torchio is a Python package containing a set of tools to efficiently
read, sample and write 3D medical images in deep learning applications
written in PyTorch,
including intensity and spatial transforms
for data augmentation and preprocessing. Transforms include typical computer vision operations
such as random affine transformations and also domain specific ones such as
simulation of intensity artifacts due to
MRI magnetic field inhomogeneity
or k-space motion artifacts.
This package has been greatly inspired by NiftyNet.
If you like this repository, please click on Star!
If you used this package for your research, please cite this repository using the information available on its Zenodo entry or use this BibTeX:
@software{perez_garcia_fernando_2020_3598622,
author = {Pérez-García, Fernando},
title = {{fepegar/torchio: TorchIO: Tools for loading,
augmenting and writing 3D medical images on
PyTorch}},
month = jan,
year = 2020,
publisher = {Zenodo},
doi = {10.5281/zenodo.3598622},
url = {https://doi.org/10.5281/zenodo.3598622}
}This package is on the Python Package Index (PyPI). To install it, just run in a terminal the following command:
$ pip install torchioImagesDataset is a reader of medical images that directly inherits from
torch.utils.Dataset.
It can be used with a
torch.utils.DataLoader
for efficient reading and data augmentation.
It receives a list of subjects, where each subject is composed of a list of
torchio.Image instances.
The paths suffix must be .nii, .nii.gz or .nrrd.
import torchio
subject_a = [
Image('t1', '~/Dropbox/MRI/t1.nrrd', torchio.INTENSITY),
Image('label', '~/Dropbox/MRI/t1_seg.nii.gz', torchio.LABEL),
]
subject_b = [
Image('t1', '/tmp/colin27_t1_tal_lin.nii.gz', torchio.INTENSITY),
Image('t2', '/tmp/colin27_t2_tal_lin.nii', torchio.INTENSITY),
Image('label', '/tmp/colin27_seg1.nii.gz', torchio.LABEL),
]
subjects_list = [subject_a, subject_b]
subjects_dataset = torchio.ImagesDataset(subjects_list)
subject_sample = subjects_dataset[0]torchio includes grid, uniform and label patch samplers. There is also an
aggregator used for dense predictions.
For more information about patch-based training, see
NiftyNet docs.
import torch
import torchio
CHANNELS_DIMENSION = 1
patch_overlap = 4
grid_sampler = torchio.inference.GridSampler(
input_array, # some NumPy array
patch_size=128,
patch_overlap=patch_overlap,
)
patch_loader = torch.utils.data.DataLoader(grid_sampler, batch_size=4)
aggregator = torchio.inference.GridAggregator(
input_array,
patch_overlap=patch_overlap,
)
with torch.no_grad():
for patches_batch in patch_loader:
input_tensor = patches_batch['image']
locations = patches_batch['location']
logits = model(input_tensor) # some torch.nn.Module
labels = logits.argmax(dim=CHANNELS_DIMENSION, keepdim=True)
outputs = labels
aggregator.add_batch(outputs, locations)
output_array = aggregator.output_arrayA patches Queue (or buffer) can be used for randomized patch-based sampling
during training.
This interactive animation
can be used to understand how the queue works.
import torch
import torchio
patches_queue = torchio.Queue(
subjects_dataset=subjects_dataset, # instance of torchio.ImagesDataset
queue_length=300,
samples_per_volume=10,
patch_size=96,
sampler_class=torchio.sampler.ImageSampler,
num_workers=4,
shuffle_subjects=True,
shuffle_patches=True,
)
patches_loader = DataLoader(patches_queue, batch_size=4)
num_epochs = 20
for epoch_index in range(num_epochs):
for patches_batch in patches_loader:
logits = model(patches_batch) # model is some torch.nn.ModuleThe transforms package should remind users of
torchvision.transforms.
They take as input the samples generated by an ImagesDataset.
