{"id":"98c4ce29-99d4-44ca-82b9-657b496705a9","arxiv_id":"2411.15539","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Reg2RG grounds each generated CT report section in a specific anatomical region by combining decoupled local texture and geometry features with global volume features in an LLM, outperforming prior methods on two chest CT datasets.","lead":"This paper introduces Reg2RG, a system that uses anatomical segmentation masks to extract region-level features from chest CT scans and feeds them to a large language model to generate radiology reports. It reports better natural language and clinical efficacy scores than several existing CT report generators on two datasets, with each report section linked to the anatomical region it describes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on an incomplete comparison: Dia-LLaMA and HILT are cited but not evaluated, and no baseline receives the same mask inputs, so the reported gains may reflect the extra segmentation supervision rather than the proposed LFD/RRA mechanism.","rationale":"The reader's weakest_assumption (mask accuracy) is a legitimate limitation that the paper acknowledges, and it is relevant to the interpretability claim. However, it is not the most load-bearing threat to the central claim: even with imperfect masks, Reg2RG still outperforms the compared baselines, so the empirical superiority claim can survive mask errors. What would falsify the central claim is a fairer comparison. The paper cites Dia-LLaMA and HILT as recent 3D CT-report methods but omits them from experiments, and no baseline uses the same segmentation masks. Since the geometry features are literally mask encodings and Table VI shows they contribute most of the CE gain, the reported superiority may be due to the additional input modality rather than the proposed LFD/RRA contributions. This is a standard internal-validity threat that a conditional acceptance should require the authors to address. Thus the verdict remains CONDITIONAL, with the condition expanded to include the omitted baselines and a mask-augmented control.","tokens_in":17572,"tokens_out":12618,"duration_ms":117855,"concrete_test":"Rerun Tables I and II with Dia-LLaMA and HILT fine-tuned under the same protocol (LLaMA2-7B decoder, 256x256x64 inputs, same data splits), and also fine-tune a strong global-feature baseline (e.g., RadFM) with the same SAT masks concatenated as an extra input channel or as an additional prompt token. If Reg2RG still beats both the omitted baselines and the mask-augmented RadFM on CE F1 and BLEU-4/ROUGE-L, the SOTA claim holds. If the mask-augmented baseline matches Reg2RG, the improvement is driven by the mask input, not the proposed LFD/RRA design.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of SOTA superiority is not yet established because the most relevant CT-report baselines are absent and the input comparison is asymmetric. The related-work section cites Dia-LLaMA [20] and HILT [9] as recent 3D CT report generation methods, but neither appears in Table I or Table II. HILT in particular targets high-resolution 3D volume encoding, which is directly related to the LFD claim of preserving local high-resolution details. More importantly, Reg2RG is the only method that receives segmentation masks: the geometry features Lg_Aj in Eq. (4) are pure mask encodings, and the texture features in Eq. (3) are mask-cropped volumes. No baseline is given access to the same masks. Table VI shows that adding geometry features raises CE F1 from 0.141 to 0.210, and the full model reaches 0.253, so a substantial part of the gain is attributable to the mask input itself rather than to the proposed region-report alignment. Without a mask-augmented global-feature baseline, the reported NLG/CE improvements could simply reflect the extra supervisory signal from segmentation maps, not the referring-and-grounding mechanism. The paper's own limitation section admits mask quality affects performance, but it does not address this comparison asymmetry.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Reg2RG, a region-guided framework for CT report generation. It uses a universal segmentation module to extract masks for anatomical regions, decouples local features into texture (mask-cropped, high-resolution) and geometry (uncropped mask encoding), and injects them alongside global features into an LLM decoder. A region-report alignment (RRA) training strategy adds a prefix predicting the region name before each region-specific report and shuffles the order of local features during training. Experiments on RadGenome-ChestCT and CTRG-Chest-548K report NLG and clinical-efficacy metrics, ablations, region recognition, and qualitative examples that support the claim that region-level grounding improves diagnostic accuracy.","tokens_in":17798,"tokens_out":5585,"duration_ms":45116,"significance":"If the results are reproducible, the paper makes a useful contribution by demonstrating that region-level features can improve CT report generation over global-only baselines, and the RRA strategy provides a concrete type of