pip install mriaug to use a 3D image library that is ~50x faster and simpler than torchio by

• only using PyTorch → full GPU(+autograd) support 🔥
• being tiny: ~200 lines of code → no room for bugs 🐛

while offering ~20 different augmentations (incl. MRI-specific operations) 🩻

👶 Normal users should use mriaug via niftiai, a deep learning framework for 3D images, since it

• provides aug_transforms3d: A convenient function that compiles all mriaugmentations!
• simplifies all the code needed for data loading, training, visualization...check it out here!

👴 Experienced users can build their own framework upon mriaug (use niftiai/augment.py as a cheat sheet)

Let's create a 3D image tensor (with additional batch and channel dimension) and apply flip3d

import torch
from mriaug import flip3d

shape = (1, 1, 4, 4, 4)
x = torch.linspace(0, 1, 4**3).view(*shape)
x_flipped = flip3d(x)
print(x[..., 0, 0])  # tensor([[[0.0000, 0.2540, 0.5079, 0.7619]]])
print(x_flipped[..., 0, 0])  # tensor([[[0.7619, 0.5079, 0.2540, 0.0000]]])

Explore the gallery to understand the usage and effect of all ~20 augmentations!

The popular libraries torchio and MONAI (utilizes torchio) often use ITK (CPU only) like this

PyTorch tensor → NumPy array → NiBabel image → ITK operation (C/C++) → NumPy array → PyTorch tensor

to augment a PyTorch tensor 😬 That's complicated and does not use the (for neural net training needed) GPU 🐌

Instead, mriaug directly uses PyTorch (CPU & GPU support) resulting in

• ~50x fewer lines of code: torchio: ~10,000 LOC, mriaug: ~200 LOC 🤓
• ~50x speedup on GPU 🔥 based on the table below (run speed.py to reproduce) 💨

Let's load an example 3D image x, show it with niftiview (used to create all images below)

define some arguments

size = (160, 196, 160)
zoom = torch.tensor([[-.2, 0, 0]])
rotate = torch.tensor([[0, .1, 0]])
translate = torch.tensor([[0, 0, .2]])
shear = torch.tensor([[0, .05, 0]])

and run all augmentations (see runall.py):