MAGIC-CT: Multiorgan Annotation and Grounded Image Captioning in CT for Cancer

Official evaluation scripts for the MAGIC-CT dataset.

📋 Overview

MAGIC-CT is a comprehensive multimodal dataset for abdominal oncology combining:

• 562 patients with contrast-enhanced CT scans
• ~1,250 annotated lesions across 8 pathologies (4 organs)
• 4,937 organ descriptions in 3 languages (EN/RU/KZ)
• Expert-validated 3D segmentation masks
• Rich clinical narratives by radiologists

Pathologies: Liver cancer (HCC), renal cancer, lung cancer, pancreatic cancer, liver cysts, kidney cysts, lung metastases, liver metastases

📊 Dataset

Download from Zenodo: https://zenodo.org/uploads/18389015

Expected structure after extraction:

data/
├── scans/
│   ├── liver_cancer/
│   │   ├── patient_001.nrrd
│   │   └── ...
│   ├── kidney_cyst/
│   ├── renal_cancer/
│   └── ...
└── segmentations/
    ├── liver_cancer/
    │   ├── patient_001.nrrd
    │   └── ...
    └── ...

🚀 Installation

# Clone repository
git clone https://github.com/maxtrubetskoy/MagicCT.git
cd MagicCT

# Install dependencies
pip install -r requirements.txt

Requirements: Python 3.8+, PyTorch 1.12+, MONAI 1.2+, CUDA 11.3+ (for GPU)

⚡ Quick Start

Evaluate Single Model

python evaluate.py --model swinunetr --data_dir data/ --output results/

Evaluate All Models (Reproduce Paper Results)

python evaluate.py --model all --data_dir data/ --output results/ --device cuda

With GPU Acceleration

python evaluate.py --model all --data_dir data/ --device cuda

Specific Cancer Types

python evaluate.py --model swinunetr --data_dir data/ \
    --cancer_types liver_cancer lung_cancer renal_cancer

Save Prediction Masks

python evaluate.py --model swinunetr --data_dir data/ --save_predictions

📋 Command-Line Options

python evaluate.py --help

Main arguments:

• --model: Model name (swinunetr, unetr, segresnet, dynunet, or all)
• --data_dir: Path to dataset directory (containing scans/ and segmentations/)
• --output: Output directory for results (default: results/)
• --cancer_types: Specific cancer types to evaluate (default: all 8 types)
• --device: Device (auto, cuda, or cpu; default: auto-detect)
• --batch_size: Batch size (default: 1)
• --workers: Number of data loading workers (default: 4)
• --roi_size: ROI size for input (default: 96 96 96)
• --save_predictions: Save prediction masks as .nrrd files

📈 Expected Results

Table 2 from Paper - Overall Model Performance:

Model	Dice (%)	HD95 (mm)	Sensitivity (%)	Specificity (%)	Inference Time (s)	Params (M)
SwinUNETR	72.3 ± 2.1	8.2 ± 2.1	75.9 ± 5.1	98.9 ± 0.7	2.8	61.9
UNETR	68.7 ± 2.3	9.4 ± 2.8	73.4 ± 4.6	99.2 ± 0.6	2.5	102.4
SegResNet	65.2 ± 1.9	11.2 ± 3.1	67.3 ± 5.6	98.7 ± 0.8	1.2	15.7
DynUNet	67.1 ± 2.0	10.1 ± 2.9	69.6 ± 5.7	99.4 ± 0.5	1.8	22.3

Table 3 from Paper - Cancer-Type Specific Performance (Dice %):

Cancer Type	SwinUNETR	UNETR	SegResNet	DynUNet
Benign Lesions
Liver Cysts	84.3 ± 3.2	80.7 ± 3.8	76.9 ± 4.1	79.3 ± 4.5
Kidney Cysts	81.0 ± 3.8	77.6 ± 4.2	73.6 ± 4.6	76.0 ± 4.8
Primary Malignancies
Hepatocellular Carcinoma	78.1 ± 4.2	74.5 ± 4.6	70.5 ± 5.1	72.8 ± 5.3
Lung Cancer	74.9 ± 4.8	71.4 ± 5.2	67.7 ± 5.5	69.6 ± 5.7
Renal Cancer	71.6 ± 5.1	68.0 ± 5.4	64.4 ± 5.8	66.2 ± 6.0
Pancreas Cancer	53.1 ± 7.8	49.3 ± 8.1	46.8 ± 8.4	47.9 ± 8.6
Metastatic Disease
Lung Metastases	63.0 ± 6.2	59.4 ± 6.8	56.5 ± 7.1	57.9 ± 7.4

