List of Papers Browse by Subject Areas Author List
Abstract
Magnetic Resonance Imaging (MRI) interpretation is fundamental to clinical decision-making, requiring radiologists to integrate multi-view anatomical planes across sequential timepoints while precisely localizing interval changes. However, existing vision-language benchmarks remain confined to single-timepoint, single-view interpretation, failing to capture the temporal-spatial reasoning essential to radiologic practice. We introduce the Time-Aware Multi-View MRI Benchmark, an evaluation framework unifying multi-view anatomical input, temporal reasoning across longitudinal scans, and structured localization guidance. The benchmark comprises 3,920 expert-verified question-answer pairs derived from 890 patients across over 3,200 longitudinal MRI timepoints, drawn from seven clinical cohorts covering glioblastoma, neurodegeneration, vestibular schwannoma, and brain metastases, in open-ended, multiple-choice, and binary formats, requiring models to identify anatomical regions of maximal change, characterize progression across sequences and views, and provide structured guidance specifying boundaries, imaging features, and confounders. Experiments across 17 vision-language models reveal moderate temporal alignment but systematic failure on change direction recognition and volumetric quantification, while multi-view inputs improve spatial localization yet degrade temporal reasoning in compact architectures. Our benchmark establishes a systematic standard for evaluating progression tracking, interval change localization, and temporal ordering essential for clinical deployment.
Code, evaluation splits, and dataset are available at https://github.com/wafaAlghallabi/Time-Aware-MRI
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3355_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/wafaAlghallabi/Time-Aware-MRI
Link to the Dataset(s)
Yale-Brain-Mets-Longitudinal: https://www.cancerimagingarchive.net/collection/yale-brain-mets-longitudinal/
UCSF-ALPTDG: https://doi.org/10.1148/ryai.230182
UCSD-PTGBM: https://www.cancerimagingarchive.net/collection/ucsd-ptgbm/
LUMIERE: https://doi.org/10.6084/m9.figshare.c.5904905
OASIS-2: https://www.oasis-brains.org/
ADNI: https://adni.loni.usc.edu/
Vestibular-Schwannoma-MC-RC: https://www.cancerimagingarchive.net/collection/vestibular-schwannoma
BibTex
@InProceedings{AlWaf_How_MICCAI2026,
author = { Al Ghallabi, Wafa AND Thawkar, Ritesh AND Ghaboura, Sara AND Thawakar, Omkar AND Saeed, Numan AND Al Nuaimi, Dana AND Alkatheeri, Ajnas AND Khan, Salman AND Khan, Fahad Shahbaz},
title = { { How Good Are Foundation Models in Longitudinal MRI Disease Progression Reasoning? } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16887},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper introduces the Time-Aware MRI Benchmark, a comprehensive evaluation framework for assessing vision-language models (VLMs) on longitudinal MRI disease progression reasoning.
- 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.Most medical VLM benchmarks focus on single-timepoint, single-view analysis. This paper compares different disease categories across multiple VLMs, incorporating longitudinal reasoning. 2.This paper uses questions that have been validated by experts.
- 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 main concern is the question generation process. The questions are generated by GPT, but details on how the questions, options, steps, and answers are produced are missing. 2.All results are evaluated using zero-shot inference; results with adapters or fine-tuning should also be reported to demonstrate potential improvements and indicate the quality of the generated Q&A.
- 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?
This paper proposes a benchmark for longitudinal temporal reasoning in medical VLM evaluation. The benchmark is large scale, clinically diverse, and well designed; results from fine tuned models would further demonstrate the quality of the proposed benchmark.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
Authors propose a new benchmark for MRI called Time-Aware Multi-View MRI Benchmark which aims to develop a framework to test models’ capabilities of temporal and spatial reasoning. It includes multiple datasets and multiple pathologies. Main contribution is to include clinically grounded tasks such as interval change localization and temporal sequence ordering within cross-plane acquisition. They also test 17 VLMs on their benchmark and make conclusions about results like poor geometric reasoning of the models.
- 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.
- They address an important issue in clinical radiology AI which is temporal understanding and to have the ability of looking at previous scans and images, and it is natural to expect more models being developed to have a better understanding of temporal and spatial signal. This benchmark can serve as framework to test these capabilities.
- It includes multiple datasets and multiple pathologies and tasks.
- Authors do an extensive test 17 VLMs on their benchmark and make implications and comments about the results.
- 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.
Even though authors propose an approach to use slices from different planes, I believe narrowing the benchmark to 2D is a limitation. As the MRI images are already in 3D and there has been models emerging for 3D, having a full 3D framework can serve as an important baseline.
