{"id":"f44c9bce-4468-421b-a1a6-35300e7fd06b","arxiv_id":"2411.18309","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A multi-view, retrieval-augmented Transformer with KAN layers reports state-of-the-art chest CT report generation on CTRG-Chest-548K, but evaluation weaknesses temper the result.","lead":"This paper presents MvKeTR, a Transformer system that generates chest CT reports by combining three anatomical views and retrieving similar past reports as medical knowledge. The authors report state-of-the-art scores on the CTRG-Chest-548K benchmark, but evaluation gaps (no error bars, test-set tuning, and mixed baseline sources) mean the gain is not firmly established.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on Table II, but the table mixes cited and re-run baselines without evidence of a shared test split or preprocessing; the implausible Reg2RG METEOR of 49.71 further indicates the comparison protocol needs verification.","rationale":"The paper's central claim is the empirical SOTA result in Table II. The reader's weakest assumption identifies the comparison protocol as the load-bearing premise; I agree. The mixed provenance of baselines is not merely a cosmetic issue: if the cited rows used the original split (which is random 80/20 in the SL-DG paper but may not be identical to the split used here), then the comparison is confounded. The re-run rows are more trustworthy as internal comparisons, but even those are not described with error bars or multiple seeds. The implausible Reg2RG METEOR strengthens the need to verify the table. That said, the paper has independent support: the ablations are internally consistent, the code is released, and the architecture is a plausible combination of known components. These issues are addressable, so a conditional verdict is appropriate. No single internal contradiction invalidates the method; the concern is about the evidence for the headline claim.","tokens_in":22921,"tokens_out":5062,"duration_ms":42500,"concrete_test":"Obtain the official test split used by SL-DG, Dia-LLaMA, and Reg2RG (e.g., from the original authors or code repositories) and rerun MvKeTR and all re-run baselines on exactly that split with the same preprocessing and CT-ViT extractor; if MvKeTR no longer exceeds the cited numbers on BLEU-4/ROUGE-L, the SOTA claim fails. Separately, recompute Reg2RG's METEOR with its public code, or ask the authors to confirm the value; if METEOR is not 49.71, correct Table II and the 'almost all metrics' wording.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table II footnotes mark SL-DG, Dia-LLaMA, and Reg2RG as cited from their original papers, while other baselines are re-run. The implementation section states that CT-ViT is used as the 3D visual extractor for both compared methods and MvKeTR, but for the cited rows the visual extractor and data split are whatever the original papers used. No evidence is provided that all rows share the same 80/20 split, preprocessing, or evaluation script. If the cited baselines were evaluated on a different test split, the claimed gains (e.g., BLEU-4 37.86 vs. 32.04 for Reg2RG) could be an artifact of split difficulty rather than model quality. Additionally, the Reg2RG METEOR of 49.71 is far outside the range of every other method (19.52-28.36) and inconsistent with its own BLEU/ROUGE levels, suggesting a data-entry error; if that value is wrong, the 'almost all metrics' conclusion is mis-stated and the table's reliability is called into question. The top-k=16 selection on Fig. 10, which plots test-set BLEU-1, is a further test-set tuning concern, but the core issue is that the comparison protocol is not demonstrated to be apples-to-apples.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MvKeTR, a Transformer-based model for chest CT report generation on the CTRG-Chest-548K dataset. The model has three main components: a Multi-View Perception Aggregator (MVPA) that fuses axial, sagittal, and coronal CT features extracted by three CT-ViT encoders using view-aware attention; a Cross-Modal Knowledge Enhancer (CMKE) that retrieves the top-k most similar reports from a CT-CLIP-encoded report bank and integrates them via cross-attention; and a report generator based on R2GenCMN with KAN layers replacing MLPs. The authors report state-of-the-art results across most automatic metrics (BLEU, METEOR, ROUGE-L), provide ablations showing the contribution of each component, include a human evaluation by two radiologists, and make the code publicly available.","tokens_in":23140,"tokens_out":4415,"duration_ms":41958,"significance":"If the empirical claims hold, MvKeTR is a plausible step forward for 3D CT report generation: the multi-view aggregation is well motivated by radiological practice, the retrieval-based knowledge enhancer is a clean way to inject domain knowledge, and the use of KANs is a timely architectural choice. The