A transform can be quickly applied to an image file using the command-line
tool torchio-transform:
$ torchio-transform input.nii.gz RandomMotion output.nii.gz --kwargs "proportion_to_augment=1 num_transforms=4"Magnetic resonance images suffer from motion artifacts when the subject moves during image acquisition. This transform follows Shaw et al., 2019 to simulate motion artifacts for data augmentation.
MRI magnetic field inhomogeneity creates slow frequency intensity variations. This transform is very similar to the one in NiftyNet.
Adds noise sampled from a normal distribution with mean 0 and standard
deviation sampled from a uniform distribution in the range std_range.
It is often used after ZNormalization, as the output of
this transform has zero-mean.
Reverse the order of elements in an image along the given axes.
Implementation of New variants of a method of MRI scale standardization adapted from NiftyNet.
This transform first extracts the values with intensity greater than the mean, which is an approximation of the foreground voxels. Then the foreground mean is subtracted from the image and it is divided by the foreground standard deviation.
Rescale intensity values in an image to a certain range.
Resample images to a new voxel spacing using nibabel.
Pad images, like in torchvision.transforms.Pad.
Crop images passing 1, 3, or 6 integers, as in Pad.
This example shows the improvement in performance when multiple workers are used to load and preprocess the volumes using multiple workers.
import time
import multiprocessing as mp
from tqdm import trange
import torch.nn as nn
from torch.utils.data import DataLoader
from torchvision.transforms import Compose
from torchio import ImagesDataset, Queue, DATA
from torchio.sampler import ImageSampler
from torchio.utils import create_dummy_dataset
from torchio.transforms import (
ZNormalization,
RandomNoise,
RandomFlip,
RandomAffine,
)
# Define training and patches sampling parameters
num_epochs = 4
patch_size = 128
queue_length = 400
samples_per_volume = 10
batch_size = 4
class Network(nn.Module):
def __init__(self):
super().__init__()
self.conv = nn.Conv3d(
in_channels=1,
out_channels=3,
kernel_size=3,
)
def forward(self, x):
return self.conv(x)
model = Network()
# Create a dummy dataset in the temporary directory, for this example
subjects_list = create_dummy_dataset(
num_images=100,
size_range=(193, 229),
force=False,
)
# Each element of subjects_list is a dictionary:
# subject_images = [
# torchio.Image('one_image', path_to_one_image, torchio.INTENSITY),
# torchio.Image('another_image', path_to_another_image, torchio.INTENSITY),
# torchio.Image('a_label', path_to_a_label, torchio.LABEL),
# ]
# Define transforms for data normalization and augmentation
transforms = (
ZNormalization(),
RandomNoise(std_range=(0, 0.25)),
RandomAffine(scales=(0.9, 1.1), degrees=10),
RandomFlip(axes=(0,)),
)
transform = Compose(transforms)
subjects_dataset = ImagesDataset(subjects_list, transform)
# Run a benchmark for different numbers of workers
workers = range(mp.cpu_count() + 1)
for num_workers in workers:
print('Number of workers:', num_workers)
# Define the dataset as a queue of patches
queue_dataset = Queue(
subjects_dataset,
queue_length,
samples_per_volume,
patch_size,
ImageSampler,
num_workers=num_workers,
)
batch_loader = DataLoader(queue_dataset, batch_size=batch_size)
start = time.time()
for epoch_index in trange(num_epochs, leave=False):
for batch in batch_loader:
# The keys of batch have been defined in create_dummy_dataset()
inputs = batch['one_modality'][DATA]
targets = batch['segmentation'][DATA]
logits = model(inputs)
print('Time:', int(time.time() - start), 'seconds')
print()Output:
Number of workers: 0
Time: 394 seconds
Number of workers: 1
Time: 372 seconds
Number of workers: 2
Time: 278 seconds
Number of workers: 3
Time: 259 seconds
Number of workers: 4
Time: 242 seconds