List of Papers Browse by Subject Areas Author List
Abstract
Large-scale benchmarks such as BraTS have driven progress in brain tumor segmentation, but they provide only masks with limited access to the clinical semantics found in radiology reports. We introduce ReportX, a paired resource of 257 clinical reports aligned to BraTS-GLI-2023 subjects, structured into a rich set of qualitative and quantitative attributes. Qualitative fields are curated by clinicians, while quantitative descriptors are automatically derived via atlas-based localization and geometric computations. We compare our annotation schema to existing report-augmented datasets and show that ReportX provides substantially broader coverage of clinically relevant factors. To exploit this supervision, we encode reports using biomedical language models and incorporate their embeddings as auxiliary semantic guidance for 3D tumor segmentation during training. Experimental results demonstrate that the proposed vision-text alignment improves segmentation performance on standard BraTS metrics, with clinically curated reports providing more consistent improvements than automatically generated or less-structured counterparts. We publicly release the dataset (https://ditto.ing.unimore.it/reportx) and the source code (https://github.com/AImageLab-zip/ReportX).
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0852_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/AImageLab-zip/ReportX
Link to the Dataset(s)
https://ditto.ing.unimore.it/reportx/
https://www.synapse.org/Synapse:syn51156910
BibTex
@InProceedings{MarKev_ReportX_MICCAI2026,
author = { Marchesini, Kevin AND Carpentiero, Omar AND Del Gaudio, Livia AND Farioli, Francesco AND Cucchiara, Rita AND Grana, Costantino AND Cuculo, Vittorio AND Bolelli, Federico},
title = { { ReportX: The BraTS Clinical Report Dataset } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16895},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper introduces ReportX, a novel text dataset for brain tumor segmentation comprising 257 paired MRI scans sourced from the BraTS-GLI-2023 challenge. The authors innovatively propose a bidirectional text annotation and verification pipeline that combines medical expert curation with atlas-based automated feature extraction. This workflow efficiently generates high-quality, highly accurate, and professional radiology reports, offering significant value for the advancement of multimodal (image-text) intelligent processing in medical imaging.
Furthermore, the authors present a vision-text alignment segmentation network guided by loss function constraints. By incorporating text exclusively as auxiliary supervision during the training phase, the model achieves superior segmentation performance during the image-only inference phase compared to baseline models trained solely on images. This design holds practical advantages for real-world clinical deployment. Finally, the provision of publicly accessible datasets and source code ensures strong reproducibility for this research.
- 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.
Well-motivated with strong clinical relevance: The authors rightly point out that many existing 3D medical multimodal datasets are overly simplistic in structure and lack sufficient clinical information. Innovating from a dataset perspective, therefore, holds significant value for real-world clinical applications.
Novel methodology and rigorous evaluation: The paper presents a novel approach for refining and validating text reports. Furthermore, it introduces internal evaluation metrics, such as the Concept Coverage Score (CCS), to quantitatively demonstrate the superior quality of the curated reports.
Highly practical model design: The proposed network exhibits strong clinical practicality. Specifically, the text-free inference phase aligns perfectly with real-world clinical diagnostic scenarios where high-quality structured reports are often unavailable at the time of inference.
Good reproducibility: The authors have provided anonymized datasets and source code, which greatly promotes transparency and will benefit future research within the open-source community.
- 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.
