more fluent and more clinically accurate than current…","keywords":["radiology report generation","optimal transport","large language models","cross-modal alignment","clinical efficacy","disease label prediction","Sinkhorn-Knopp","MIMIC-CXR"],"falsifier":"Take the MIMIC-CXR test set, run the disease prediction module on each image, and compare its predicted labels against CheXbert-extracted labels from the ground-truth reports. Then generate reports twice, once with predicted labels and once with ground-truth labels; if clinical efficacy scores drop substantially in the predicted-label condition, the test-time pipeline does not deliver the reported clinical accuracy.","tokens_in":15914,"feed_emoji":"🩻","tokens_out":7020,"duration_ms":64786,"temperature":0.7,"pith_summary":"Radiology report generation from chest X-rays faces two problems: LLMs write fluent prose that often misses or invents findings, and raw image features do not sit naturally in the text space where LLMs work. This paper proposes OTDRG, which extracts disease labels from reports, projects image patches and labels into the LLM embedding space, and uses optimal transport to find a global alignment between the two. The aligned features are combined into a prompt and the LLM is fine-tuned with LoRA to generate the report. The authors claim this improves both language quality and clinical accuracy over prior methods on MIMIC-CXR and IU X-Ray, with ablations showing the transport alignment module is the main driver of the clinical gains.","feed_headline":"Optimal transport sharpens AI-written radiology reports","feed_subtitle":"Aligning X-ray features to disease labels before LLM generation lifts both fluency and clinical accuracy.","key_machinery":"The load-bearing component is FIOT (Feature Integration by Optimal Transport). It builds a Euclidean cost matrix between projected image patch features and projected disease-label features, solves for an optimal transport plan using the Sinkhorn-Knopp algorithm with an entropy term, and returns the Sinkhorn distance as an auxiliary loss. The aligned features are concatenated into a prompt of the form `</IMG> $f_p(I) \\oplus f_p(L)$ </IMG>`, which is what the LLM is fine-tuned on via LoRA; the transport plan itself also supplies the heatmaps the paper uses to show that the model attends to disease-relevant image regions.","core_discovery":"OTDRG's central claim is that a cross-modal optimal transport plan between X-ray image patch features and disease-label features, computed in the LLM's embedding space, bridges the distribution gap that otherwise makes LLMs prioritize fluency over diagnostic content. The transport plan is obtained with the Sinkhorn-Knopp algorithm with entropic regularization, and the aligned features are fed into the LLM as an image-text integrated prompt along with the disease labels. At test time, a separately trained classifier predicts the disease labels from the image, replacing the ground-truth labels used during training. The authors report best-in-comparison NLG scores on both datasets and, on MIMIC-CXR, clinical efficacy precision/recall/F1 second only to PromptMRG, with the FIOT module credited for the largest part of the clinical improvement.","pith_inferences":["Because the disease prediction module is never evaluated on its own, the end-to-end results are conditional on it: a direct accuracy, precision, and recall report for predicted labels would determine how much of the clinical gain survives in real deployment.","The same label-conditioned optimal transport alignment could be tested on other structured clinical outputs, such as pathology or discharge summaries, where a small set of discrete findings plays the role of disease labels.","A cleaner attribution of the FIOT module's contribution would compare it against a control that concatenates the same disease labels without optimal transport; the current ablations do not include such a control."],"forward_implications":["LLM-based report generators can be steered toward clinical accuracy by conditioning generation on aligned disease labels rather than on raw image features alone.","On MIMIC-CXR, OTDRG reports the highest BLEU-1, BLEU-4, METEOR, and ROUGE-L among the compared methods and a clinical F1 of 0.436, behind only PromptMRG; on IU X-Ray it reports top BLEU-1 and METEOR.","The ablation sequence shows clinical F1 rising from 0.263 in the baseline to 0.436 for the full model, with the FIOT module adding the largest precision and recall gain.","The transport-plan heatmaps concentrate on disease-related areas of the X-ray, linking the alignment objective to the content of the generated report."],"supporting_citations":[{"why":"Supplies the 14-disease label vocabulary and the CheXbert labeler used to extract disease labels and to compute clinical efficacy metrics.","marker":"[34]"},{"why":"PromptMRG, the diagnosis-driven prompting method that is the strongest clinical-efficacy comparator and a source of the disease-label conditioning idea.","marker":"[16]"},{"why":"R2GenGPT, the frozen-LLM radiology report generation baseline that OTDRG builds on and compares against.","marker":"[43]"},{"why":"LoRA, the low-rank adaptation method used to fine-tune the LLM with few trainable parameters.","marker":"[14]"},{"why":"Sinkhorn-Knopp algorithm, the computational engine behind the entropic optimal transport plan in FIOT.","marker":"[33]"},{"why":"Swin Transformer, the visual encoder that produces the image patch features aligned to labels.","marker":"[25]"},{"why":"MIMIC-CXR, the main dataset for training and evaluating the model.","marker":"[17]"},{"why":"IU X-Ray, the second benchmark dataset used for cross-dataset evaluation.","marker":"[10]"},{"why":"Focal loss, the training objective for the disease label prediction module.","marker":"[23]"}],"fun_headline_variants":["Optimal transport aligns X-ray and text for better radiology reports","OTDRG uses optimal transport to boost clinical accuracy of LLM radiology reports","Cross-modal optimal transport aligns features for clinically accurate reports","Optimal transport bridging vision and language improves radiology report quality"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"At test time the generated reports depend on disease labels predicted from the X-ray by a small classifier, but the paper reports no accuracy for that classifier; if those predicted labels are unreliable, the claimed clinical gains may shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Optimal transport aligns X-ray and text for better radiology reports","OTDRG uses optimal transport to boost clinical accuracy of LLM radiology reports","Cross-modal optimal transport aligns features for clinically accurate reports","Optimal transport bridging vision and language improves radiology report quality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00092,"raw_usage":{"total_tokens":3941,"prompt_tokens":934,"completion_tokens":3007,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":2933}},"tokens_in":550,"tokens_out":3007,"duration_ms":20913,"temperature":1.0,"reasoning_tokens":2933,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:59:22.720379+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the MIMIC-CXR test set, run the disease prediction module on each image, and compare its predicted labels against CheXbert-extracted labels from the ground-truth reports. Then generate reports twice, once with predicted labels and once with ground-truth labels; if clinical efficacy scores drop substantially in the predicted-label condition, the test-time pipeline does not deliver the reported clinical accuracy.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the 14-disease label vocabulary and the CheXbert labeler used to extract disease labels and to compute clinical efficacy metrics."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PromptMRG, the diagnosis-driven prompting method that is the strongest clinical-efficacy comparator and a source of the disease-label conditioning idea."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Sinkhorn-Knopp algorithm, the computational engine behind the entropic optimal transport plan in FIOT."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Swin Transformer, the visual encoder that produces the image patch features aligned to labels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Focal loss, the training objective for the disease label prediction module."}],"review_version":1}