List of Papers Browse by Subject Areas Author List
Abstract
Manual reporting of 3D MRI studies is time-consuming, yet end-to-end, structured report generation for 3D liver MRI remains underexplored due to volumetric complexity and scarce paired data. We propose MRI2Rep, an autoregressive framework for liver MRI report generation. From 3{,}929 real-world MRI–report pairs acquired over a 10-year single-institution cohort, a Report-to-Label Canonicalization (RLC) module converts the free-text reports into structured, closed-vocabulary diagnostic sequences without lesion-level annotations. On a held-out test set, MRI2Rep achieves 76.0\% case-level sensitivity, 29.4\% lesion-level F1 (vs.\ $\leq$8.3\% for adapted medical vision–language baselines), and 82.4\% liver-level accuracy. In a blinded reader study, two radiologists rated 75\%/70\% of AI-generated reports clinically acceptable (vs.\ 95\%/100\% for originals), while our automated LLM-based judge (LLM-Eval, 61.8\%) applies a stricter standard, validating it as a conservative proxy. To our knowledge, this is the first end-to-end LI-RADS-structured reporting system for 3D liver MRI. The
code is available at https://github.com/Alena-Xinran/MRI2Rep.git
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2760_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Alena-Xinran/MRI2Rep
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiXin_MRI2Rep_MICCAI2026,
author = { Li, Xinran AND Yang, Junlin AND Shewarega, Annabella AND Zhou, Zongwei AND Chapiro, Julius AND Duncan, James S. AND Staib, Lawrence H.},
title = { { MRI2Rep: Autoregressive Structured Report Generation for 3D Liver MRI } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16879},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose an autoregressive framework, MRI2Rep, that generates structured radiology reports based on 3D MRI volumes of a liver tumors. The contribution is three-fold: 1) a report-to-label canonicalization (RLC) module is proposed to first convert free-text reports to a standard vocabulary to be used for training; 2) an encoder-decoder model is trained to predict these sequences from the MRI sequences; and 3) a hand-crafted sentence template module is used to convert the output sequence back into a report with full sequences. The method is evaluated on a datasets of 3,929 report-image pairs of liver MRI, and compared to four VLMs.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Use of domain-knowledge through LI-RADS and structuring of free-text data: mapping both the free text reports and the outputs of the model therefore on LI-RADS appears a strong choice, as this is a main instrument used by radiologists for accessing liver lesions. 2.Clinical reader study: letting two radiologists score the generated reports is a good way of accessing whether the reports meet clinical criteria. 3.Systematic error analysis on test set to discover patterns: gives insight in the methods and direction for future improvements.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The dataset is not described, we only now the number of samples. This makes it very difficult to assess the results: we have no idea where this dataset originates from, how reports where generated, how the scans were acquired, etc. The evaluation setup is also poorly described. A held-out test set is mentioned, but its not mentioned what the size is, how this dataset was created from the main dataset, etc. A cross-validation setup would also have been much stronger, as this could have been a lucky or unlucky split, which we also cant estimate the risk of as we dont know the dataset. No statistical testing is used as well. 2.Comparison with other methods. Related work on VLMs includes only a few models, four (M3D, CT2Rep, Merlin, RadFM) are used in the comparison experiments. Other models that seem more suitable as these are either newer, trained / evaluated on more similar data (liver MRI) or more data are for example: https://doi. org/10.1016/j. media. 2026.103992, https://doi. org/10.48550/arXiv. 2509.21249.Highly related are also these VLMs specifically made for Liver MRI: https://doi. org/10.3390/app152312578, https://doi. org/10.1016/j. ebiom. 2025.106060, 3.No clear problem is introduced in the introduction. The authors claim that “automated report generation for 3D MRI remains largely unexplored”, which seems to be the problem that is being addressed, but that is incorrect, see above. Hence its unclear what technical or clinical problem the proposed methods are exactly addressing. 3.The organization of the paper can be improved. The abstract misses essential components (see below), the introduction does not include a clear problem statement (see above), results are reported in the introduction which seems to be a repetition of the abstract, the related work section just bullet points some methods without a clear explanation of them, results section does include methods / statistics, 4.It would be good to perform ablation experiments for both the RLC and the model. Would the other VLMs achieve a similar performance if the RLC setup and vocabulary would already be used? 5.Performance for liver tumor phenotyping, the most important task, is poor across all models including the proposed method (F1 score of 29.4). Hence none of the methods seem to achieve satisfactory results. 6.Poor reproducibility: no code or data published. 7.No motivation is provided for using the specific selected sequences (arterial and portal venous phase). Why not use complete MRI exams, mimicking clinical practice? Routinely used other sequences such as a T2-weighted MRI, T1 in-out phase, DWI / ADC are now a crucial missing component in the analysis. 8.No motivation is provided for the encoder-decoder architecture. How does this compare to the literature? 9.The RLC module is poorly described. Section 3.1 explains what the model does, but not how it performs this task. Besides that is methodologically unclear, its also not reproducible. In the results it seems that LLMs are used based on this module, but thats not explicitly described. 10.It should be acknowledged that LI-RADS is only applicable for cirrhotic patients. 11.Abstract: The abstract appears incomplete, there is no discussion / conclusion. Generally the abstract could be improved, the performance of the baselines is not mentioned, and its unclear how this blinded reader study was performed and what “cilinical acceptability” means or what LLM-Eval is.
