Helper Module for Deep Learning.

pynet data augmentation overview¶

Credit: A Grigis

pynet contains a set of tools to efficiently augment 3D medical images that is crutial for deep learning applications. It includes random affine/non linear transformations, simulation of intensity artifacts due to MRI magnetic field inhomogeneity or k-space motion artifacts, and others.

Load the data¶

We load the Brats dataset and select the first MRI brain image.

import os
import sys
if "CI_MODE" in os.environ:
    sys.exit()

import time
import numpy as np
import nibabel
import random
from pynet.datasets import DataManager, fetch_toy, fetch_brats
from pynet.preprocessing import rescale, downsample

datasetdir = "/tmp/toy"
data = fetch_toy(datasetdir=datasetdir)
image = nibabel.load(data.t1w_path)
image = rescale(downsample(image.get_data(), scale=4), dynamic=(0, 255))

Define deformations¶

We now declare MRI brain deformation functions. The deformation can be combined with the Transformer class.

from pynet.augmentation import add_blur
from pynet.augmentation import add_noise
from pynet.augmentation import add_ghosting
from pynet.augmentation import add_spike
from pynet.augmentation import add_biasfield
from pynet.augmentation import add_motion
from pynet.augmentation import add_offset
from pynet.augmentation import flip
from pynet.augmentation import affine
from pynet.augmentation import deformation
from pynet.augmentation import Transformer

compose_transforms = Transformer(with_channel=False)
compose_transforms.register(
    flip, probability=0.5, axis=0, apply_to=["all"])
compose_transforms.register(
    add_blur, probability=1, sigma=4, apply_to=["all"])
transforms = {
    "add_blur": (add_blur, {"sigma": 4}),
    "add_noise": (add_noise, {"snr": 5., "noise_type": "rician"}),
    "flip": (flip, {"axis": 0}),
    "affine": (affine, {"rotation": 5, "translation": 0, "zoom": 0.05}),
    "add_ghosting": (add_ghosting, {"n_ghosts": (4, 10), "axis": 2,
                                   "intensity": (0.5, 1)}),
    "add_spike": (add_spike, {"n_spikes": 1, "intensity": (0.1, 1)}),
    "add_biasfield": (add_biasfield, {"coefficients": 0.5}),
    "deformation": (deformation, {"max_displacement": 4, "alpha": 3}),
    "add_motion": (add_motion, {"rotation": 10, "translation": 10,
                                "n_transforms": 2, "perturbation": 0.3}),
    "add_offset": (add_offset, {"factor": (0.05, 0.1)}),
    "compose_transforms": (compose_transforms, {}),
}

Test transformations¶

We now apply the transformations on the loaded image. Results are directly displayed in your browser at http://localhost:8097.

from pynet.plotting import Board

board = Board(port=8097, host="http://localhost", env="data-augmentation")
for cnt in range(10):
    print("Iteration: ", cnt)
    for key, (fct, kwargs) in transforms.items():
        images = np.asarray([image, np.clip(fct(image, **kwargs), 0, 255)])
        images = images[..., images.shape[-1] // 2]
        images = np.expand_dims(images, axis=1)
        board.viewer.images(
            images, opts={"title": key, "caption": key}, win=key)
    time.sleep(1)

Data augmentation¶

We now illustrate how we can use the Transformer in combinaison with the DataManager to perform data augmentation during training. Results are directly displayed in your browser at http://localhost:8097.

datasetdir = "/neurospin/nsap/processed/deepbrain/tumor/data/brats"
data = fetch_brats(datasetdir=datasetdir)

board = Board(port=8097, host="http://localhost", env="data-augmentation")
compose_transforms = Transformer()
compose_transforms.register(
    flip, probability=0.5, axis=0, apply_to=["input", "output"])
compose_transforms.register(
    add_blur, probability=1, sigma=4, apply_to=["input"])
manager = DataManager(
    input_path=data.input_path,
    metadata_path=data.metadata_path,
    output_path=data.output_path,
    number_of_folds=2,
    batch_size=2,
    test_size=0.1,
    sample_size=0.1,
    sampler=None,
    add_input=True,
    data_augmentation_transforms=[compose_transforms])
loaders = manager.get_dataloader(
    train=True,
    validation=False,
    fold_index=0)
for dataitem in loaders.train:
    print("-" * 50)
    print(dataitem.inputs.shape, dataitem.outputs.shape, dataitem.labels)
    images = [dataitem.inputs[0, 0].numpy(), dataitem.inputs[0, 1].numpy(),
              dataitem.outputs[0, 0].numpy(), dataitem.outputs[0, 1].numpy(),
              dataitem.outputs[0, 4].numpy(), dataitem.outputs[0, 5].numpy()]
    images = np.asarray(images)
    images = np.expand_dims(images, axis=1)
    images = images[..., images.shape[-1] // 2]
    images = rescale(images, dynamic=(0, 255))
    board.viewer.images(
        images, opts={"title": "transformer", "caption": "transformer"},
        win="transformer")
    time.sleep(2)

Total running time of the script: ( 0 minutes 0.000 seconds)

Gallery generated by Sphinx-Gallery