| | --- |
| | license: cc0-1.0 |
| | language: |
| | - en |
| | - ru |
| | - kz |
| | pretty_name: "MAGIC-CT: Multiorgan Annotation and Grounded Image Captioning in CT for Cancer" |
| | tags: |
| | - medical-imaging |
| | - computed-tomography |
| | - cancer-detection |
| | - 3d-segmentation |
| | - multimodal |
| | - radiology |
| | --- |
| | |
| | # MAGIC-CT: Multiorgan Annotation and Grounded Image Captioning in CT for Cancer |
| |
|
| | ## Description |
| |
|
| | **MAGIC-CT** is a comprehensive, multimodal dataset designed to advance artificial intelligence in abdominal oncology. It addresses the critical need for resources that bridge the gap between radiological imaging and clinical language by pairing high-resolution 3D Computed Tomography (CT) scans with expert-authored radiology reports. |
| |
|
| | The dataset includes 562 patients with various abdominal tumors, covering 7 distinct pathologies across 4 organs: |
| | * **Liver:** Liver Cysts, Liver Cancer (Hepatocellular Carcinoma) |
| | * **Lungs:** Lung Metastases, Lung Cancer |
| | * **Kidneys:** Kidney Cysts, Renal Cancer |
| | * **Pancreas:** Pancreatic Cancer |
| |
|
| | Each patient case features at least one annotated CT scan with detailed 3D segmentation masks that delineate tumor boundaries and key anatomical structures. These visual data are paired with a radiologist-authored report that provides a comprehensive narrative of organ-specific findings and the overall abdominal status. With over 1,250 annotated lesions and 850 textual findings, MAGIC-CT offers the granular spatial-textual alignment required for training robust multimodal AI systems. |
| |
|
| | This resource enables significant advancements in AI-driven tumor characterization, automated report generation, and metastasis tracking, with profound implications for the future of precision oncology. |
| |
|
| | ## Dataset Access |
| |
|
| | The full MAGIC-CT dataset is hosted on Google Drive due to its large size. You can access and download it using the following link: |
| |
|
| | **[Download the MAGIC-CT Dataset from Google Drive](https://drive.google.com/drive/folders/1SwR7A6OkW1DsgjSdcZTtmxBsa0lKSbpH?usp=drive_link)** |
| |
|
| | ### Dataset Structure |
| |
|
| | The dataset is organized into three main folders: `scans`, `segmentations`, and `reports`. |
| |
|
| | * `/scans`: Contains all imaging data in anonymized `.nrrd` format. |
| | * `/segmentations`: Contains the corresponding 3D segmentation masks, also in `.nrrd` format. |
| | * `/reports`: Contains structured `.json` files with patient metadata and narrative reports in English (`en`), Russian (`ru`), and Kazakh (`kz`). |
| |
|
| | An example of the `.json` report structure is as follows: |
| | ```json |
| | { |
| | "encrypted_patient_id": "abc12345", |
| | "age": 65, |
| | "gender": "male", |
| | "screening_date": "2023-04-17", |
| | "ru": { |
| | "liver": "...", |
| | "pancreas": "...", |
| | "kidneys": "..." |
| | }, |
| | "kz": { |
| | "liver": "...", |
| | "pancreas": "...", |
| | "kidneys": "..." |
| | }, |
| | "en": { |
| | "liver": "The liver has a homogeneous structure; no focal lesions detected.", |
| | "pancreas": "Mass identified in the pancreatic body, measuring 5.3×3.5 cm...", |
| | "kidneys": "Small cortical cysts in both kidneys, no hydronephrosis." |
| | } |
| | } |
| | ``` |
| |
|
| | ## Citation Information |
| |
|
| | If you use the MAGIC-CT dataset in your research, please cite our paper: |
| |
|
| | ```bibtex |
| | @article{TBA} |
| | ``` |
| |
|
| | --- |
| |
|
| | *This dataset was developed by the Department of Computer Science, School of Engineering and Digital Sciences at Nazarbayev University, Astana, Kazakhstan.* |