- Please rate the clarity and organization of this paper
Poor
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The two major factors leading to my decision where the absent description of the dataset and poor evaluation setup, which makes it difficult to assess the results, and that many related work was missed making the novelty of the paper unclear and some of the claims unjustified.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
While I appreciate the rebuttal which addresses some of my concerns, unfortunately there are simply too many aspects missing in the paper for me to accept it in its current form.
Specifically w.r.t. the author feedback: [C1] The dataset and evaluation should have been better described in the paper. k-fold CV may be not the most practical method due to computational resources, but would be much stronger. [C2] Related work should have been better reported in the paper. See also other comments: none of the methods indeed solve this exact problem, but a better comparison / combination could have been made. [C3] Indeed the proposed method triples the performance of baselines, but an F1 of 29.4 is still very poor and thus not a satisfactory result. [C4] It’s clear what the goal of the RLC is and the idea is strong, but the description of the method remians poor and not reproducible, and no code is provided. [C6] Agreed that the choice for ART/PV is clinically well-motivated, but technically it would have been stronger to include all standard sequences, even those that are not used for LI-RADS classification. But this is a minor concern.
Review #2
- Please describe the contribution of the paper
This work proposes the MRI2Rep framework for templated liver MRI report generation. The framework consists of three components: the Report-to-Label Canonicalization (RLC) module which converts reports into structured diagnostic sequences, an encoder-decoder model which uses uses the input MRI volume to produce the diagnostic sequence, and a sequence-to-report rendering module to deterministically map the diagnostic sequence to output report using hand-crafted sentence templates.
RLC consistently improves the radiologist retention rate of extracted lesion labels across three LLM backends, and various ablation studies were conducted to demonstrate the accuracy of predicted lesions. Radiologists accepted 70-75% of AI-generated reports, compared with automated LLM-Eval which was stricter with an acceptance rate of 61.8%.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The authors have included extensive ablation study results in table 2 to justify modelling choices such as the inclusion of of a visual encoder and ART-to-PV pre-training.
Succinct justifications were given in various places which adds to credibility, e.g. inter-rater agreement for the clinical reader study.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The related work section is fairly limited. For instance, recent works in chest X-ray report generation have used structured label extraction as an auxiliary task [1] or basis for generating the full report [2]. Additionally, LLM-based label extraction from radiology reports could benefit from citing more directly relevant works such as [3].
The motivation for introducing stylistic variation in the sequence-to-report rendering stage isn’t clear. Was it to prevent radiologists from identifying AI-generated reports in the reader study? It would be useful to add a sentence to clarify the link if this is the case.
