List of Papers Browse by Subject Areas Author List
Abstract
Cancer screening benefits greatly from longitudinal and multimodal data. However, most public CT datasets are single time-point and image-only. They lack longitudinal follow-up, radiology reports, pathology results, and clinical variables. This makes them poorly suited to screening, where cancers must be detected early at a low rate of false positives per scan. We present CancerVerse, the first large-scale open-source longitudinal and multimodal dataset for multicancer screening. It contains 24,168 CT scans from 13,863 patients, with voxel-wise tumor annotations for 13 cancer types in the pelvis, abdomen, and chest. Each cancer patient has 1 to 32 longitudinal CT scans, paired with radiology reports, pathology results, and clinical variables. A 8,799-patient verified healthy cohort with more than one year of follow-up enables realistic specificity estimation. We formulate screening as tumor detection and segmentation, and evaluate sensitivity at fixed false positives (FP) per scan. Models trained on longitudinal and multimodal data outperform those trained on single time-point, image-only data, with a 9% absolute sensitivity gain at 0.5 FP/scan. These results show that effective cancer screening depends on modeling the full longitudinal patient trajectory with multimodal data, rather than isolated imaging snapshots.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0562_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
https://huggingface.co/datasets/BodyMaps/CancerVerse
BibTex
@InProceedings{LiWen_CancerVerse_MICCAI2026,
author = { Li, Wenxuan AND Bassi, Pedro R. A. S. AND Zhou, Xinze AND Wasserthal, Jakob AND Hamamci, Ibrahim E. AND Er, Sezgin AND Menze, Bjoern H. AND Akan, Gulhan E. AND Płotka, Szymon AND Prządo, Jakub AND Chen, Qi AND Tang, Yucheng AND Xu, Daguang AND Sitek, Arkadiusz AND Wang, Kang AND Yang, Yang AND Yuille, Alan L. AND Zhou, Zongwei},
title = { { CancerVerse: A Fully Open Longitudinal and Multimodal Dataset for Multicancer Screening } },
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
The authors present CancerVerse, which they claim to be the first large-scale, open-source longitudinal and multimodal dataset designed specifically for multicancer screening. The dataset includes 23,742 3D CT scans from 9,316 patients, with voxel-wise annotations for 13 cancer types across the pelvis, abdomen, and chest. Importantly, it moves beyond standard single time-point diagnostic datasets by incorporating longitudinal patient trajectories (1–19 scans per patient), paired with non-imaging data such as radiology reports, clinical variables, and pathology results. Furthermore, it introduces a verified healthy cohort with over one year of follow-up, enabling realistic evaluation of false-positive rates. The authors also provide a benchmark showing that a longitudinal and multimodal model achieves approximately a 9% absolute improvement in sensitivity at 0.5 false positives per scan compared to image-only baselines.
- 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.
Addressing a Critical Dataset Gap: Most existing public CT datasets (e. g. , KiTS, LiTS, PanTS) are single-organ and single time-point, typically capturing tumors at confirmed (often later) stages. CancerVerse advances the field by providing longitudinal data and voxel-wise annotations for 13 cancer types, including nine types that previously lacked public CT datasets with voxel-wise annotations.
Clinically Realistic Evaluation Design: By formulating evaluation around fixed false positives per scan (FROC analysis) rather than standard ROC curves, the benchmark better reflects real clinical constraints in screening settings, where false positives are costly.
Multimodal Integration: The dataset incorporates radiology reports, pathology, and clinical variables—features that are rare in public datasets. The use of report-based supervision is a practical design choice that may improve scalability.
Commitment to Open Science: Releasing the dataset under a CC BY-NC 4.0 license, along with standardized formats and trained segmentation backbones, will provide a valuable resource for the MICCAI 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.
Missing Data Dynamics: While the authors acknowledge incomplete multimodal data availability, the paper lacks detailed missingness statistics. Providing availability percentages for clinical variables and laboratory tests across the 9,316 patients would help assess the dataset’s practical usability.
Scope of Longitudinal Modeling: The dataset’s longitudinal structure is a key strength. While the current work evaluates detection at different time points, extending the benchmark to temporal forecasting tasks—such as disease progression modeling or future cancer prediction—would better demonstrate its full potential.