🔗 Model Weights

Models use MONAI framework with pretrained weights:

SwinUNETR (Recommended for best results):

• Download: swin_unetr.base_5000ep_f48_lr2e-4_pretrained.pt
• Place in: pretrained/ folder
• Pre-trained on 5,050 CT scans with self-supervised learning

Other models (UNETR, SegResNet, DynUNet):

• Use MONAI's built-in architectures
• No additional weights required

# Optional: Download SwinUNETR pretrained weights for better results
mkdir -p pretrained
cd pretrained
wget https://github.com/Project-MONAI/MONAI-extra-test-data/releases/download/0.8.1/swin_unetr.base_5000ep_f48_lr2e-4_pretrained.pt
cd ..

📁 Output Structure

results/
├── swinunetr/
│   ├── results.csv       # Per-case metrics (Dice, HD95, sensitivity, specificity)
│   ├── summary.json      # Aggregate statistics (mean ± std)
│   └── predictions/      # (optional) .nrrd prediction masks
├── unetr/
├── segresnet/
├── dynunet/
└── comparison.csv        # Model comparison table

🔬 Technical Details

Dataset Statistics:

• Total patients: 562 (492 with reports)
• Age: 63 ± 14 years (range: 19-92)
• Gender: 52% male, 46% female
• Imaging: Contrast-enhanced CT (Philips Ingenuity)
• Contrast agents: Ultravist 370, Gadovist
• Radiation dose: 7-15 mSv

Annotations:

• Annotators: 7 radiologists
• Inter-annotator agreement: Cohen's κ = 0.74 ("Substantial")
• Software: 3D Slicer
• Format: .nrrd (NRRD format with spatial metadata)

📖 Citation

If you use this dataset or code, please cite:

@article{popov2025magicct,
  title={{MAGIC-CT}: Multiorgan Annotation and Grounded Image Captioning in {CT} for Cancer},
  author={Popov, Maxim and Iklassov, Zangir and Baimagambet, Zhanas and Jakipov, Murat and Andreyeva, Xeniya and Akhtar, Muhammad and Tak\'{a}\v{c}, Martin and Jamwal, Prashant},
  year={2026},
  doi={10.5281/zenodo.17549293},
  url={https://zenodo.org/uploads/18389015}
}

👥 Authors

Corresponding Author: Maxim Popov (maxim.popov@nu.edu.kz)

1. Maxim Popov - Nazarbayev University, Kazakhstan
2. Zangir Iklassov - Mohamed bin Zayed University of AI (MBZUAI), UAE
3. Zhanas Baimagambet - Nazarbayev University, Kazakhstan
4. Murat Jakipov - National Research Oncology Center (NROC), Kazakhstan
5. Xeniya Andreyeva - National Research Oncology Center (NROC), Kazakhstan
6. Muhammad Akhtar - Nazarbayev University, Kazakhstan
7. Martin Takáč - Mohamed bin Zayed University of AI (MBZUAI), UAE
8. Prashant Jamwal - Nazarbayev University, Kazakhstan (Team Lead)

🙏 Acknowledgments

This work was supported by:

• Collaborative Research Program of Nazarbayev University (Grant No. 111024CRP2007)
• National Research Oncology Center (NROC), Astana, Kazakhstan
• MONAI Consortium for framework support

📄 License

• Code: MIT License (see LICENSE)
• Dataset: CC0 1.0 Universal (Public Domain)

🐛 Issues & Questions

• Issues: GitHub Issues
• Email: zangir.iklassov@mbzuai.ac.ae
• Dataset: https://zenodo.org/uploads/18389015

🔗 Links

• Paper: Coming soon (under review)
• Dataset: https://zenodo.org/uploads/18389015
• Code: https://github.com/maxtrubetskoy/MagicCT
• MONAI: https://monai.io

---

Last Updated: February 2026