[1] Jianguo Liang. ‘Multi-Task Learning for Radiology Report Generation with Structured Findings Consistency’. Computer Science Bulletin 8, no. 01 (2025): 477–89.https://doi.org/10.71465/csb178.[2] Hou, Wenjun, Kaishuai Xu, Yi Cheng, Wenjie Li, and Jiang Liu. ORGAN: Observation-Guided Radiology Report Generation via Tree Reasoning. Version 1.2023.https://doi.org/10.48550/ARXIV.2306.06466 [3] Castro, Daniel C., Aurelia Bustos, Shruthi Bannur, et al. ‘PadChest-GR: A Bilingual Chest X-Ray Dataset for Grounded Radiology Report Generation’. arXiv:2411.05085.Preprint, arXiv, 7 November 2024.https://doi.org/10.48550/arXiv.2411.05085.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
Section 4.1 RLC Validation: Is there a reason for phrasing the metric as retention rate rather than e.g. precision? If I understand correctly, an AI-produced label is retained if it is supported by the original report, i.e. retention rate measures the proportion of true positives to all predicted positives.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The presented framework is well-supported by extensive experiments. From what I understand (first point in the main contributions) the authors intend to make the dataset available, which would be a useful resource.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have clarified the novelty compared to prior works in the literature, though it would be useful to mention these in the paper itself. The work itself appears scientifically sound.
Review #3
- Please describe the contribution of the paper
This article focuses on the task of automatic report generation for 3D liver MRI, and proposes the MRI2Rep framework to address the difficulties of “high volume data dimension, large report supervision noise, and lack of lesion level annotation”. The core idea is to first convert free text radiation reports into closed set, auditable structured sequence labels through Report to Label Canonical (RLC), then train a 3D CNN+Transformer encoder decoder model to generate structured diagnostic sequences from multi-phase MRI volume data through autoregression, and finally render them into natural language reports using templates.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The most impressive aspect of the paper is “end-to-end generation of structured reports”. One of the biggest innovations of this article is RLC, which transforms free text reports into standardized label sequences and achieves a retention rate of around 98% in radiologists’ tests, demonstrating the high quality of label extraction itself.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The structured report generation appears complete, but the core lesion level performance is still not high. 2.RLC is strong, but the entire system heavily relies on the quality of front-end text normalization. 3.”No need for lesion level spatial annotation” is an advantage, but it also limits the accuracy of lesion localization. 4.The model is fragile in scenarios of “post-treatment changes” and “complex medical history”.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The experiment was based on 3929 pairs of MRI report samples, of which 3830 had complete multi-phase sequences. The final model achieved 76.0% case level sensitivity, 29.4% lesion level Lesion-F1, 82.4% liver background classification accuracy on the test set, and achieved a clinical acceptance rate of 75%/70% in the blind evaluation of dual radiologists.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
1.Regarding Lesion-F1, it is only 29.4% The author’s explanation carries a certain degree of plausibility: 29.4% is not a standard binary classification F1 score, but rather a strict match of type + pos, meaning both lesion type and Couinaud segment location must be accurately determined simultaneously. Moreover, it is a set-F1 score across 5 LI-RADS categories × 9 anatomical regions, making it significantly more challenging than simply determining “presence or absence of lesions.” The author further emphasizes that other 3D-VLM baselines achieve a maximum of only 8.3%, while MRI2Rep has improved to 29.4%, clearly demonstrating its notable advantage in similar tasks. However, our original skepticism still holds some validity: significant relative improvement does not equate to absolute performance being sufficiently high. For clinical report generation, lesion type, location, and quantity are core facts. The Lesion-F1 score of 29.4% and Triplet-F1 score of 28.1% still indicate instability in the model’s precise structured diagnosis. The original paper’s Table 2 indeed shows MRI2Rep’s Lesion-F1 at 29.4% and Triplet-F1 at 28.1%. Overall, we agree with the author’s explanation regarding task difficulty and relative improvement, but from an absolute performance perspective, the 29.4% lesion-level F1 score still indicates that the model has a certain gap in generating reliable lesion-level structured reports. 