bio-volumentations 1.2.0

Library for 3D-5D augmentations of volumetric multi-dimensional biomedical images and their annotations

Bio-Volumentations

Bio-Volumentations is an image augmentation and preprocessing package for 3D (volumetric), 4D (time-lapse volumetric or multi-channel volumetric), and 5D (time-lapse multi-channel volumetric) biomedical images and their annotations.

The library offers a wide range of efficiently implemented image transformations. This includes both preprocessing transformations (such as intensity normalisation, padding, and type casting) and augmentation transformations (such as affine transform, noise addition and removal, and contrast manipulation).

The Bio-Volumentations library is a suitable tool for image manipulation in machine learning applications. It can be used with any major Python deep learning library, including PyTorch, PyTorch Lightning, TensorFlow, and Keras.

This library builds upon wide-spread libraries such as Albumentations and TorchIO (see the Contributions section below). Therefore, it can easily be adopted by developers.

Installation

Install the package from pip using:

pip install bio-volumentations

See the project's PyPi page for more details.

Requirements

• NumPy
• SciPy
• Scikit-image
• Matplotlib
• SimpleITK

Usage

Importing

Import the library to your project using:

import bio_volumentations as biovol

How to Use Bio-Volumentations?

The Bio-Volumentations library processes 3D, 4D, and 5D images. Each image must be represented as a numpy.ndarray and must conform to the following conventions:

• The order of dimensions is [C, Z, Y, X, T], where C is the channel dimension, T is the time dimension, and Z, Y, and X are the spatial dimensions.
• The three spatial dimensions (Z, Y, X) must be present. To transform a 2D image, please create a dummy Z dimension first.
• The channel (C) dimension is optional. If it is not present, the library will automatically create a dummy dimension in its place, so the output image shape will be [1, Z, Y, X].
• The time (T) dimension is optional and can only be present if the channel (C) dimension is also present in the input data. To process single-channel time-lapse images, please create a dummy C dimension first.

Thus, an input image is interpreted in the following ways based on its shape:

1. [Z, Y, X] ... a single-channel volumetric image;
2. [C, Z, Y, X] ... a multi-channel volumetric image;
3. [C, Z, Y, X, T] ... a single-channel as well as multi-channel volumetric image sequence.

The shape of the output image is either [C, Z, Y, X] (cases 1 & 2) or [C, Z, Y, X, T] (case 3).

The images are type-casted to a floating-point datatype before transformations, irrespective of their actual datatype.

For the specification of image annotation conventions, please see below.

It is strongly recommended to use Compose to create and use transformation pipelines.
The Compose class automatically checks and adjusts image format, datatype, stacks individual transforms to a pipeline, and outputs the image as a contiguous array. Optionally, it can also convert the transformed image to a desired format.
If you call transformations outside of Compose, we cannot guarantee the all assumptions are checked and enforced, so you might encounter unexpected behaviour.

Below, there are several examples of how to use this library. You are also welcome to check our documentation pages.

Example: Transforming a Single Image

To create the transformation pipeline, you just need to instantiate all desired transformations (with the desired parameter values) and then feed a list of these transformations into a new Compose object.

Optionally, you can specify a datatype conversion transformation that will be applied after the last transformation in the list, e.g. from the default numpy.ndarray to a torch.Tensor. You can also specify the probability of actually applying the whole pipeline as a number between 0 and 1. The default probability is 1 (always apply). See the docs for more details.

The Compose object is callable. The data is passed as a keyword argument, and the call returns a dictionary with the same keywords and corresponding transformed images. This might look like an overkill for a single image, but will come handy when transforming images with annotations. The default key for the image is image.

import numpy as np
from bio_volumentations import Compose, RandomGamma, RandomRotate90, GaussianBlur

# Create the transformation pipeline using Compose
aug = Compose([
        RandomGamma(gamma_limit = (0.8, 1,2), p = 0.8),
        RandomRotate90(axes = [1, 2, 3], p = 1),
        GaussianBlur(sigma = 1.2, p = 0.8)
      ])

# Generate an image - shape [C, Z, Y, X]
img = np.random.rand(1, 128, 256, 256)

# Transform the image
# Please note that the image must be passed as a keyword argument to the transformation pipeline
# and extracted from the outputted dictionary.
data = {'image': img}
aug_data = aug(**data)
transformed_img = aug_data['image']

Example: Transforming Image Tuples

Sometimes, it is necessary to consistently transform a tuple of corresponding images. To that end, Bio-Volumentations define several target types:

• image for the image data (any datatype allowed, gets converted to floating-point by default);
• mask for integer-valued label images (expected integer datatype); and
• float_mask for real-valued label images (expected floating-point datatype).

The mask and float_mask target types are expected to have the same shape as the image target except for the channel (C) dimension which must not be included. For example, for images of shape [150, 300, 300], [1, 150, 300, 300], and [4, 150, 300, 300], the corresponding mask and float_mask must be of shape [150, 300, 300]. If you want to use a multichannel mask or float_mask, you have to split it into a set of single-channel masks or float_masks, respectively, and input them as stand-alone targets (see below).

If a Random... transform receives multiple targets on its input in a single call, the same transformation parameters are used to transform all of these targets. For example, RandomAffineTransform applies the same geometric transformation to all target types in a single call.