Confounding Variables in Dataset Comparison: While the proposed ReportX dataset contains rich clinical semantics, the experimental design fails to conclusively prove that this richness is the primary driver of the performance gains. As shown in Figure 2 and Table 2, the sample sizes across the compared datasets are highly imbalanced (e. g. , N=257 for ReportX vs. N=46 for AutoRG). Without a strictly controlled ablation study (e. g. , comparing exactly 65 overlapping samples across all datasets), it is impossible to determine whether the improvements stem from the quality/richness of the reports or simply the larger quantity of text supervision. Lack of Methodological Novelty and Feature Space Analysis: The vision-text alignment module relies primarily on a contrastive learning objective. Although the authors introduce minor adaptations (e. g. , a positive-only formulation and VICReg regularization), the underlying mechanism remains highly standard, offering limited technical novelty. Furthermore, forcing a global alignment between highly disparate modalities (a text Transformer and a 3D U-Net) raises concerns about representational degradation. Specifically, it remains unclear whether aggressively aligning the image encoder’s features with text embeddings compromises the encoder’s original spatial representation ability required for dense 3D segmentation. The paper lacks quantitative analysis or qualitative visualizations (e. g. , t-SNE or feature activation maps) to alleviate this concern. Insufficient Baselines for Segmentation: Although the paper proposes a practical text-free inference setting, it only compares the proposed method against its own vanilla 3D U-Net baseline. To convincingly demonstrate the value of introducing textual supervision, the authors must compare their results against well-established, state-of-the-art (SOTA) image-only segmentation baselines (e. g. , nnU-Net, Swin UNETR). Without this, it is hard to gauge the true impact of the text modality on pushing the upper bound of current segmentation performance. Ad-hoc Engineering Tricks: The sudden introduction of the “whitening” technique in the experimental phase seems disconnected from the core contributions of the paper. While empirically beneficial (as shown in Table 2), it acts more like an ad-hoc engineering trick to fix BERT’s anisotropy rather than a deeply integrated methodological innovation, diluting the focus of the proposed framework.
- Please rate the clarity and organization of this paper
Satisfactory
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My overall recommendation for this paper is a Weak Reject. I genuinely appreciate the authors’ substantial effort in curating the BraTS-ReportX dataset. Providing clinically validated, highly structured radiology reports for the BraTS-GLI-2023 challenge is a valuable contribution to the medical vision-language community, and the clinical motivation behind this work is well-founded.However, my decision is primarily driven by significant methodological flaws and critical gaps in the experimental evaluation, which severely undermine the paper’s core algorithmic claims. The major factors leading to this score are as follows:Fundamentally Unfair Dataset Comparison (Confounding Variables): The paper claims that the superior quality of ReportX leads to better segmentation performance. However, the comparison in Table 2 is conducted with highly unbalanced sample sizes. Because the experiments do not control for the volume of text supervision, it is scientifically impossible to conclude whether the observed performance gains are driven by the higher quality of the reports or simply the larger quantity of available paired data. A strictly size-matched ablation study is entirely missing.Lack of Strong Image-Only Baselines: The authors attempt to demonstrate the value of text-guided segmentation, but they only compare their approach against a vanilla 3D U-Net baseline. To convincingly prove that introducing the text modality provides a meaningful boost in this domain, the method must be benchmarked against established, state-of-the-art (SOTA) image-only segmentation frameworks (e.g., nnU-Net, Swin UNETR). Without this, it is difficult to justify the necessity of the proposed cross-modal alignment.Absence of Multimodal Baselines: The experimental section does not compare the proposed text-vision alignment method with any other contemporary multimodal or vision-language segmentation algorithms. This makes it impossible to assess where the proposed technical approach stands compared to existing literature.Conclusion:While the proposed dataset is a commendable resource, a paper introducing a new dataset alongside a novel training framework must rigorously prove the utility of both. From an experimental evaluation perspective, the validation lacks the rigor and appropriate baselines required to substantiate the authors’ claims. Therefore, I cannot recommend acceptance in its current state.
- Reviewer confidence
Very confident (4)
- [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.
Based on the author’s statements during the rebuttal phase, I have resolved my concerns regarding the issue of unfair comparisons. Although I believe the paper still has shortcomings in terms of evaluation and methodology, it provides complete textual data and reproducible code. Given the need to advance this field, I believe the paper is acceptable.
Review #2
- Please describe the contribution of the paper
Provide high-quality reports on the well-known BRATS dataset, demonstrating its value through intrinsic metrics as well as performance on downstream tasks.
- 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 BRATS dataset is a well-established and widely used benchmark with proven quality. However, it often lacks high-quality radiology reports. The addition of such reports would substantially increase its utility and impact within the research community.