Limited Methodological Novelty: The modeling framework combines established techniques, including temporal aggregation and report-based supervision. As such, the primary contribution lies in dataset curation and benchmarking. The authors may consider emphasizing this more clearly rather than positioning the work as an algorithmic innovation.
Modality Limitations: The dataset is primarily CT-based. Future extensions could incorporate additional imaging modalities (e. g. , MRI or PET) to support more comprehensive multiparametric screening.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 presents a potentially impactful dataset contribution (CancerVerse), addressing a clear gap with longitudinal, multimodal, and multicancer data, along with a clinically realistic evaluation protocol. This resource is likely to be valuable to the MICCAI community.
However, the methodological novelty is limited, and key details (e.g., multimodal fusion, missing data statistics) are insufficiently described. Additionally, the longitudinal aspect is under-explored beyond detection, and the dataset remains CT-only.
Overall, the dataset and benchmark justify weak acceptance, but the limited technical depth and analysis place the paper only marginally above the acceptance threshold.
- 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
Review #2
- Please describe the contribution of the paper
This paper presents CancerVerse, a large-scale, fully open, longitudinal and multimodal CT dataset for multicancer screening. The dataset contains 23,742 3D CT scans from 9,316 patients, paired with voxel-wise tumor annotations for 13 cancer types, radiology reports, pathology results, clinical variables, laboratory tests, and a verified healthy cohort with more than one year of follow-up. Beyond the dataset release, the paper formulates screening as multicancer tumor detection/segmentation under fixed false positives per scan and benchmarks image-only, longitudinal-only, multimodal-only, and longitudinal+multimodal models. The results suggest that incorporating longitudinal trajectories and non-imaging data improves screening performance, with a reported absolute sensitivity gain of 9% at 0.5 FP/scan over an image-only baseline. Overall, this is an ambitious and timely paper. The dataset appears potentially very valuable to the community, especially given the combination of longitudinal scans, multimodal alignment, verified healthy follow-up, and voxel-wise multicancer annotations. The effort behind curation and annotation is impressive. The benchmarking setup is also better aligned with screening than many prior diagnostic-style evaluations.
- 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.The dataset itself is a substantial contribution. Its scale, longitudinal structure, multimodal alignment, and voxel-wise annotations make it highly compelling and help fill an important gap in public resources for screening-oriented medical AI.
2.The paper also provides a comprehensive benchmark that clearly demonstrates the value of longitudinal and multimodal data for multicancer screening.
3.The validation is thorough and convincing. In particular, the inclusion of a verified healthy cohort and external datasets strengthens the credibility and practical relevance of the findings.
4.The paper is generally well written and well organized, with clear motivation and polished presentation.
- 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 handling of heterogeneous data is insufficiently described. For instance, different patients have varying numbers of CT scans, but the manuscript does not clearly explain how such variable-length longitudinal data are processed. In addition, the feature extraction pipeline for non-imaging data (e.g., reports, pathology, and clinical variables) remains unclear.
2.Benign cases are particularly important in cancer screening, as they are essential for a realistic evaluation of false positives and specificity. It would therefore be helpful for the paper to discuss their composition and role in the paper.
3.Some aspects of the evaluation protocol are unclear. First, it is not clear whether the sensitivity reported in Table 1 is calculated at the patient level or the lesion level. Second, the definition of negative samples for specificity and FP/case evaluation is ambiguous. Are these metrics computed only on the verified healthy cohort, or are patients with other cancer types also treated as negative samples when evaluating a specific cancer type?
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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.
(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?
1.The planned open release of the dataset and trained models would be a valuable resource for the community. 2.The experiments are well designed and include sufficient comparisons. 3.The manuscript is polished and well written.
- 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 rebuttal addresses my main concerns. Overall, these clarifications improve the paper’s clarity and strengthen my confidence in the experimental setup.
Review #3
- Please describe the contribution of the paper
The main contribution of this paper is the publication of CancerVerse, a large-scale longitudinal multi-cancer CT dataset, including images, masks and accompanying data like radiology reports.