Some transformations, such as RandomGaussianNoise or RandomGamma, are only defined for the image target and leave the mask and float_mask targets unchanged. Please consult the documentation of the individual transforms for more details.

The image tuples are fed to the Compose object call as keyword arguments and extracted from the outputted dictionary using the same keys. The default key values are image, mask, and float_mask.

import numpy as np
from bio_volumentations import Compose, RandomGamma, RandomRotate90, GaussianBlur

# Create the transformation using Compose
aug = Compose([
        RandomGamma(gamma_limit = (0.8, 1,2), p = 0.8),
        RandomRotate90(axes = [1, 2, 3], p = 1),
        GaussianBlur(sigma = 1.2, p = 0.8)
      ])

# Generate image and a corresponding labeled image
img = np.random.rand(1, 128, 256, 256)
lbl = np.random.randint(0, 1, size=(128, 256, 256), dtype=np.uint8)

# Transform the images
# Please note that the images must be passed as keyword arguments to the transformation pipeline
# and extracted from the outputted dictionary.
data = {'image': img, 'mask': lbl}
aug_data = aug(**data)
transformed_img, transformed_lbl = aug_data['image'], aug_data['mask']

Example: Transforming Multiple Images of the Same Target Type

Although there are only three target types, you can input an arbitrary number of images to any transformation. To achieve this, you have to define the value of the targets argument when creating a Compose object.

The value of targets must be a list with exactly 3 items: a list with keys of image-type targets, a list with keys of mask-type targets, and a list with keys of float_mask-type targets. The specified keys will then be used to input the images to the transformation call as well as to extract the transformed images from the outputted dictionary.

The keys can be any valid dictionary keys; most importantly, they must be unique across all target types. You don't need to feed an image for each target to the transformation call: in our example below, we have four targets (two image, one mask, and one float_mask), but we only transform three images.

You cannot define your own target types; that would require re-implementing all existing transforms.

import numpy as np
from bio_volumentations import Compose, RandomGamma, RandomRotate90, GaussianBlur

# Create the transformation using Compose: do not forget to define targets
aug = Compose([
        RandomGamma( gamma_limit = (0.8, 1,2), p = 0.8),
        RandomRotate90(axes = [1, 2, 3], p = 1),
        GaussianBlur(sigma = 1.2, p = 0.8)
    ], 
    targets= [ ['image' , 'image1'] , ['mask'], ['float_mask'] ])

# Generate the image data: two images and a single int-valued mask
img = np.random.rand(1, 128, 256, 256)
img1 = np.random.rand(1, 128, 256, 256)
lbl = np.random.randint(0, 1, size=(128, 256, 256), dtype=np.uint8)

# Transform the images
# Please note that the images must be passed as keyword arguments to the transformation pipeline
# and extracted from the outputted dictionary.
data = {'image': img, 'image1': img1, 'mask': lbl}
aug_data = aug(**data)
transformed_img = aug_data['image']
transformed_img1 = aug_data['image1']
transformed_lbl = aug_data['mask']

Example: Adding a Custom Transformation

Each transformation inherits from the Transform class. You can thus easily implement your own transformations and use them with this library. You can check our implementations to see how this can be done. For example, Flip can be implemented as follows:

import numpy as np
from typing import List
from bio_volumentations import DualTransform

class Flip(DualTransform):
    def __init__(self, axes: List[int] = None, always_apply=False, p=1):
        super().__init__(always_apply, p)
        self.axes = axes

    # Transform the image
    def apply(self, img, **params):
        return np.flip(img, params["axes"])

    # Transform the int-valued mask
    def apply_to_mask(self, mask, **params):
       # The mask has no channels
        return np.flip(mask, axis=[item - 1 for item in params["axes"]])
    
    # Transform the float-valued mask
    # By default, float_mask uses the implementation of mask, unless it is overridden (see the implementation of DualTransform).
    #def apply_to_float_mask(self, float_mask, **params):
    #    return self.apply_to_mask(float_mask, **params)

    # Get transformation parameters. Useful especially for RandomXXX transforms to ensure consistent transformation of image tuples.
    def get_params(self, **data):
        axes = self.axes if self.axes is not None else [1, 2, 3]
        return {"axes": axes}

Implemented Transforms

A List of Implemented Transformations

Point transformations:

GaussianNoise 
PoissonNoise
GaussianBlur 
RandomGaussianBlur
RandomGamma 
RandomBrightnessContrast 
HistogramEqualization 
Normalize
NormalizeMeanStd

Geometrical transformations:

AffineTransform
Resize 
Scale
Flip 
CenterCrop 
Pad
RandomAffineTransform
RandomScale 
RandomRotate90
RandomFlip 
RandomCrop

Other:

Float
Contiguous

Contributions

Authors of the Bio-Volumentations library: Samuel Šuľan, Lucia Hradecká, Filip Lux.

• Lucia Hradecká: [e-mail]
• Filip Lux: [e-mail]

The Bio-Volumentations library is based on the following image augmentation libraries:

• Albumentations
• Volumentations
• Volumentations: Continued Development
• Volumentations: Enhancements
• Volumentations: Further Enhancements
• TorchIO

We would thus like to thank their authors, namely the Albumentations team, Pavel Iakubovskii, ZFTurbo, ashawkey, Dominik Müller, and TorchIO contributors.

Citation

TBA