- 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 design of the downstream task lacks novelty. The authors rely on established techniques such as VICReg, positive pair-based training, and Dice Focal Loss, all of which are often used in the field. While the ablation study provides insight into which components contribute to performance, the primary objective of the paper appears to be demonstrating the effectiveness of the BratsReportX dataset. However, this objective could likely be achieved using alternative approaches, whether simpler or more sophisticated. In essence, the work mainly focuses on comparing datasets rather than introducing methodological innovation. The contribution, therefore, lies in leveraging existing techniques to highlight the value of the proposed dataset.
Another weakness is the lack of clarity and detail in the experimental setup:
1.It is unclear how accuracy is computed when validating the automatically derived fields. If clinicians evaluate the same cases using predefined queries in a fully automated pipeline, it is not obvious how discrepancies arise. Given that the information is derived from the same segmentation masks, one would expect consistent outputs.
2.The evaluation relies heavily on self-defined metrics. It is unclear why commonly used metrics for report quality, such as TBFact, BERTScore, or RaTEScore (as used in BT-Report), were not included. Incorporating at least one widely accepted metric would strengthen the evaluation, even acknowledging their limitations.
3.In Table 2, how were these values calculated for the 251 test cases? In particular, how can the reported standard deviation (i.e., ± value) be equal to 0?
4.The performance differences reported in Tables 2 and 3 are marginal. It would be important to conduct statistical significance testing to support the claims. Without such analysis, small variations or outliers could affect the ranking (e.g., Table 2 TC: 84.4 ± 0.2).
- Please rate the clarity and organization of this paper
Satisfactory
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Given the importance of high-quality radiology reports, especially in relatively underexplored domains such as MRI, I am inclined toward a weak accept. However, there are concerns regarding the experimental setup and the reported results. I would update my rating after these issues are adequately addressed.
- Reviewer confidence
Very confident (4)
- [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 clarified most of the points I raised, including the reported metrics in the results and the novelty of their method. If these clarifications and the additional statistical analysis are properly incorporated into the final version of the paper, I believe the work would make a valuable contribution to the community.
Review #3
- Please describe the contribution of the paper
This paper proposed a new multimodal dataset ReportX combining brain tumor dataset with radiology report. The authors also provide detailed quantitative analysis of the dataset quality along with sufficient improvement for downstream 3D segmentation tasks.
- 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.To the best of my knowledge, this is one of the pioneering works that used neuroradiologists, following standard clinical protocol for report generation, instead of LLM-based prior solutions. 2.It is also interesting that the report is only required while training, whereas the inference is purely image-based, making it practical. 3.The authors use SigLip-inspired vision-text alignment strategy to mitigate the drawbacks of existing VLM alignment strategies.
- 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.Although the work is pioneering, the scale is relatively small (257 samples only). An expansion to the entire Brats dataset would be nice. 2.The experimental rigor could have been improved by adapting it to more sophisticated and advanced segmentation pipeline (like transformer). 3.The authors proposed using different BERT encoders for text embeddings, however, the experimental results section missing how the performance changes based on different encoders. 4.The paper is missing sufficient clinical details about the annotation and dataset generation. Especially this part: “The dataset was divided into….. annotated by both clinicians” needs further clarification. How are the two splits obtained and details about the 50 overlapping cases would have been helpful.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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 paper has a strong dataset contribution along with sufficient baseline evaluation, justifying my positive score. My score isn’t higher because the dataset size is comparatively small and the paper is missing some clinical details.
- 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 justified my concerns, hence I keep my score.
Author Feedback
Reviewers recognized the strong motivation, clinical relevance, and pioneering nature of our ReportX dataset [All], as well as the novelty of our report-validation approach [R1]. As clarified throughout the paper, the primary contribution of this work is the dataset itself, together with its validation pipeline. As noted by R1 and R2, our text-free inference design has strong practical value. All reviewers recognized our efforts toward reproducibility, including the release of both code and dataset.