interpretability. The code release and use of public datasets are strengths. However, the significance is currently limited because the most relevant 3D CT baselines are not included in the comparison, the comparison to baselines is asymmetric with respect to mask supervision, and no statistical significance is reported. The interpretability claim is also weakened by low region-recognition accuracy for lung and pleura.","major_comments":[{"comment":"The SOTA comparison omits Dia-LLaMA [20] and HILT [9], both described in Section II-A as recent 3D CT report generation methods. Since HILT in particular targets high-resolution volume encoding and is directly related to the LFD claim, its absence means the 'outperforms several state-of-the-art methods' claim in the Abstract is not yet established against the most relevant baselines. Please include these models or provide a justified reason for their exclusion.","section":"Section IV-C, Tables I and II"},{"comment":"The ablation does not control for the additional mask supervision. Settings (a)-(c) all use mask-derived local features, while the baseline (RadFM) receives only the global volume. The geometry features Lg_Aj in Eq. (4) are pure mask encodings, and the texture features in Eq. (3) are mask-cropped volumes. A simple baseline that injects the segmentation masks (or a global feature concatenated with a mask summary) into the same LLM decoder is needed to attribute the gains to the LFD/RRA mechanisms rather than to the extra mask information.","section":"Section IV-D, Table VI"},{"comment":"No error bars, multiple seeds, or significance tests are reported. On CTRG-Chest-548K, the differences over the second-best method are small (e.g., BLEU-4 32.04 vs 30.89 for RadFM) and ROUGE-L is actually lower than several baselines (47.76 vs 50.93 for R2GenGPT). The paper should report variance across at least 3 seeds or a significance test for the headline claims.","section":"Section IV-C, Tables I and II"},{"comment":"The region recognition F1 for lung and pleura is about 0.44, yet Table IV shows lung is among the most frequently reported regions (1514 of 1564 test reports). This means the claimed interpretability via grounding is unreliable for a major organ. The paper should quantify how often reports for lung/pleura are grounded in the correct region, and discuss whether the interpretability benefit actually holds for these common regions.","section":"Section IV-C.3 and Table III"}],"minor_comments":[{"comment":"The dataset name is inconsistent: Section IV-A refers to 'CTRG-Chest-548K' while Section IV-B, Table I, and other places use 'CTRG-Chest-584K'; please unify.","section":"Section IV-A/B and Table I"},{"comment":"The text contains a placeholder 'Toy firgure' near the abstract/first page; this should be removed.","section":"Figure 1 caption area"},{"comment":"The captions contain the typo 'Metrcis' instead of 'Metrics'; please correct.","section":"Tables V and VI captions"},{"comment":"The acronym 'RRA' is used in the abstract but is not defined until Section III-D; define it at first use.","section":"Abstract and Section III-D"},{"comment":"The text cites Qwen2.5-14B but reference [33] is titled 'Qwen2 technical report'; please update the citation to the correct Qwen2.5 technical report or correct the model name.","section":"Reference [33]"},{"comment":"Figure 2 is dense and the shuffle mechanism is hard to follow; a higher-resolution figure or a step-by-step schematic would improve readability.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical system paper, but the SOTA claim is currently overreaching because the most relevant 3D CT baselines are missing and the comparison is asymmetric with respect to mask supervision. If the authors add Dia-LLaMA and HILT, include a mask-conditioned baseline, and report variance, the paper would be suitable for IEEE TMI. The novelty claim 'first referring and grounding for CT report generation' appears defensible given that RGRG is for CXR and MedRegA uses bounding boxes rather than region features, but the authors should still position this carefully in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: useful engineering for CT report generation, but the headline SOTA claim rests on an asymmetric comparison that should be fixed before the paper is taken at face value.\n\nThe genuinely new piece is the application of region-level referring and grounding to 3D CT report generation: extracting anatomical masks, decoupling texture from geometry, and training the model to first recognize the region and then generate that region's report. The RRA prefix trick gives a checkable link between a report and its region, which global-only baselines lack. The ablations are directionally consistent, and the LFD idea of cropping the masked volume to keep resolution while encoding the uncropped mask for geometry is sensible and cheap. That is honest, competent work.\n\nNow the soft spots, in proportion. The comparison is incomplete in a way that matters: Dia-LLaMA and HILT are cited as recent 3D CT report generators but never evaluated. HILT is specifically about high-resolution 3D encoding—exactly what LFD claims to address—so its absence distinctly weakens the 'outperforms several SOTA' claim. More important, every baseline gets only global features; Reg2RG is the only method with segmentation masks. The ablations show that adding geometry (the mask encoding) alone raises CE F1 from 0.141 to 0.210, so a large share of the gain could be the mask supervision rather than LFD or RRA. A mask-augmented global-feature baseline would settle this, and its absence is a real hole. Secondary issues: no error bars or seeds, so single-run differences of 1–3 BLEU points are within noise; the dataset name flips between CTRG-Chest-548K and 584K; and lung/pleura region recognition at ~0.44 F1 is acknowledged but still undercuts the interpretability story for two of ten regions. The paper's own limitation section admits mask sensitivity, which is honest.\n\nFor someone working on CT report generation or grounded medical VLMs, this is a useful data point and a reasonable starting point. I'd send it to peer review—a good referee will ask for the missing baselines and a mask-augmented comparison, and the revision would materially improve the paper. I would not cite it as a SOTA reference until that comparison exists.","headline":"Useful region-guided CT report generator, but the SOTA claim is undercut by an asymmetric baseline comparison that misses the most relevant 3D methods.","tokens_in":18365,"tokens_out":3612,"would_cite":false,"duration_ms":33203,"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":"A region-guided framework for CT report generation claims to outperform global-feature-only methods by grounding each report section in the anatomical region it describes.","keywords":["CT report generation","referring and grounding","region-level understanding","chest CT","large language model","local feature decoupling","region-report alignment","interpretability"],"falsifier":"Corrupt the input masks on a held-out set—swap the lung and pleura masks, erode or dilate them, or replace them with oracle ground-truth masks—and measure region recognition accuracy and clinical F1; if performance does not degrade substantially, the mask dependency is weaker than claimed, and if oracle masks do not improve results, the local features are not carrying the diagnostic signal.","tokens_in":1553,"feed_emoji":"🩻","tokens_out":2137,"duration_ms":66923,"temperature":0.7,"pith_summary":"This paper proposes Reg2RG, a framework for chest CT report generation that focuses on individual anatomical regions rather than only the whole volume. It uses a universal segmentation module to carve out each region, decouples the region's texture from its geometry, and feeds both, together with global context, into a large language model. A region-report alignment training step makes the model first name the region it is looking at and then write that region's findings, so each report section is explicitly grounded in the image region it describes. The paper reports that this region-guided design outperforms global-feature-only baselines on standard language metrics and on abnormality-detection precision, recall, and F1, with the grounding step doubling as an interpretability check.","feed_headline":"Region-guided CT reports beat global-only methods on clinical accuracy","feed_subtitle":"The framework grounds each report section in the anatomical region it describes, using an LLM as decoder.","key_machinery":"The central object is the decoupled region feature pair: texture features $L^t_{A_j}=f_A(f_V(\\mathrm{Crop}(M_{A_j}\\odot V)))$ and geometry features $L^g_{A_j}=f_P(f_M(M_{A_j}))$, concatenated into $L_{A_j}$. The mask $M_{A_j}$ from the universal segmentation module defines the referring region; texture preserves high-resolution details of the cropped masked volume, and geometry preserves the original size and position from the uncropped mask. These local features are shuffled and injected into the LLM prompt with a recognition prefix, so the model must identify the region before writing its report. The region-report alignment (RRA) training objective is the language-modeling loss over this restructured sequence, and this is what carries the claimed referring and grounding behavior.","core_discovery":"Reg2RG's central claim is that CT report generation should be treated as a referring-and-grounding task over anatomical regions, not a volume-to-text mapping. The paper argues that global volume features alone cause decoders to miss abnormalities, and that region-level local features—texture from cropped masked volumes plus geometry from the uncropped mask—give the model the information needed to diagnose each region. The region-report alignment strategy forces the LLM to recognize the anatomical area from the local features before generating that region's report, and shuffling region order prevents the model from relying on a fixed sequence. On two large chest CT datasets, the paper reports that Reg2RG outperforms prior 3D and 2D methods on BLEU, METEOR, and ROUGE-L, and on the region-grounded dataset it also achieves higher