Helper Module for Deep Learning.
Source code for pynet.augmentation.spatial
# -*- coding: utf-8 -*-
##########################################################################
# NSAp - Copyright (C) CEA, 2020
# Distributed under the terms of the CeCILL-B license, as published by
# the CEA-CNRS-INRIA. Refer to the LICENSE file or to
# http://www.cecill.info/licences/Licence_CeCILL-B_V1-en.html
# for details.
##########################################################################
"""
Common functions to transform image.
Code: https://github.com/fepegar/torchio
"""
# Import
import numpy as np
from scipy.spatial.transform import Rotation
from scipy.ndimage import map_coordinates
from .transform import compose
from .transform import gaussian_random_field
from .transform import affine_flow
from .utils import interval
[docs]def affine(arr, rotation=10, translation=10, zoom=0.2, order=3, dist="uniform",
seed=None):
""" 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.
"""
rotation = interval(rotation)
translation = interval(translation)
random_rotations = random_generator(
rotation, arr.ndim, dist=dist, seed=seed)
random_translations = random_generator(
translation, arr.ndim, dist=dist, seed=seed)
np.random.seed(seed)
random_zooms = np.random.uniform(
low=(1 - zoom), high=(1 + zoom), size=arr.ndim)
random_rotations = Rotation.from_euler(
"xyz", random_rotations, degrees=True)
random_rotations = random_rotations.as_matrix()
affine = compose(random_translations, random_rotations, random_zooms)
shape = arr.shape
flow = affine_flow(affine, shape)
locs = flow.reshape(len(shape), -1)
transformed = map_coordinates(arr, locs, order=order, cval=0)
return transformed.reshape(shape)
[docs]def flip(arr, axis=None, seed=None):
""" 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.
"""
if axis is None:
np.random.seed(seed)
axis = np.random.randint(low=0, high=arr.ndim, size=1)[0]
return np.flip(arr, axis=axis)
[docs]def deformation(arr, max_displacement=4, alpha=3, order=3, seed=None):
""" 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.
"""
kwargs = {"seed": seed}
flow_x = gaussian_random_field(
arr.shape[:2], alpha=alpha, normalize=True, **kwargs)
flow_x /= flow_x.max()
flow_x = np.asarray([flow_x] * arr.shape[-1]).transpose(1, 2, 0)
if seed is not None:
kwargs = {"seed": seed + 2}
flow_y = gaussian_random_field(
arr.shape[:2], alpha=alpha, normalize=True, **kwargs)
flow_y /= flow_y.max()
flow_y = np.asarray([flow_y] * arr.shape[-1]).transpose(1, 2, 0)
if seed is not None:
kwargs = {"seed": seed + 4}
flow_z = gaussian_random_field(
arr.shape[:2], alpha=alpha, normalize=True, **kwargs)
flow_z /= flow_z.max()
flow_z = np.asarray([flow_z] * arr.shape[-1]).transpose(1, 2, 0)
flow = np.asarray([flow_x, flow_y, flow_z])
flow *= max_displacement
ranges = [np.arange(size) for size in arr.shape]
locs = np.asarray(np.meshgrid(*ranges)).transpose(0, 2, 1, 3).astype(float)
locs += flow
locs = locs.reshape(len(locs), -1)
transformed = map_coordinates(arr, locs, order=order, cval=0)
return transformed.reshape(arr.shape)
[docs]def random_generator(interval, size, dist="uniform", seed=None):
""" 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.
"""
if dist == "uniform":
np.random.seed(seed)
random_variables = np.random.uniform(
low=interval[0], high=interval[1], size=size)
# max height occurs at x = exp(mean - sigma**2)
# FWHM is found by finding the values of x at 1/2 the max height =
# exp((mean - sigma**2) + sqrt(2*sigma**2*ln(2))) - exp((mean - sigma**2)
# - sqrt(2*sigma**2*ln(2)))
elif dist == "lognormal":
np.random.seed(seed)
sign = np.random.randint(0, 2, size=size) * 2 - 1
sign = sign.astype(np.float)
np.random.seed(seed)
random_variables = np.random.lognormal(mean=0., sigma=1., size=size)
random_variables /= 12.5
random_variables *= (sign * interval[1])
else:
raise ValueError("Unsupported sampling distribution.")
return random_variables
Follow us
© 2019, pynet developers .
Inspired by AZMIND template.
Inspired by AZMIND template.