Datasets:

chehablab
/

miniMSD512

dataset_info:
  features:
    - name: organ
      dtype: string
    - name: image
      dtype: image
    - name: binary_mask
      dtype: image
    - name: classes_mask
      dtype: image
    - name: volume_id
      dtype: int32
    - name: slice_id
      dtype: int32
  splits:
    - name: train
      num_bytes: 4888752475.780907
      num_examples: 51891
  download_size: 4611427588
  dataset_size: 4888752475.780907
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
license: cc-by-4.0
task_categories:
  - image-segmentation
language:
  - en
tags:
  - organs
  - medical
  - ct
  - mri
pretty_name: Mini Medical Segmentation Decathlon 244
size_categories:
  - 10K<n<100K

Processed and Reduced Medical Segmentation Decathlon Dataset

The miniMSD dataset is a medical image segmentation benchmark covering 10 human organs. It is derived from the Medical Segmentation Decathlon (MSD) by converting volumetric scans from NIfTI (NII) format into serialised 2D RGB images, alongside their corresponding segmentation masks. The dataset is provided in multiple resolution variants (244 and 512), enabling easier use, off-the-shelf accessibility, and flexible experimentation.

Dataset Details

The dataset covers 10 human body organs, listed below. Each organ includes up to 40 volumes, with each volume consisting of a variable number of image slices. Each dataset entry contains the following components: the organ type, the image, a binary mask, a detailed (multi-class) mask, a volume ID, and a slice ID. The image, binary mask, and detailed mask are all provided as PIL images. The binary mask contains two labels: 0 for background and 1 for the target region. The detailed mask contains multiple labels (0, 1, 2, 3, …), where each label corresponds to a specific anatomical structure. The mapping of label indices to structures is provided below.

Organ	Number of Volumes	Total Slices	Avg. Slices per Volume	% of Total Slices
Prostate	32	1204	37.625	1.26%
Heart	20	2271	113.550	2.38%
Hippocampus	40	2754	68.850	2.89%
HepaticVessel	40	5796	144.900	6.08%
BrainTumour	40	6200	155.000	6.51%
Spleen	40	6964	174.100	7.31%
Pancreas	40	7068	176.700	7.42%
Colon	40	7344	183.600	7.71%
Lung	40	22510	562.750	23.62%
Liver	40	33200	830.000	34.83%

Labels Mapping

BrainTumour

0: background
1: necrotic / non-enhancing tumor
2: edema
3: enhancing tumor

Heart

0: background
1: left atrium

Liver

0: background
1: liver
2: tumor

Hippocampus

0: background
1: anterior
2: posterior

Prostate

0: background
1: peripheral zone
2: transition zone

Lung

0: background
1: nodule

Pancreas

0: background
1: pancreas
2: tumor

HepaticVessel

0: background
1: vessel
2: tumor

Spleen

0: background
1: spleen

Colon

0: background
1: colon

Uses

from datasets import load_dataset
miniMSD244 = load_dataset("chehablaborg/miniMSD244", split="train")

sample_id = 312
organ = miniMSD244[sample_id]["organ"]
image = miniMSD244[sample_id]["image"]
binary_mask = miniMSD244[sample_id]["binary_mask"]
classes_mask = miniMSD244[sample_id]["classes_mask"]
plt.imshow(image, cmap="grey")
plt.show()

Citation

Please mention us in an acknowledgement chehablab.com and cite the original authors of the dataset

@misc{msd2019,
      title={A large annotated medical image dataset for the development and evaluation of segmentation algorithms}, 
      author={Amber L. Simpson and Michela Antonelli and Spyridon Bakas and Michel Bilello and Keyvan Farahani and Bram van Ginneken and Annette Kopp-Schneider and Bennett A. Landman and Geert Litjens and Bjoern Menze and Olaf Ronneberger and Ronald M. Summers and Patrick Bilic and Patrick F. Christ and Richard K. G. Do and Marc Gollub and Jennifer Golia-Pernicka and Stephan H. Heckers and William R. Jarnagin and Maureen K. McHugo and Sandy Napel and Eugene Vorontsov and Lena Maier-Hein and M. Jorge Cardoso},
      year={2019},
      eprint={1902.09063},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/1902.09063}, 
}

License

This work is licensed under a Creative Commons CC BY SA License.

Chehab lab @ 2026