paper includes useful ablations, a public dataset, a public code link, and a human evaluation, which are all strengths. However, the central claim of surpassing prior state-of-the-art rests on a comparison table whose baselines have mixed provenance and at least one implausible entry, and the reported gains are not accompanied by variance or significance information. The architecture is coherent, but the empirical evidence needs substantial verification before the SOTA claim can be accepted.","major_comments":[{"comment":"The SOTA claim is not established because Table II mixes directly cited results (SL-DG, Dia-LLaMA, Reg2RG, marked with double-dagger) with results the authors re-ran using public code (marked with asterisk). Section IV-A states only that 80% of the data is randomly allocated to training and 20% to testing, and Section IV-C says CT-ViT is used as the visual extractor for both compared methods and MvKeTR. For the cited rows, however, the test split, preprocessing, and evaluation script are those of the original papers, and no evidence is given that they coincide with the authors' split. If the cited baselines were evaluated on a different test set, the claimed improvements could be an artifact of split difficulty. The authors should either re-run all baselines under the same protocol or clearly report the provenance and split for each row and restrict the SOTA claim to comparable settings.","section":"Section IV-A, IV-C, Table II"},{"comment":"The METEOR value of 49.71 for Reg2RG is far outside the range of all other methods (19.52-28.36) and is inconsistent with Reg2RG's own BLEU and ROUGE scores in the same row (BLEU-1 49.63, BLEU-4 32.04, ROUGE-L 47.76). This is likely a data-entry error or a metric-implementation mismatch. Because the abstract and Section IV-D claim that MvKeTR surpasses SOTA across 'almost all metrics,' and the only exception is this METEOR value, the correctness of this entry is load-bearing. The authors must verify the value against the original Reg2RG paper or re-run Reg2RG on their split, and correct the table and the surrounding text if needed.","section":"Table II, Reg2RG row"},{"comment":"The hyperparameter top-k is set to 16 because it peaks on the test set, as shown in Fig. 10, which plots BLEU-1 on CTRG-Chest-548K. Selecting a hyperparameter on the test set inflates the reported scores and makes the comparison with baselines unfair if those baselines did not receive the same test-set tuning. The authors should select top-k on a validation split and report the corresponding test performance, or at minimum provide a sensitivity analysis that distinguishes validation-based selection from test-set selection.","section":"Section IV-H2, Fig. 10"},{"comment":"All quantitative results are reported as single numbers without error bars, confidence intervals, or significance tests. Given that the dataset contains only 1,804 image-report pairs and the split is random, the differences between MvKeTR and strong baselines such as CAMANet could be within run-to-run variance. The authors should report results over multiple seeds with mean and standard deviation, and perform a significance test (e.g., paired bootstrap) for the main comparisons in Table II and the ablation comparisons in Table III.","section":"Section IV-B, IV-D, Tables II and III"},{"comment":"The 'AVG. Δ' values in Table III are not reproducible from the listed metric values. For Ours-MLP, the arithmetic mean of the six per-metric relative improvements over BASE is approximately 16.7%, not the reported 14.1%; for Ours, the arithmetic mean is approximately 22.5%, not 23.6%. The base row also gives inconsistent numbers for R2GenCMN (48.29/28.42 in Table III versus 28.42 in the text). The authors should state the averaging formula and correct the table, because the claimed 9.5% gap between Ours-MLP and Ours is used to support the KAN contribution.","section":"Table III, Ours-MLP row"},{"comment":"The human evaluation is based on only 10 randomly selected test cases and two radiologists, with no inter-rater agreement measure and no statistical test for the reported 'significant margins' and '32% improvement.' For a clinical-facing claim, this sample is too small and the scoring procedure is underspecified. The authors should either report per-case scores, inter-rater reliability (e.g., Cohen's kappa), and a paired significance test, or temper the clinical-utility claim accordingly.","section":"Section IV-F, Table IV"}],"minor_comments":[{"comment":"The sentence 'The 14.5% BLEU-4 degradation without CMKE confirms that retrieved knowledge primarily enhances model generalization' appears to misstate the ablation: BASE+CMKE improves BLEU-4 by about 21% over BASE (34.43 vs. 28.42), and 14.5% is the average improvement over all metrics. This should be reworded to avoid confusion.","section":"Section IV-G"},{"comment":"Equation (22) is called a 'Bayesian posterior estimation,' but the displayed factorization p(Y|X,K) ∝ p(X|Y)p(K|Y) is not a valid Bayesian posterior as written (the right-hand side conditions on Y and omits the prior and evidence terms). This is a narrative interpretation of feature concatenation rather than a derivation, and the framing should be removed or substantially revised.","section":"Section III-F, Eq. (22)"},{"comment":"The abstract contains the typo 'TansfoRmer,' and the paper refers to 'Vit-Transformer' in Section IV-D; these should be corrected for consistency.","section":"Abstract and Title"},{"comment":"Figures 8 and 10 contain font-encoding artifacts (unprintable glyph sequences) in the submitted PDF, making the axis labels and captions unreadable in places; the figures should be regenerated with proper fonts.","section":"Fig. 8 and Fig. 10"},{"comment":"The notation for CT patch extraction is inconsistent: the text says 'extracting (28)×(28)×(28) non-overlapping patches' but earlier defines patches Z_v as 8×8×8, and the relationship between the temporal/spatial patch sizes Pt, Ph, Pw and these numbers is not clarified. Please align the notation.","section":"Notation in Section III-B"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the architecture is interesting, but the comparison table needs careful verification before publication. I would ask the authors for the raw outputs of all re-run baselines, the exact Reg2RG METEOR value from the original paper or their own run, and the split/provenance of every row in Table II. The test-set selection of top-k and the absence of variance estimates should also be addressed in revision. The manuscript is not ready for acceptance in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the architecture is a genuinely new combination for 3D chest CT report generation: view-aware multi-view attention, cross-modal retrieval of similar reports, and KAN layers instead of MLPs. The ablations are internally consistent and show each component contributes. The authors also provide code and run a robustness study with rotations and artifacts, which is more than most papers in this subfield do. Second, the central claim that MvKeTR surpasses prior SOTA is not established by the evidence as presented.\n\nThe core problem is Table II. Rows marked with a double-dagger are cited from the original papers, while others are re-run with CT-ViT as the visual extractor. There is no evidence that all rows share the same 80/20 split, preprocessing, or evaluation script. For the cited baselines (SL-DG, Dia-LLaMA, Reg2RG), the extractor and split are whatever the original authors used. If those numbers came from a different test set, the claimed BLEU-4 gain of 37.86 vs. 32.04 for Reg2RG is not apples-to-apples. The Reg2RG METEOR of 49.71 is also far outside the range of every other method (19.52-28.36) and inconsistent with its own BLEU/ROUGE levels; that looks like a data-entry error. If it is wrong, the 'almost all metrics' conclusion is mis-stated and the whole table becomes suspect.\n\nThere are also smaller but real issues. Top-k=16 is selected by peaking on the test set (Fig. 10), which is a form of test-set tuning. No error bars or significance tests appear anywhere. The dataset is only 1,804 pairs, so variance matters. Equation (22) frames feature concatenation as Bayesian posterior estimation, but that is a narrative device, not a derivation. And the artifact experiments that show performance improving under motion and ring artifacts are surprising enough to need more explanation.\n\nNone of this is fatal. The architecture is plausible, the ablations are coherent, and the flaws are addressable. This paper deserves a serious referee, but the referee should insist on a clean comparison protocol: re-run all baselines under identical conditions, add multiple seeds and confidence intervals, choose top-k on a validation split, and correct the metric anomaly. After those fixes, the MvKeTR claim may well hold. I would send it to peer review, but not in its current form.","headline":"A sensible new architecture for 3D CT report generation with internally consistent ablations, but the SOTA claim rests on a comparison table that mixes cited and re-run baselines and contains a likely erroneous METEOR value.","tokens_in":23719,"tokens_out":1305,"would_cite":false,"duration_ms":13356,"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":"MvKeTR reads chest CTs from three anatomical planes, retrieves similar reports, and outperforms prior generators on nearly every metric.","keywords":["chest CT report generation","multi-view perception","knowledge enhancement","Kolmogorov-Arnold networks","view-aware attention","CT-CLIP retrieval","CT-ViT","radiology report generation"],"falsifier":"Re-run all Table II baselines under the exact 80/20 split, 224-cubed volumes, CT-ViT backbone, and beam size used for MvKeTR, and check whether the MvKeTR margins over SL-DG, Dia-LLaMA, and Reg2RG survive; if the cited rows shift when run in this setting, the state-of-the-art claim collapses. In the same experiment, choose top-k on a validation split rather than the test set to see whether the peak at k=16 is real.","tokens_in":22691,"feed_emoji":"🫁","tokens_out":8033,"duration_ms":68159,"temperature":0.7,"pith_summary":"This paper aims to show that automatic chest CT report generation improves when a model reads a volume the way a radiologist does: from the axial, coronal, and sagittal planes at once, and with reference to similar prior cases. The proposed architecture, MvKeTR, combines a Multi-View Perception Aggregator that fuses the three views through view-aware attention with a Cross-Modal Knowledge Enhancer that retrieves the most similar reports via a CT-CLIP model and feeds them in through cross-attention. It replaces MLP layers with Kolmogorov-Arnold Networks to capture fine high-frequency image details with fewer parameters. On the public CTRG-Chest-548K dataset the paper reports the best scores in nearly every automatic metric, including BLEU-1 58.36, BLEU-4 37.86, and ROUGE-L 54.25. If the comparison holds, the result matters because CT reports are time-consuming to write and error-prone, and the design offers a workflow-aligned template for automating them.","feed_headline":"Three-plane reading lifts chest CT report scores","feed_subtitle":"The new model fuses axial, coronal, and sagittal views plus retrieved similar reports, scoring 58.36 BLEU-1 on CTRG-Chest-548K.","key_machinery":"The load-bearing machinery is the pair of perception and knowledge branches on top of a shared sequence-to-sequence backbone. The Multi-View Perception Aggregator treats each anatomical plane as a separate token sequence and modifies ordinary attention by adding a learnable view embedding $E_v$ to the query–key logits, so the model can weight the axial, coronal, and sagittal evidence differently; the three streams are then concatenated and fused by a KAN layer. The Cross-Modal Knowledge Enhancer retrieves the $k=16$ most similar reports with a frozen CT-CLIP model, feeds their embeddings as keys and values in cross-attention with axial CT tokens, and preserves a residual connection so retrieved knowledge cannot override direct visual evidence. Kolmogorov-Arnold Networks, whose activations are learnable spline functions, replace MLPs throughout both modules; the paper argues from prior theoretical results that KANs scale linearly rather than quadratically in parameters and have less spectral bias, helping the model capture high-frequency lesion features.","core_discovery":"The paper's central claim is that a chest CT report generator should mirror a clinician's diagnostic routine, and that doing so yields state-of-the-art results. The model predicts a report from a 224-cubed CT volume by running three independent CT-ViT extractors on the axial, coronal, and sagittal reorderings of the volume; a view-aware attention mechanism adds a learnable view embedding to the attention logits, and the three normalized outputs are concatenated and fused through a KAN layer. A second branch retrieves the top-16 most similar reports from the CT-RATE bank using a frozen CT-CLIP encoder, then attends to those report embeddings with the axial visual features, again with residual connections and a KAN layer. The two feature sets are concatenated and fed to a three-layer cross-modal-memory transformer decoder to generate the findings and impression. On the CTRG-Chest-548K test set the paper reports that this design outperforms previous CT report generators on almost every metric, and the ablations attribute the gains to all three components, with multi-view perception contributing the largest single improvement.","pith_inferences":["The paper does not test retrieval banks drawn from CTRG-Chest-548K itself; a natural follow-up is to isolate whether the gains come from cross-dataset knowledge transfer or from any similar-case prior.","The top-k=16 choice peaks on the test set (Fig. 10); an honest hyperparameter check would select k on a validation split before reporting test numbers.","If the multi-view advantage transfers, the same view-token and view-aware attention design could be applied to other volumetric modalities such as MRI, where coronal and sagittal planes also carry complementary diagnostic information.","Because METEOR is the one metric where MvKeTR trails (Reg2RG scores 49.71), the 'almost all metrics' claim depends on metric weighting; evaluating with disease-level recall would test whether the qualitative nodule-detection gains generalize."],"forward_implications":["If the reported numbers are taken at face value, multi-view perception is the largest single contributor: BASE+MVPA improves the average over all six metrics by 17.2%, versus 14.5% for knowledge enhancement