precision, recall, and F1 for abnormality labels, with region recognition serving as an early-stage validation of which anatomy each report is grounded in.","pith_inferences":["If masks are the bottleneck, improving the segmentation module or learning region proposals jointly should directly raise region recognition and clinical F1, especially for lung and pleura where the paper's own Table III shows F1 around 0.44.","The shuffling mechanism implies the model learns region identity from local features rather than from report text, so the approach could transfer to datasets without region-grounded reports by using a text-to-region splitter to create supervision.","A natural extension is to move from organ-level regions to lesion-level regions; the paper's limitation section notes that organ-level features miss or mislocate lesions, making lesion segmentation the next testbed.","If clinical efficacy gains persist under mask corruption, the framework would be robust enough for deployment with imperfect universal segmenters; if not, mask quality becomes the main cost driver of the whole pipeline."],"forward_implications":["If the central claim holds, CT report generators no longer need to encode every diagnostic cue into a single global embedding; region-level detail can be delegated to local feature extractors.","Region recognition before report generation gives a built-in check: a report section is only trustworthy if the model correctly names the region it refers to.","The decoupling strategy shows that high-resolution texture and mask-derived geometry can be preserved without a large increase in compute, making region-level report generation practical for 3D volumes.","Explicit grounding of each finding to an anatomical region makes the generated reports more useful to radiologists, because the visual basis of each statement can be inspected.","The reported gains in clinical efficacy suggest that region-guided features matter for abnormality detection even when the language decoder is held fixed."],"supporting_citations":[{"why":"Supplies the region-grounded chest CT report dataset with anatomical masks used for training and evaluation.","marker":"[16]"},{"why":"Universal segmentation module whose masks define the referring regions from which local texture and geometry features are computed.","marker":"[6]"},{"why":"Provides the pretrained volume encoder and adapter used to extract global and local texture features, and also serves as a baseline.","marker":"[4]"},{"why":"Baseline global-feature 3D CT report generator whose volume-to-text approach the paper argues misses region-level abnormalities.","marker":"[3]"},{"why":"2D baseline that achieves some of the strongest comparison results and is used in the NLG and clinical efficacy tables.","marker":"[22]"},{"why":"Supplies the chest CT dataset and the text classifier used to extract abnormality labels for clinical efficacy evaluation.","marker":"[15]"},{"why":"Large language model used to split the second dataset's reports into region-level sections, creating supervision for the region-guided training.","marker":"[33]"},{"why":"The large language model decoder that generates the region-specific reports and whose in-context referencing ability the framework relies on.","marker":"[38]"}],"fun_headline_variants":["CT reports that see region-by-region outperform global scans","Grounding each CT region before writing its report boosts accuracy","Anatomy-grounded CT reporting: region-level features beat whole-volume","For CT reports, region grounding beats global context alone"],"cache_read_input_tokens":20480,"weakest_assumption_plain":"The load-bearing premise is that the segmentation module's masks are accurate enough to define the regions that drive diagnosis, because every local feature is built by masking and cropping the volume; a wrong or incomplete mask grounds the report in the wrong anatomy and the claimed interpretability collapses.","fun_headline_variants_meta":{"raw":{"variants":["CT reports that see region-by-region outperform global scans","Grounding each CT region before writing its report boosts accuracy","Anatomy-grounded CT reporting: region-level features beat whole-volume","For CT reports, region grounding beats global context alone"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000706,"raw_usage":{"total_tokens":3210,"prompt_tokens":1004,"completion_tokens":2206,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":2138}},"tokens_in":620,"tokens_out":2206,"duration_ms":13164,"temperature":1.0,"reasoning_tokens":2138,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:10:33.330257+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Corrupt the input masks on a held-out set—swap the lung and pleura masks, erode or dilate them, or replace them with oracle ground-truth masks—and measure region recognition accuracy and clinical F1; if performance does not degrade substantially, the mask dependency is weaker than claimed, and if oracle masks do not improve results, the local features are not carrying the diagnostic signal.","supporting_citations":[],"review_version":1}