- 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 dataset is of unprecedented size, especially considering the accompanying information in the form of segmentation masks, radiology reports, other health related data and its longitudinal nature.
The dataset includes annotations for 13 cancer types, some of which are rare to non-existent in available public datasets.
The authors benchmark different widely adopted segmentation models on their proposed dataset, showing superior performance of a longitudinal model with multimodal information over all the baseline methods.
- 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.
While the proposed dataset is interesting there are several issues I have with this paper:
Dataset Construction: As the dataset is the main contribution of this paper, I would have liked to see more details on the generation of voxel-wise tumor masks, which is a very time-consuming task for this large of a dataset. Many recent large scale datasets and collections struggle with exactly that actually struggle with annotation quality even for easier target structures like organs (e.g. https://huggingface.co/datasets/mrmrx/CADS-dataset, https://github.com/MrGiovanni/AbdomenAtlas). Consequentially, I would have liked to see more details on quality assurance and control and the annotation procedure in general.
Evaluation: The authors train segmentation models on the dataset, however, their main evaluation focuses on sensitivity and specificity, disregarding the main objective of segmentation. Dice is only shown as a mean over all structures in Fig. 3, however, it would be more helpful to see this per structure.
Additionally, it seems like a case is counted as a true positive, if there is any prediction in the image, which might be heavily misleading. Accordingly, the metrics sensitivity at 0.25/0.5 FP/scan don’t really make sense in this setting, as this mixes up instance wise and case wise measurements and is not what is classically reported as FROC in detection metrics.
Interrater evaluation: The comparison against interrater performance should be reported on the same split of the test set. Fig. 3 currently reports this against the performance on the full test set, which could be heavily misleading due to different distributions and frequencies of the tumor classes.
Baseline Models: While the comparison against a large suit of baseline models is appreciated, it is not really sensible for the ULS family of models, which are not developed for fully automatic segmentation but always assume a lesion at the center of the patch. This also raises a question regarding the validity of the results as this should usually result in a significantly higher sensitivity due to this bias of always predicting something.
Overstated claims: The authors claim that they provide the first publicly available annotations for 9 of the 13 tumor types. While this is true for most of the tumor classes, some of these are actually covered by other datasets already, e.g. esophagus and uterus in AbdomenAtlas 2.0 (https://github.com/BodyMaps/AbdomenAtlas2.0).
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
The dataset is very interesting in general. However, I have several issues with the benchmarking on the dataset and am not fully convinced of the quality of the annotations.
- 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 answered most of my questions. I would have liked to be able to inspect the dataset before making a decision but as this is not possible in MICCAI.and the described annotation workflow sounds convincing I am recoomending to accept the paper. However, I would like to recommend submitting similar papers to e.g. the NeurIPS E&D track in the future as this allows reviewers to inspect datasets before accepting the paper.
Author Feedback
We thank the reviewers and meta-reviewer. R1 and R3 recommend acceptance. R4 calls CancerVerse “of unprecedented size… 13 cancer types, some rare to non-existent in public datasets… superior performance over all baselines.” The meta-reviewer calls it “an important contribution to the MICCAI community.”
R1: Q1-Missing Data Dynamics. Availability across 9,316 patients: reports 100%, pathology 58%, clinical variables 64%. We will add this to the camera-ready. Q2-Scope of Longitudinal Modeling. Agreed. We hope the release of CancerVerse can encourage more exploration from our community. Q3-Limited Methodological Novelty. Our paper presents a dataset-and-benchmark contribution; no “novel method” is claimed. The longitudinal-and-multimodal model tested in this paper combines two existing models (longitudinal model, Rokuss, MICCAI’24; multimodal model, Bassi, MICCAI’25). Q4-Modality Limitations. Agreed. Additional imaging modalities (e.g., MRI) are planned for future versions.
