Medical image datasets#

TorchIO offers tools to easily download publicly available datasets from different institutions and modalities.

The interface is similar to torchvision.datasets.

If you use any of them, please visit the corresponding website (linked in each description) and make sure you comply with any data usage agreement and you acknowledge the corresponding authors’ publications.

If you would like to add a dataset here, please open a discussion on the GitHub repository:

Discuss

IXI#

The Information eXtraction from Images (IXI) dataset contains “nearly 600 MR images from normal, healthy subjects”, including “T1, T2 and PD-weighted images, MRA images and Diffusion-weighted images (15 directions)”.

Note

This data is made available under the Creative Commons CC BY-SA 3.0 license. If you use it please acknowledge the source of the IXI data, e.g. the IXI website.

`IXI`#

class torchio.datasets.ixi.IXI(root: str | Path, transform: Transform | None = None, download: bool = False, modalities: Sequence[str] = ('T1', 'T2'), **kwargs)[source]#

Bases: SubjectsDataset

Full IXI dataset.

Parameters:

root – Root directory to which the dataset will be downloaded.
transform – An instance of Transform.
download – If set to True, will download the data into root.
modalities – List of modalities to be downloaded. They must be in ('T1', 'T2', 'PD', 'MRA', 'DTI').

Warning

The size of this dataset is multiple GB. If you set download to True, it will take some time to be downloaded if it is not already present.

Example

>>> import torchio as tio
>>> transforms = [
...     tio.ToCanonical(),  # to RAS
...     tio.Resample((1, 1, 1)),  # to 1 mm iso
... ]
>>> ixi_dataset = tio.datasets.IXI(
...     'path/to/ixi_root/',
...     modalities=('T1', 'T2'),
...     transform=tio.Compose(transforms),
...     download=True,
... )
>>> print('Number of subjects in dataset:', len(ixi_dataset))  # 577
>>> sample_subject = ixi_dataset[0]
>>> print('Keys in subject:', tuple(sample_subject.keys()))  # ('T1', 'T2')
>>> print('Shape of T1 data:', sample_subject['T1'].shape)  # [1, 180, 268, 268]
>>> print('Shape of T2 data:', sample_subject['T2'].shape)  # [1, 241, 257, 188]

`IXITiny`#

class torchio.datasets.ixi.IXITiny(root: str | Path, transform: Transform | None = None, download: bool = False, **kwargs)[source]#

Bases: SubjectsDataset

This is the dataset used in the main notebook. It is a tiny version of IXI, containing 566 \(T_1\)-weighted brain MR images and their corresponding brain segmentations, all with size \(83 \times 44 \times 55\).

It can be used as a medical image MNIST.

Parameters:

root – Root directory to which the dataset will be downloaded.
transform – An instance of Transform.
download – If set to True, will download the data into root.

EPISURG#

`EPISURG`#

class torchio.datasets.episurg.EPISURG(root: str | Path, transform: Transform | None = None, download: bool = False, **kwargs)[source]#

Bases: SubjectsDataset

EPISURG is a clinical dataset of \(T_1\)-weighted MRI from 430 epileptic patients who underwent resective brain surgery at the National Hospital of Neurology and Neurosurgery (Queen Square, London, United Kingdom) between 1990 and 2018.

The dataset comprises 430 postoperative MRI. The corresponding preoperative MRI is present for 268 subjects.

Three human raters segmented the resection cavity on partially overlapping subsets of EPISURG.

If you use this dataset for your research, you agree with the Data use agreement presented at the EPISURG entry on the UCL Research Data Repository and you must cite the corresponding publications.

Parameters:

root – Root directory to which the dataset will be downloaded.
transform – An instance of Transform.
download – If set to True, will download the data into root.