alone.","Replacing MLP layers with KAN layers accounts for a 9.5% gap in average improvement between the full model and the MLP variant, suggesting the activation choice matters beyond simply adding parameters.","The knowledge branch's residual design means retrieved reports act as a prior that can be overridden; the paper argues this preserves the model's ability to report novel findings not present in retrieved cases.","View-aware attention degrades gracefully under rotation and artifact perturbations in the paper's controlled experiments, with the largest rotation-induced drop being 3.80% and some artifact conditions slightly improving scores.","The modular design is compatible with alternative 3D extractors; Table V shows CT-ViT works best but 3D ViT, CT-Net, and U-Net can be swapped in."],"supporting_citations":[{"why":"Supplies the CTRG-Chest-548K dataset and the 80/20 split that define the evaluation; also provides the SL-DG comparison row.","marker":"[8]"},{"why":"Supplies CT2Rep as the prior 3D CT report generation framework and provides the CT-ViT visual extractor used in all runs.","marker":"[9]"},{"why":"Provides CT-RATE and the CT-CLIP encoders used to retrieve the top-k similar reports for knowledge enhancement.","marker":"[53]"},{"why":"Introduces Kolmogorov-Arnold Networks, the learnable-spline activation architecture that replaces MLPs in both modules.","marker":"[15]"},{"why":"Provides the expressiveness, parameter-scaling, and spectral-bias theorems the paper cites to justify KANs over MLPs.","marker":"[50]"},{"why":"R2GenCMN is the base report generator and the strongest re-run baseline; the ablation uses it as the BASE configuration.","marker":"[5]"}],"fun_headline_variants":["Multi-view + knowledge fusion lifts chest CT report scores","CT model reads like a radiologist, retrieves cases, beats SOTA","Three-plane reading plus case knowledge boosts CT reports","Radiologist-like multi-view fusion improves CT report accuracy","Multi-view perception and knowledge enhancement set new SOTA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole comparison rests on the assumption that the numbers in Table II are mutually comparable, yet the rows marked with a double dagger are copied from their original papers while the rest were re-run, and the paper gives no evidence that every row used the same split, preprocessing, and visual extractor.","fun_headline_variants_meta":{"raw":{"variants":["Multi-view + knowledge fusion lifts chest CT report scores","CT model reads like a radiologist, retrieves cases, beats SOTA","Three-plane reading plus case knowledge boosts CT reports","Radiologist-like multi-view fusion improves CT report accuracy","Multi-view perception and knowledge enhancement set new SOTA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000494,"raw_usage":{"total_tokens":2476,"prompt_tokens":1046,"completion_tokens":1430,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":662,"completion_tokens_details":{"reasoning_tokens":1359}},"tokens_in":662,"tokens_out":1430,"duration_ms":13160,"temperature":1.0,"reasoning_tokens":1359,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:18:50.229302+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run all Table II baselines under the exact 80/20 split, 224-cubed volumes, CT-ViT backbone, and beam size used for MvKeTR, and check whether the MvKeTR margins over SL-DG, Dia-LLaMA, and Reg2RG survive; if the cited rows shift when run in this setting, the state-of-the-art claim collapses. In the same experiment, choose top-k on a validation split rather than the test set to see whether the peak at k=16 is real.","supporting_citations":[{"cited_title":"Work like a doctor: Unifying scan localizer and dynamic generator for automated computed tomography report generation,","cited_arxiv_id":null,"evidence_quote":"Supplies the CTRG-Chest-548K dataset and the 80/20 split that define the evaluation; also provides the SL-DG comparison row."},{"cited_title":"Ct2rep: Automated radiology report generation for 3d medical imaging,","cited_arxiv_id":null,"evidence_quote":"Supplies CT2Rep as the prior 3D CT report generation framework and provides the CT-ViT visual extractor used in all runs."},{"cited_title":"KAN: Kolmogorov–arnold networks,","cited_arxiv_id":null,"evidence_quote":"Introduces Kolmogorov-Arnold Networks, the learnable-spline activation architecture that replaces MLPs in both modules."},{"cited_title":"On the expressiveness and spectral bias of KANs,","cited_arxiv_id":null,"evidence_quote":"Provides the expressiveness, parameter-scaling, and spectral-bias theorems the paper cites to justify KANs over MLPs."},{"cited_title":"Cross-modal memory networks for radiology report generation,","cited_arxiv_id":null,"evidence_quote":"R2GenCMN is the base report generator and the strongest re-run baseline; the ablation uses it as the BASE configuration."}],"review_version":1}