BrainWebPET#

class deepinv.datasets.BrainWebPET(root=None, subject_ids=None, download=True, transform=None, pet_class=None, contrast=(), random_degradations_kwargs=None, lesion_diameters=None, lesion_kwargs=None, seed=0)[source]#

Bases: ImageDataset

BrainWeb PET phantoms.

Loads synthetic 3D volumes from BrainWeb dataset [1], of shape (1, 127, 344, 344). The dataset has been adapted to emission tomography, and returns an emission and attenuation map, at the Siemens Biograph mMR isotropic resolution of 2.0863 mm per voxel.

Passing lesion_diameters adds high activity lesions with brainweb.add_lesions and includes a lesion_mask in the returned params, where the background is labelled 0 and lesions are labelled from 1 onwards.

This dataset relies on the original implementation of Casper da Costa-Luis: <casperdcl/brainweb>`_. Install it with pip install brainweb. See the original implementation for a detailed description of the keyword arguments.

Note

For a version of this dataset dedicated to magnetic resonance imaging, which contains more contrast options, see deepinv.datasets.BrainWebMRI.

Parameters:

  • root (str, pathlib.Path, None) – Dataset directory. Defaults to the DeepInv cache.

  • subject_ids (int, collections.abc.Sequence[int], None) – Subjects to include in the dataset. Defaults to None which includes all subjects.

  • download (bool) – Download missing subjects. Defaults to True.

  • pet_class (type[brainweb.Act], None) – BrainWeb PET activity preset. Defaults to brainweb.FDG.

  • lesion_diameters (list[float], None) – Lesion diameters in mm. Defaults to None, which adds no lesions.

  • contrast (str, collections.abc.Sequence[str]) – Contrasts to include in the returned parameters. Valid values are "T1" and "T2". Defaults to an empty tuple.

  • lesion_kwargs (dict, None) – Keyword arguments for brainweb.add_lesions.

  • random_degradations_kwargs (dict, None) – Keyword arguments for brainweb.get_mmr_fromfile controlling random structural degradations.

  • transform (collections.abc.Callable, None) – Optional transform to apply to the returned volumes.

  • seed (int, None) – Seed used when adding random lesions.

Example:

>>> import brainweb
>>> import numpy as np
>>> from deepinv.datasets import BrainWebPET
>>> class RandomFDG(brainweb.FDG):
...     greyMatter = lambda: np.random.normal(128, 8)
>>> dataset = BrainWebPET(
...     root="data/brainweb_pet",
...     random_degradations_kwargs={"petNoise": 0.5, "petSigma": 2},
...     contrast=["T1", "T2"],
...     pet_class=RandomFDG,
...     lesion_diameters=[15, 7],
...     lesion_kwargs={"intensity": [200, 150], "blur": [0, 0], "thresh": 30},
... )
>>> emission, params = dataset[0]
>>> emission.shape == params["attenuation"].shape
True

References: