Helper Module for Deep Learning.

Common functions to transform image. Code: https://github.com/fepegar/torchio

pynet.augmentation.spatial.affine(arr, rotation=10, translation=10, zoom=0.2, order=3, dist='uniform', seed=None)[source]¶

Random affine transformation.

The affine translation & rotation parameters are drawn from a lognormal distribution - small movements are assumed to occur more often and large movements less frequently - or from a uniform distribution.

Parameters

arr: array

the input data.

rotation: float or 2-uplet, default 10

the rotation in degrees of the simulated movements. Larger values generate more distorted images.

translation: float or 2-uplet, default 10

the translation in voxel of the simulated movements. Larger values generate more distorted images.

zoom: float, default 0.2

the zooming magnitude. Larger values generate more distorted images.

order: int, default 3

the order of the spline interpolation in the range [0, 5].

dist: str, default ‘uniform’

the sampling distribution: ‘uniform’ or ‘lognormal’.

seed: int, default None

seed to control random number generator.

Returns

transformed: array

the transformed input data.

pynet.augmentation.spatial.deformation(arr, max_displacement=4, alpha=3, order=3, seed=None)[source]¶

Apply dense random elastic deformation.

Reference: Khanal B, Ayache N, Pennec X., Simulating Longitudinal Brain MRIs with Known Volume Changes and Realistic Variations in Image Intensity, Front Neurosci, 2017.

Parameters

arr: array

the input data.

max_displacement: float, default 4

the maximum displacement in voxel along each dimension. Larger values generate more distorted images.

alpha: float, default 3

the power of the power-law momentum distribution. Larger values genrate smoother fields.

order: int, default 3

the order of the spline interpolation in the range [0, 5].

seed: int, default None

seed to control random number generator.

Returns

transformed: array

the transformed input data.

pynet.augmentation.spatial.flip(arr, axis=None, seed=None)[source]¶

Apply a random mirror flip.

Parameters

arr: array

the input data.

axis: int, default None

apply flip on the specified axis. If not specified, randomize the flip axis.

seed: int, default None

seed to control random number generator.

Returns

transformed: array

the transformed input data.

pynet.augmentation.spatial.random_generator(interval, size, dist='uniform', seed=None)[source]¶

Random varaible generator.

Parameters

interval: 2-uplet

the possible values of the generated random variable.

size: uplet

the number of random variables to be drawn from the sampling distribution.

dist: str, default ‘uniform’

the sampling distribution: ‘uniform’ or ‘lognormal’.

seed: int, default None

seed to control random number generator.

Returns

random_variables: array

the generated random variable.