2.Regarding the RLC dependency issue The author emphasizes that RLC is not merely a simple black-box LLM call, but rather a LI-RADS-guided prompting approach combined with deterministic triplet assembly and per-token evidence sentences. Moreover, it maintains a retention rate of at least 91% across three LLM backends, demonstrating the strong stability of RLC. The issue does not lie in “RLC being inaccurate,” but rather in the model primarily learning the closed-set label space defined by RLC, rather than the full semantics of complete free-text reports. Therefore, the high quality of RLC demonstrates its effective semantic normalization at the front end, but it does not mean the system is entirely free from the expressive constraints imposed by a closed-set label space. Complex medical histories, post-treatment changes, follow-up comparisons, ambiguous expressions, and cross-sentence references—once inadequately encoded by the RLC system—become difficult for subsequent models to learn. The original paper also acknowledges that its output is a closed-set structured sequence, rendered into reports via templates. Overall, the author demonstrates that RLC achieves high-quality label extraction and mitigates concerns about “front-end label noise.” However, this does not eliminate the system’s reliance on the RLC design space. Thus, RLC serves as both a key contribution to improving label quality and auditability in this paper, while also defining the scope of learnable and expressive diagnostic semantics to some extent. 3.3.Advantages and Disadvantages of “No Need for Lesion-Level Spatial Annotation” The author emphasizes that the “lack of lesion-level spatial annotations” stems from considerations of clinical applicability, as retrospective real-world data typically lacks bounding boxes or segmentations. This claim is valid, and the original paper explicitly highlights “no need for lesion-level spatial annotations” as a core advantage. However, the experimental results also suggest that this weakly supervised approach may limit the lesion localization capability. The model’s performance on multi-lesion cases significantly declined in the paper, and the authors acknowledged in their rebuttal that the remaining gap is related to the lack of lesion-level spatial annotations. Overall, the absence of lesion-level spatial annotation is a practical advantage of MRI2Rep, but it may also be a significant factor limiting its lesion localization performance. The authors’ explanation justifies this design, yet its limitations in complex, multi-lesion scenarios remain to be acknowledged. Therefore, this is more of a trade-off between practical usability and fine-grained localization capability, rather than a mere methodological flaw. 4.Regarding post-treatment changes and complex medical history The author’s response states that post treatment/complex history fragility has been acknowledged and they are ‘lies outside the LI-RADS major feature scope’. This explanation makes sense at the method boundary, but from the perspective of generating real clinical reports, it is still a limitation that needs to be taken seriously. Because real radiology reports do not only assess the major features of LI-RADS, but often require a combination of past treatments, postoperative changes, follow-up comparisons, and medical history background. The error analysis in the original paper also mentioned that 34.1% of false positives occurred in previous ablation cases, indicating that this issue is not an occasional error, but rather an important source of model false positives. Overall, the author categorizes post-treatment changes as outside the scope of LI-RADS major feature, which helps to clarify the boundaries of the method presented in this article; But for the task of “automatic report generation”, treatment history and follow-up information are not optional additional information, but important components of the real diagnostic context. Therefore, this issue should not be simply regarded as an out of boundary problem, but rather as an important limitation for subsequent clinical generalization and safety evaluation. 5.Regarding datasets and generalization The author added that the data comes from a 10-year retrospective cohort of a single tertiary hospital, which enhances the transparency of the data source description; However, from the perspective of model generalization, this also suggests that there is still a lack of external multi center validation. Set over split can improve the coverage of test set categories and enhance the representativeness of internal testing, but it still cannot replace external validation across centers, devices, and scanning protocols. Therefore, the current results better demonstrate the effectiveness of the model on the internal test set, but cannot fully prove its cross center generalization ability.
Author Feedback
We thank the AC and Reviewers (R1-R3). We address six AC-priority points.
[C1] Dataset & evaluation setup (AC; R1.W1). Dataset: 3,929 multi-phase liver-MRI report-image pairs from a 10-year retrospective cohort at one tertiary academic hospital, on Siemens scanners (predominantly 3T, with earlier 1.5T) using extracellular GBCA (e.g., gadoterate); reports by board-certified abdominal radiologists; IRB-approved (consent waiver). RLC maps free-text reports to closed-vocabulary labels; 3,830 with complete ART and PV are retained (Sec.3.1). Stratified 80/10/10 set-cover split on (type,pos) yields 383 test cases covering rare LI-RADS categories. k-fold CV is impractical for our 3D autoregressive pipeline; set-cover mitigates split variance, and the 21-point Lesion-F1 margin over the best baseline (8.3 to 29.4) is well outside any plausible bootstrap CI at n=383.Code released upon publication (R1.W6).
[C2] Problem statement, related work & novelty (AC; R1.W2/W3; R2.W1). Sec.1 names two challenges: noisy report-level supervision on small 3D cohorts and volumetric/multi-phase complexity. Our novelty is closed-vocabulary, LI-RADS-structured 3D-MRI report generation, not 3D-MRI reporting in general. Of R1’s 4 cited works, the liver-MRI papers (Appl Sci 2025; eBioMedicine 2025) target classification/VQA, and Media 2026 + arXiv:2509.21249 are general 3D-MRI representation learning; none does end-to-end LI-RADS-structured reporting. CXR structured-finding works (ORGAN, PadChest-GR; R2.refs) are 2D and do not enforce LI-RADS reasoning.
[C3] Lesion-level F1 = 29.4 (AC; R1.W5; R3.W1/W4). 29.4 is strict set-F1 over (type,pos) across 5 LI-RADS by 9 Couinaud sectors - far harder than binary detection. Under the identical protocol (Sec.4.2), all four 3D-VLM baselines score Lesion-F1 at most 8.3 (Tab.2); MRI2Rep more than triples this. Case-level sensitivity 76.0%; blinded radiologists accept 70-75% of reports (k=0.97). The remaining gap reflects the deliberate absence of lesion-level spatial annotations, a framework choice for clinical applicability. Post-treatment / complex-history fragility is acknowledged and lies outside LI-RADS major-feature scope.
[C4] Dependence on RLC / front-end quality (AC; R3.W2; R1.W9). RLC is an LI-RADS-guided LLM prompting protocol (Sec.3.1) with deterministic triplet assembly and per-token evidence sentences - not a black-box call. Retention at least 91% across three independent LLM backends (Tab.1); the 98.1% headline averages two annotators. Small backend variability shows the structured logic, not any single LLM, drives quality.
[C5] Ablation comparability with VLMs (R1.W4). Tab.2 evaluates baselines under matched output taxonomy: their free-text outputs are mapped to the same closed (type,pos,qty) labels via identical Claude-3.5 extraction (Sec.4.2). The gap (Lesion-F1 8.3 or lower vs. 29.4) reflects model capability under matched vocabulary, not taxonomy mismatch.
[C6] Sequence choice & architecture (R1.W7/W8). Our standard extracellular GBCA protocol has no hepatobiliary phase: LI-RADS major features (APHE, washout) are by definition assessed on ART/PV and mandatory for HCC categorisation, a clinically appropriate, not arbitrary, choice. T2 HASTE and T1 Dixon are acquired but ancillary, not major-feature inputs. The without-PV ablation (29.4 to 13.5; Tab.2) supports this. Encoder-decoder follows 3D-captioning practice (CT2Rep, Merlin); the no-visual-encoder ablation (to 10.4) justifies the design.
[Minor] R1.W10: LR-1 to LR-5 apply to at-risk patients; non-HCC tokens (cyst, hemangioma, targetoid) and a cirrhosis liver-background token extend coverage. R1.W11: “clinical acceptability”=radiologist accept/reject of the rendered report (Sec.4.3); “LLM-Eval”=Claude-3.5 LLM-judge over rendered vs reference reports, stricter (61.8%) than radiologists (70-75%). R2: “retention” equals precision numerically but stresses auditability vs. source. Stylistic variation prevents lexical artefacts unblinding readers.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
I recommend Invite for Rebuttal. Reviewer 1 was clearly negative and argued that the dataset description and evaluation setup are too incomplete to assess the work reliably. I share Reviewer 1’s concern that the dataset description is insufficient. Reviewer 2 was positive and found the framework promising, but still pointed out limited related-work coverage. Reviewer 3 also supported moving forward, while noting that lesion-level performance is still not strong and that the full system depends heavily on front-end normalization quality.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper proposes a structured report generation framework for 3D liver MRI, with Report-to-Label Canonicalization as a central contribution. The initial concerns focused on insufficient dataset and evaluation description, limited related-work coverage, low lesion-level F1, dependence on RLC quality, and limited reproducibility. The rebuttal clarified the dataset source, split, MRI sequences, RLC design, baseline mapping protocol, and the interpretation of the strict lesion-level F1 metric. Reviewers 2 and 3 support acceptance after rebuttal and consider the framework scientifically sound and valuable.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors have addressed the reviewers’ comments well. 2 reviewers acknowledged the significance of the work with detailed justifications. Therefore acceptance of the paper is concluded.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors have adequately addressed most of the reviewers’ concerns. However, the dataset description mentioned in the rebuttal should be detailed further and included in the final version.
