Angelou0516/FLARE22-CT-Abd

FLARE22 CT Abdominal Organ Segmentation Dataset (Preprocessed)

Dataset Description

This is a preprocessed version of the FLARE22 (Fast and Low-resource semi-supervised Abdominal oRgan sEgmentation) dataset, optimized for SAM/SAM2 evaluation.

• Original Dataset: FLARE22 on Zenodo
• Preprocessing: CT windowing, resizing to 1024x1024, multi-organ masks
• Format: RGB images (PNG) + grayscale masks with organ class IDs

Dataset Statistics

• Splits: train
• Total Samples: 4543 CT slices
• Image Size: 1024 x 1024 pixels
• Organs: 13 classes (liver, kidneys, spleen, pancreas, aorta, etc.)

Organ Classes

Class ID	Organ Name (English)	Organ Name (中文)
0	Background	背景
1	Liver	肝脏
2	Right Kidney	右肾
3	Spleen	脾脏
4	Pancreas	胰腺
5	Aorta	主动脉
6	Inferior Vena Cava	下腔静脉
7	Right Adrenal Gland	右肾上腺
8	Left Adrenal Gland	左肾上腺
9	Gallbladder	胆囊
10	Esophagus	食管
11	Stomach	胃
13	Left Kidney	左肾

Note: Class 12 (Duodenum/十二指肠) is removed by default due to scattered distribution.

Usage

Load with 🤗 Datasets

from datasets import load_dataset

# Load dataset
dataset = load_dataset("Angelou0516/FLARE22-CT-Abd", split="val")

# Access sample
sample = dataset[0]
image = sample['image']  # PIL Image (RGB)
mask = sample['label']   # PIL Image (grayscale with class IDs)
case_id = sample['case_id']

Use with EasyMedSeg

from dataloader.image.flare22_dataset import FLARE22Dataset

# Will automatically download from HuggingFace
dataset = FLARE22Dataset(
    mode='download',
    split='val',
    hf_repo_id='Angelou0516/FLARE22-CT-Abd'
)

sample = dataset[0]
image = sample['image']  # PIL Image
mask = sample['mask']    # numpy array

Evaluation with SAM/SAM2

# Clone EasyMedSeg
git clone https://github.com/your-org/EasyMedSeg.git
cd EasyMedSeg

# Install dependencies
pip install -r requirements.txt

# Run evaluation
python scripts/evaluation.py --config configs/test_flare22_image.yaml

Preprocessing Details

The original FLARE22 NII.gz CT volumes were preprocessed with:

1. CT Windowing: Window level = 40, width = 400 (abdominal soft tissue)
2. Normalization: Pixel values normalized to [0, 1] range
3. Resizing: Bilinear interpolation to 1024x1024
4. Mask Processing:
   • Nearest-neighbor interpolation (preserves class IDs)
   • Small object removal (< 100 pixels in 2D, < 1000 in 3D)
5. RGB Conversion: Grayscale CT converted to 3-channel RGB

Citation

FLARE22 Dataset

@article{ma2022fast,
  title={Fast and Low-GPU-memory abdomen CT organ segmentation: The FLARE challenge},
  author={Ma, Jun and others},
  journal={Medical Image Analysis},
  year={2022}
}

SAM

@article{kirillov2023segment,
  title={Segment Anything},
  author={Kirillov, Alexander and others},
  journal={ICCV},
  year={2023}
}

SAM2

@article{ravi2024sam2,
  title={SAM 2: Segment Anything in Images and Videos},
  author={Ravi, Nikhila and others},
  journal={arXiv preprint arXiv:2408.00714},
  year={2024}
}

License

This preprocessed dataset inherits the license from the original FLARE22 dataset.

• Original License: CC BY 4.0
• Usage: Research and educational purposes
• Attribution: Please cite the original FLARE22 paper when using this dataset

Links

• 🏥 FLARE22 Challenge
• 📊 Original Dataset
• 🤖 EasyMedSeg Framework
• 📝 Dataset Card

Acknowledgments

• FLARE22 Challenge Organizers
• Original dataset contributors
• EasyMedSeg framework developers