- 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
I am not sure if the robust percentile-based intensity normalization (2-98%) would be the best way to do normalization, which may create inconsistencies of the image appearance with cases of hyper and hypointensities.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I believe this paper would interest many people in MICCAI community, as the trends are going further for 3D imaging and VLMs. This paper has a potential to drive this forward by proposing a new benchmark to test model capabilities of clinical duties such as temporal reasoning.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The release of this benchmark represents a valuable and timely contribution to the field, especially in the domain of longitudinal MRI disease progression. It introduces a comprehensive framework for evaluating temporal–spatial reasoning in longitudinal, multi-view MRI interpretation and benchmarks the performance of 17 vision-language models.
- 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 benchmark comprises over 3,200 longitudinal MRI timepoints across diverse diseases with radiologist-style questions, requiring temporal–spatial reasoning, and benchmarks the performance of 17 vision-language models.
- 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.
Major： 1) Since this is a benchmark study, it would be beneficial to include a comprehensive summary table that clearly presents all key information, rather than relying solely on textual descriptions. This table could include details such as MRI modalities, age at scan, number of slices per MRI, number of time points, number of diseases, number of questions, and types of questions. While these aspects are mentioned in the text, a structured table would provide a clearer and more accessible overview for readers. 2) Are there any annotation discrepancies, and if so, how are they addressed? 3) The question design would benefit from further illustration. Although Fig. 3 provides three examples, including complete examples covering different question types would help readers better understand the range, diversity, and difficulty of the questions. 4) While evaluation metrics are provided across the five tasks, further clarification is needed regarding the question types for each task, as well as the rationale for selecting specific metrics. 5) The results exhibit substantial variation across tasks. A discussion of failure cases, as well as differences in model performance across tasks, would provide more insightful analysis rather than simply summarizing the results. I understand that this may be due to space limitations; however, I believe the authors should provide a comprehensive supplementary file when submitting the paper.
Minor: Texts in figure 1 are too small.
- 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 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?
Overall, the release of this benchmark represents a valuable and timely contribution to the field, especially in this Longitudinal MRI Disease Progression domain. However, the presentation of the benchmark itself requires further clarification (see above weakness). The illustration of the dataset and task design is not sufficiently detailed, and the rationale behind the chosen evaluation metrics across tasks remains unclear. The results analysis would benefit from deeper discussion.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We sincerely thank the reviewers and meta-reviewer for their thoughtful and constructive feedback. We are pleased to address the key points raised.
[On Question Generation Process (R1, Meta-reviewer)] We acknowledge that additional detail on the QA generation pipeline would strengthen the paper. In the camera-ready, we will expand Section 2.3 to clarify that GPT-5 generated candidate questions conditioned on temporal image sequences and structured clinical metadata (inter-scan intervals, sequence identifiers, RANO annotations where available). Critically, every generated QA pair underwent independent dual review by two board-certified radiologists who verified diagnostic accuracy, temporal consistency, and clinical relevance with a 72% acceptance rate. The radiologist review process ensures that any stylistic patterns from the generation model do not introduce clinically inaccurate content.
[On Fine-tuned Model Results (R1)] This is a valuable suggestion. Our benchmark is designed as a zero-shot evaluation framework to assess out-of-the-box temporal reasoning capabilities of current VLMs. Fine-tuning experiments are an important future direction that we will explicitly discuss in the camera-ready, framing them as a natural next step the benchmark enables.
[On 2D vs 3D Limitation (R2)] We agree that native 3D VLM evaluation is an important frontier. As noted in Section 2.2, mainstream VLMs are currently constrained to 2D inputs. Our multi-view (axial/coronal/sagittal) design deliberately approximates radiological 3D reasoning within existing architectural constraints. We will strengthen the discussion of this limitation and position 3D-native evaluation as explicit future work.
[On Intensity Normalization (R2)] Percentile-based normalization (p2–p98) with sequence-specific windowing was chosen to robustly handle outlier intensities common in multi-site longitudinal data. We acknowledge the concern regarding hyper/hypointense cases and will add clarification that sequence-specific windowing parameters were tuned per cohort to preserve clinically relevant signal ranges.
[On Annotation Discrepancies (R3)] Inter-radiologist disagreements were resolved through a structured adjudication protocol: flagged cases were revised by the generation pipeline and resubmitted for dual review, or discarded. We will make this explicit in the camera-ready.
We appreciate the reviewers recognizing this as a timely and valuable contribution to longitudinal MRI reasoning evaluation. We are committed to delivering a polished camera-ready that addresses the raised concerns using the half-page extension.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
This paper proposes a benchmark for evaluation of VLM in the context of longitudinal disease progression with MR which is being praised as a timely and important contribution across all reviewers. The evaluation is performed on a large dataset using many models, different tasks and pathologies. There are aspects of the paper that would need some polishing in terms notably of the clarification of the question generation and their repartition across tasks. It would be beneficial to include results of variation across tasks instead of only including a summary.