Sample size [R1] The report dataset sizes are: ReportX-257, TextBraTS-365, AutoRG-235, and BTReport-1,251.Only 40 cases are shared across all datasets. In Fig.2(b), we report pairwise CCS between ReportX and each competitor, using all available pairs to ensure the fairest possible comparison; restricting the analysis to the 40 common cases would yield the same pattern, but with substantially fewer points in the plot. For training, we likewise use all reports available in each dataset to make the best possible use of each. Despite being comparable to or smaller than competitors, ReportX achieves the best DSC and HD95 gains, showing that quality, not quantity, drives performance.
Novelty and baselines [All] To our knowledge, report-guided 3D brain tumor segmentation with image-only inference has no prior work in the 3D MRI domain. Most vision-language segmentation methods use text at inference time (VoxTell, [27]), while the closest related works are either not segmentation-specific report-based pre-training [19,34] or leverage different supervision settings [6]. Given the lack of prior work, we design the first architecture for this task and adapt SigLIP to a positive-only 3D multi-GPU setting. Tab.3 shows that each component contributes effectively. Moreover, we intentionally built our segmentation framework on standard SOTA modules, a 3D UNet (same architecture as nnU-Net) and SigLIP loss with VICReg, to isolate the dataset’s contribution from additional architectural complexity.
Whitening [R1] BERT embedding anisotropy is a well-documented phenomenon [10]. We see whitening as a well-grounded correction rather than an engineered trick and, as shown in Tab.2, it yields consistent benefit across all datasets and encoders.
Text encoders [R2,R3] Results in Tab.2 are averaged across the 3 text encoders. Std ≈ 0 (i.e., <.05) reflects that encoder choice has a negligible effect in this setting, a positive finding demonstrating invariance across encoders.
Alignment and stat. analysis [R1,R2] Following well-established literature approaches [13,19], we perform vision-text alignment in a separate projection space via linear layers. Moreover, the contrastive loss serves as regularization rather than the primary loss. Tab.2 shows that our approach outperforms the baseline across datasets. Statistical analysis on considered metrics confirms our findings: i.e., for average DSC the Friedman test is significant (n=251, α=.05, p=.002). Nemenyi test: Critical Distance (CD)=.385; mean ranks: ReportX-Ours(1.57)<BTReport(2.77)<AutoRG(2.96)<TextBraTS(3.35)<Baseline(4.35). All pairwise differences are significant according to CD, except AutoRG vs. BTReport. Overall, the test confirms that the baseline performs worst, while ReportX significantly outperforms all competing methods.
Annotations [R2,R3] The 257 cases are split between 2 clinicians using institution-stratified sampling; the 50 overlapping cases follow the same criteria. On these 50 subjects, we validate the methods for automatically extracting the “derivable fields” (a small subset of the entire report): clinicians were asked to manually annotate these fields to measure accuracy against the automatic outputs.
Report metrics [R2] BERTScore, TBFact, and RaTEScore evaluate generated reports against references, whereas ours are clinician-written. We adapt RadFact for inter-rater consistency and introduce CCS/TA as coverage metrics grounded in neuroradiology ontologies.
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.
The paper provides a BraTS-ReportX dataset, which introduces clinically enriched radiology reports for multimodal medical learning. However, the experimental design suffers from significant confounding factors, weak and incomplete baselines, making the claimed performance gains difficult to interpret.
- 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 concerns regarding the method and experiments have been largely resolved, and all reviewers agree to accept the paper.
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.
My decision is based solely on my own read through the paper, the original review comments and the rebuttal. Since the original reviewers did not provide a new score (as far as I am aware it was not mandatory), my comments are based on whether I believe the authors have truly responded to the major concerns.
Starting from the paper itself, I do agree with the reviewers that there is a clear data contribution, even if the scope is limited to 257 cases and not the whole dataset. In my opinion, this is a strong point and the main reason for a temptatively high score.
In my opinion, parts of the rebuttal could have been expanded to provide more clarity (I will admit that I am not too familiar with parts of the methodology), but as pointed by reviewer #2 major concerns have been addressed. I am also slightly worried about the sample size difference post-rebuttal, but I think the merits of the dataset contribution are more important than some of the validation shortcomings and comparison fairness.
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 clarified most of the concerns raised by the reviewers, who have all changed their scores to “Accept” after the rebuttal.