| --- |
| license: cc-by-nc-nd-4.0 |
| pretty_name: "CancerVerse" |
| size_categories: |
| - 10K<n<100K |
| tags: |
| - medical-imaging |
| - computed-tomography |
| - oncology |
| - cancer |
| - radiology-reports |
| - longitudinal |
| - multimodal |
| - vision-language |
| task_categories: |
| - image-classification |
| - image-segmentation |
| - image-to-text |
| --- |
| |
| # 🩻 CancerVerse |
|
|
| ### The first **longitudinal**, **multimodal** CT dataset spanning **13 malignant tumor types** |
|
|
| CancerVerse pairs whole-body abdominal/pelvic **CT volumes** with the **radiologists' own free-text reports**, follows patients **across time**, and is being released in stages — culminating in **expert voxel-level tumor masks** and a deep **clinical & longitudinal annotation layer**. It is built to power the next generation of cancer-aware medical AI: tumor detection and segmentation, report generation and grounding, longitudinal disease modeling, and treatment-response prediction. |
|
|
| > **24,168 CT volumes · 13,863 patients · 24,168 paired radiology reports · 15 years of follow-up · 13 cancer types — fully de-identified.** |
|
|
| CancerVerse is built and maintained by the [BodyMaps](https://huggingface.co/BodyMaps) team. |
|
|
| --- |
|
|
| ## 📊 CancerVerse by the Numbers |
|
|
| | | | |
| |---|---| |
| | 🧊 **24,168** | de-identified 3D **CT volumes** | |
| | 🧑🤝🧑 **13,863** | unique **patients** | |
| | 📝 **24,168** | paired **radiology reports** | |
| | 🗓️ **15 years** | of imaging history (**Jul 2010 → Jun 2025**) | |
| | 🔁 **up to 26** | imaging **timepoints** for a single patient | |
| | 🎚️ **4** | CT **contrast phases** (portal-venous, arterial early/late, non-contrast …) | |
| | 🎯 **13** | **malignant tumor types** covered | |
|
|
| --- |
|
|
| ## 🚀 Staged Release — what's ready today, what's coming next |
|
|
| CancerVerse is continuously evolving. We are releasing it in stages so the community can start building today while the richest layers are finalized. |
|
|
| ### ✅ Available now |
|
|
| - **De-identified CT volumes** — 24,168 3D scans from 13,863 patients, across multiple scanners and **4 contrast phases**. |
| - **Paired radiology reports** — the radiologist's full free-text impression for 24,168 scans. Real clinical language, ready for report generation, information extraction, and image–text pre-training. |
| - **Scan-level context** — sex, age, exam date, contrast phase, scanner, and study/series/protocol descriptions for every volume. |
| - **Longitudinal linkage** — stable de-identified patient IDs + exam dates let you reconstruct each patient's imaging timeline (up to **26 timepoints**, spanning up to **14.4 years**). |
|
|
| ### 🔜 Coming soon |
|
|
| - 🧬 **Structured tumor features** *(by August 2026)* — per-organ malignancy status, largest-lesion **diameters, locations, attenuation**, and **tumor staging**. |
| - 🔬 **Pathology results** *(by August 2026)* — ICD-10 codes. |
| - 🎯 **Expert manual segmentation annotations** *(by September 2026)* — voxel-level, radiologist-verified **tumor** annotations. |
| - 📋 **Clinical & longitudinal layer** *(by September 2026)* — cancer histology, TNM stage, treatments received, surgical resections, **RECIST treatment response**, metastatic sites, disease-status narratives, follow-up recommendations, comorbidities, and devices/implants. |
|
|
| > 💡 *Cancer information is already in your hands today — it lives in the paired radiology reports. The structured, machine-readable labels and masks make it turnkey, and arrive in the next releases.* |
|
|
| --- |
|
|
| ## 🎯 The 13 malignant tumor types |
|
|
| CancerVerse is curated around 13 abdominal and pelvic malignancies: |
|
|
| `liver` · `kidney` · `pancreas` · `colon` · `adrenal gland` · `bladder` · `duodenum` · `esophagus` · `gallbladder` · `prostate` · `spleen` · `stomach` · `uterus` |