R3: Q1-Handling of heterogeneous data. Variable-length CTs are processed by a shared nnU-Net encoder followed by a temporal-difference layer (Rokuss’24). Single-scan patients fall back to the image-only. Reports are encoded by Llama-3-70B with tumor-attribute auxiliary labels (Bassi’25). Clinical variables are encoded by an MLP. Modalities are fused via cross-attention. Q2-Benign cases. Composition: 77% with no cancer and other disease, 23% with only benign lesions (e.g., IPMN). Role: computing specificity and FP/scan. Q3-Evaluation protocol. Sensitivity in Tab. 1 is lesion-level. Negatives are both the verified healthy cohort (>1 year follow-up) and patients with other cancer types.
R4: Q1-Dataset Construction. R4 cites annotation-quality issues in CADS and AbdomenAtlas, but both are organ datasets that heavily used AI-assisted annotation pipelines, acknowledged by their papers. CancerVerse is fundamentally different. We manually annotate voxel-wise tumor masks across 13 cancer types. No AI-generated masks. Every lesion was drawn by at least one human expert. 28 expert annotators. More than 13 months of work. Quality assurance: a representative multicancer subset was annotated independently by three experts to verify consistency (Sec. 3.1). Q2-Evaluation. (i) We will add per-cancer Dice to the camera-ready (liv/kid/pan/col/spl/bla/gal/adr/eso/sto/duo/ute/pros = 67/64/48/41/52/45/46/55/49/38/39/63/57). In cancer screening, sensitivity and specificity matter most; precise contouring is downstream. (ii) This is a misunderstanding. We use instance-wise evaluation for both positive and negative patients, precisely following the classical definition of FROC. That means one case can have more than one false positive instance. Q3-Interrater evaluation. This is an excellent point. We have now reported a mean Dice of 50.8% on the same split. We will update Fig. 3 accordingly. Q4-Baseline Models. ULS models have a known limitation: they require the lesion at the center of the patch (Grauw, MedIA’25). We evaluate using a sliding window on whole CT volumes (the realistic screening setting), where the lesion is usually not at the window center, which is why ULS sensitivity in Tab. 1 is low. We report every baseline at its operating point on the ROC curve (nearest to top-left) for fair comparison. We will add this analysis to the camera-ready. Q5-Overstated claim. We revise “9 novel cancer types” to “7” and cite AbdomenAtlas 2.0. To our knowledge, among publicly released whole-volume abdominal CT datasets, CancerVerse is one of only two exceeding 20,000 first-time-released CT scans (alongside Merlin, n=25,494; Nature’26). It is the only public dataset combining voxel-wise tumor annotations across 13 cancer types with longitudinal trajectories, paired radiology and pathology reports, clinical variables, and a verified healthy follow-up cohort. The full dataset will be released on 09/01 if accepted, ahead of MICCAI’26.We respectfully ask for reconsideration.
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.
This paper proposes a very large-scale multimodal dataset for screening, which represents an important contribution to the MICCAI community. Two reviewers recognized this contribution and gave accept ratings. However, Reviewer 3 raised several concerns regarding dataset construction, especially regarding the tumor annotation, interrater comparison, baseline models, and overstated claims—all of which should be more clearly addressed and illustrated in the paper.
- 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.
This paper introduces a comprehensive longitudinal and multimodal dataset for multi-cancer screening. Given the consensus among reviewers and the immense value of the extensive tumor annotation and dataset construction efforts, I recommend accepting this paper. This work promises to be a highly impactful resource for the MICCAI community.
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 paper presents a potentially impactful open-source dataset contribution (CancerVerse), which includes a large-scale longitudinal, multimodal and multi-cancer screening dataset (23,742 CT scans from 9,316 patients, paired with tumor masks for 13 cancer types, radiology reports, pathology results, clinical variables, laboratory tests). CancerVerse represents a large advancement over the existing public CT datasets, which are mostly single time-point and single-modality. Plus, the tumor masks are annotated manually to ensure its high quality. Considering the scarcity of public mulit-cancer and multi-modal dataset, CancerVerse could be highly valuable in oncology research (cancer screening and diagnosis) for the medical imaging community.
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.
All reviewers agreed on acceptance and acknowledged the value of a large-scale, multi-modal, multi-organ, and multi-timepoint dataset. Although there are still concerns about the methodological novelty and not being able to inspect the dataset before making a decision, I think the contribution overweighs the limitation.