Warning

The size of this dataset is multiple GB. If you set download to True, it will take some time to be downloaded if it is not already present.

get_labeled() → SubjectsDataset[source]#: Get dataset from subjects with manual annotations.

get_paired() → SubjectsDataset[source]#: Get dataset from subjects with pre- and post-op MRI.

get_unlabeled() → SubjectsDataset[source]#: Get dataset from subjects without manual annotations.

Kaggle datasets#

`RSNAMICCAI`#

class torchio.datasets.rsna_miccai.RSNAMICCAI(root_dir: str | Path, train: bool = True, ignore_empty: bool = True, modalities: Sequence[str] = ('T1w', 'T1wCE', 'T2w', 'FLAIR'), **kwargs)[source]#

Bases: SubjectsDataset

RSNA-MICCAI Brain Tumor Radiogenomic Classification challenge dataset.

This is a helper class for the dataset used in the RSNA-MICCAI Brain Tumor Radiogenomic Classification challenge hosted on kaggle. The dataset must be downloaded before instantiating this class (as opposed to, e.g., torchio.datasets.IXI).

This kaggle kernel includes a usage example including preprocessing of all the scans.

If you reference or use the dataset in any form, include the following citation:

U.Baid, et al., “The RSNA-ASNR-MICCAI BraTS 2021 Benchmark on Brain Tumor Segmentation and Radiogenomic Classification”, arXiv:2107.02314, 2021.

Parameters:

root_dir – Directory containing the dataset (train directory, test directory, etc.).
train – If True, the train set will be used. Otherwise the test set will be used.
ignore_empty – If True, the three subjects flagged as “presenting issues” (empty images) by the challenge organizers will be ignored. The subject IDs are 00109, 00123 and 00709.

Example

>>> import torchio as tio
>>> from subprocess import call
>>> call('kaggle competitions download -c rsna-miccai-brain-tumor-radiogenomic-classification'.split())
>>> root_dir = 'rsna-miccai-brain-tumor-radiogenomic-classification'
>>> train_set = tio.datasets.RSNAMICCAI(root_dir, train=True)
>>> test_set = tio.datasets.RSNAMICCAI(root_dir, train=False)
>>> len(train_set), len(test_set)
(582, 87)

`RSNACervicalSpineFracture`#

class torchio.datasets.rsna_spine_fracture.RSNACervicalSpineFracture(root_dir: str | Path, add_segmentations: bool = False, add_bounding_boxes: bool = False, **kwargs)[source]#

Bases: SubjectsDataset

RSNA 2022 Cervical Spine Fracture Detection dataset.

This is a helper class for the dataset used in the RSNA 2022 Cervical Spine Fracture Detection hosted on kaggle. The dataset must be downloaded before instantiating this class.

MNI#

`ICBM2009CNonlinearSymmetric`#

class torchio.datasets.mni.ICBM2009CNonlinearSymmetric(load_4d_tissues: bool = True)[source]#

Bases: SubjectMNI

ICBM template.

More information can be found in the website.

Parameters:: load_4d_tissues – If True, the tissue probability maps will be loaded together into a 4D image. Otherwise, they will be loaded into independent images.

Example

>>> import torchio as tio
>>> icbm = tio.datasets.ICBM2009CNonlinearSymmetric()
>>> icbm
ICBM2009CNonlinearSymmetric(Keys: ('t1', 'eyes', 'face', 'brain', 't2', 'pd', 'tissues'); images: 7)
>>> icbm = tio.datasets.ICBM2009CNonlinearSymmetric(load_4d_tissues=False)
>>> icbm
ICBM2009CNonlinearSymmetric(Keys: ('t1', 'eyes', 'face', 'brain', 't2', 'pd', 'gm', 'wm', 'csf'); images: 9)

`Colin27`#

class torchio.datasets.mni.Colin27(version=1998)[source]#: Bases: SubjectMNI

Medical image datasets#

IXI#

EPISURG#

Kaggle datasets#

MNI#

ITK-SNAP#

3D Slicer#

FPG#

MedMNIST#