|
|
| --- |
|
|
| ## 📦 What's inside this release |
|
|
| Each row is **one CT volume**, paired with that scan's radiology report and acquisition context. |
|
|
| | Field | Description | |
| |---|---| |
| | `CancerVerse ID` | Unique de-identified identifier for the CT volume (one per scan). | |
| | `Patient ID` | De-identified patient identifier — **links a patient's scans across time**. | |
| | `Encrypted Accession Number` | De-identified study/visit identifier. | |
| | `sex` | Patient sex (M / F). | |
| | `age` | Patient age at the time of the scan (years). | |
| | `exam_date` | Date of the CT exam (enables longitudinal ordering). | |
| | `phase` | Contrast phase (e.g., portal-venous, arterial early/late, non-contrast). | |
| | `scanner` | Scanner model / vendor. | |
| | `study_description`, `series_description`, `protocol_name` | Acquisition protocol details. | |
| | `report` | **Free-text radiology report**. | |
|
|
| The CT volumes themselves are provided alongside this table and are linked by `CancerVerse ID`. |
|
|
| --- |
|
|
| ## ⬇️ Download |
|
|
| ### 1. Set up the environment |
|
|
| ```bash |
| conda create -n huggingface python=3.11 -y |
| conda activate huggingface |
| python -m pip install --upgrade pip |
| python -m pip install --upgrade "huggingface_hub[cli]" |
| hf --version |
| hf auth login |
| ``` |
|
|
| ### 2. Download CancerVerse |
|
|
| ```bash |
| mkdir CancerVerse |
| cd CancerVerse |
| bash download_cancerverse.sh |
| ``` |
|
|
| --- |
|
|
| ## ✨ Why CancerVerse? |
|
|
| - **Truly longitudinal.** Most public CT datasets are single snapshots. CancerVerse follows real patients over **up to 14.4 years and 26 timepoints**, enabling disease-progression and treatment-response research. |
| - **Multimodal from day one.** Every scan ships with its **radiologist report** — pixels *and* clinical language — ideal for vision-language pre-training, report generation, and grounded detection. |
| - **Cancer-centric.** Purpose-built around **13 malignant tumor types**, with expert masks and structured oncology labels on the way. |
| - **Real-world scale & diversity.** 13,863 patients, multiple scanners, and 4 contrast phases — clinical heterogeneity, not a sanitized benchmark. |
| - **Responsibly de-identified.** No protected health information; stable de-identified IDs preserve patient timelines for research. |
|
|
| --- |
|
|
| ## 📜 License |
|
|
| CancerVerse is released under **CC BY-NC-ND 4.0**. Please review the license before use. |
|
|
| --- |
|
|
| ## 📝 Citation |
|
|
| A dedicated CancerVerse data paper is forthcoming. In the meantime, if you use CancerVerse, please cite: |
|
|
| ```bibtex |
| @inproceedings{li2024well, |
| title={How well do supervised models transfer to 3d image segmentation}, |
| author={Li, Wenxuan and Yuille, Alan and Zhou, Zongwei}, |
| booktitle={The Twelfth International Conference on Learning Representations}, |
| volume={1}, |
| year={2024} |
| } |
| |
| @article{li2024abdomenatlas, |
| title={AbdomenAtlas: A large-scale, detailed-annotated, \& multi-center dataset for efficient transfer learning and open algorithmic benchmarking}, |
| author={Li, Wenxuan and Qu, Chongyu and Chen, Xiaoxi and Bassi, Pedro RAS and Shi, Yijia and Lai, Yuxiang and Yu, Qian and Xue, Huimin and Chen, Yixiong and Lin, Xiaorui and others}, |
| journal={Medical Image Analysis}, |
| volume={97}, |
| pages={103285}, |
| year={2024}, |
| publisher={Elsevier} |
| } |
| ``` |
|
|
| --- |
|
|
| ## 📣 Stay tuned |
|
|
| CancerVerse is actively growing — **masks and structured oncology labels are coming next.** ⭐ Follow [BodyMaps on Hugging Face](https://huggingface.co/BodyMaps) and watch this dataset to be notified when each release lands. |
|
|
