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 and padding) and augmentation transformations (such as affine transform, noise addition and removal, and contrast manipulation).
Bio-Volumentations are a handy tool for image manipulation in machine learning applications.
The library can transform 3D to 5D images with image-based and point-based annotations,
gives you fine-grained control over the transformation pipelines,
and can be used with any major Python deep learning library
(including PyTorch, PyTorch Lightning, TensorFlow, and Keras)
in a wide range of applications including classification, object detection, semantic & instance
segmentation, and object tracking.
Bio-Volumentations build upon widely used libraries such as Albumentations and TorchIO
(see the Contributions and Acknowledgements section below) and are accompanied by
detailed documentation and a user guide.
Therefore, they can easily be adopted by developers.
Simply install the package from pip using:
pip install bio-volumentations
That's it :)
For more details, see the project's PyPI page.
Releases can also be found at the project's GitHub page.
Bio-Volumentations are being actively developed by the CBIA @ FI MUNI, therefore
CBIA's official repository at GitLab is the primary repository
at the moment. The project's GitHub repository holds releases
of our library and will be updated at important milestones and with new library versions.
To check out our library on test data, you can run the example provided in the example folder.
There, you will find a test sample consisting of a 3D image (image.tif) with an associated binary mask
(segmentation_mask.tif), a runnable Python script, and the transformed sample (image_transformed.tif and
segmentation_mask_transformed.tif).
To run the example, please download the example folder and
install the bio-volumentations, tiffile and imagecodecs packages to your Python environment.
Then run the following from the command line:
cd example
python transformation_example.py
The script will generate a new randomly transformed sample and save it into the image_transformed.tif and
segmentation_mask_transformed.tif files. These files can be opened using ImageJ.
This example uses data from the Fluo-N3DH-CE dataset [1] from the Cell Tracking Challenge repository [2].
[1] Murray J, Bao Z, Boyle T, et al. Automated analysis of embryonic gene expression with cellular resolution in C. elegans. Nat Methods 2008;5:703–709. https://doi.org/10.1038/nmeth.1228.
[2] Maška M, Ulman V, Delgado-Rodriguez P, et al. The Cell Tracking Challenge: 10 years of objective benchmarking. Nat Methods 2023;20:1010–1020. https://doi.org/10.1038/s41592-023-01879-y. Repository: https://celltrackingchallenge.net/3d-datasets/.
Import the library to your project using:
import bio_volumentations as biovolThe 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.
Thus, an input image is interpreted in the following ways based on its dimensionality:
- 3D: a single-channel volumetric image [Z, Y, X];
- 4D: a multi-channel volumetric image [C, Z, Y, X];
- 5D: a single- or multi-channel volumetric image sequence [C, Z, Y, X, T].
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 being transformed, irrespective of their actual datatype.
For the specification of image annotation conventions, please see below.
The transformations are implemented as callable classes inheriting from an abstract Transform class.
Upon instantiating a transformation object, one has to specify the parameters of the transformation.
All transformations work in a fully 3D fashion. Individual channels and time points of a data volume
are usually transformed separately and in the same manner; however, certain transformations can also work
along these dimensions. For instance, GaussianBlur can perform the blurring along the temporal dimension and
with different strength in individual channels.
The data can be transformed by a call to the transformation object.
It is strongly recommended to use Compose to create and use transformation pipelines.
An instantiated Compose object encapsulates the full transformation pipeline and provides additional support:
it automatically checks and adjusts image format and datatype, outputs the image as a contiguous array, and
can optionally 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.
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 transformation objects 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 (i.e., the pipeline is applied in each call).
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 keyword and the corresponding transformed image. This might look like an overkill for a single image,
but it will come handy when transforming images with annotations. The default key for an 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']Sometimes, it is necessary to transform an image with some associated additional targets.
To that end, Bio-Volumentations define several target types:
imagefor the image data;maskfor integer-valued label images;float_maskfor real-valued label images;keypointsfor a list of key points; andvaluefor non-transformed values.
For more information on the format of individual target types, see the Getting Started guide
Please note that there must always be an image-type target in the sample.
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 other targets unchanged. Please consult the
documentation of the individual transforms for more details.
The associated targets are fed to the Compose object call as keyword arguments and extracted from the outputted
dictionary using the same keywords.
The default key values are 'image', 'mask', 'float_mask', 'keypoints', and 'value'.
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']You can input arbitrary number of inputs to any transformation. To achieve this, you have to define the keywords
for the individual inputs when creating the Compose object.
The specified keywords will then be used to input the images to the transformation call as well as to extract the
transformed images from the outputted dictionary.
Specifically, you can define image-type target keywords using the img_keywords parameter - its value
must be a tuple of strings, each string representing a single keyword. Similarly, there are mask_keywords,
fmask_keywords, value_keywords, and keypoints_keywords parameters for the other target types.
Setting any of these parameters overwrites its default value.
The keywords can be any valid dictionary keys, and they must be unique.
You do not need to use all specified keywords in a transformation call. However, at least the target with
the 'image' keyword must be present in each transformation call.
In our example below, we only transform three targets even though we defined four target keywords explicitly
(and there are some implicit keywords as well for the other target types).
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)
],
img_keywords=('image', 'abc'), mask_keywords=('mask',), fmask_keywords=('nothing',))
# 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, 'abc': img1, 'mask': lbl}
aug_data = aug(**data)
transformed_img = aug_data['image']
transformed_img1 = aug_data['abc']
transformed_lbl = aug_data['mask']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)
# Set 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}However, you cannot define your own target types, as that would require re-implementing all existing transforms.
Intensity-based point transformations:
Normalize
NormalizeMeanStd
HistogramEqualization
GaussianNoise
PoissonNoise
RandomBrightnessContrast
RandomGammaIntensity-based local transformations:
GaussianBlur
RandomGaussianBlur
RemoveBackgroundGaussianGeometric (spatial) transformations:
AffineTransform
Resize
Scale
Rescale
Flip
Pad
CenterCrop
RandomAffineTransform
RandomScale
RandomRotate90
RandomFlip
RandomCropOther transformations:
Contiguous
StandardizeDatatype
ConversionToFormatHere, we present the execution times (mean ± standard deviation) of individual transformations from our library with respect to input image size.
The shape (size) of inputs was [1, 32, 32, 32, 1] (32k voxels), [4, 32, 32, 32, 5] (655k voxels), [4, 64, 64, 64, 5] (5M voxels), and [4, 128, 128, 128, 5] (42M voxels), respectively. The runtimes, presented in milliseconds, were averaged over 100 runs. All measurements were done on a single workstation with a Ryzen 7-3700X CPU @ 3.60GHz.
| Transformation | 32k voxels | 655k voxels | 5M voxels | 42M voxels |
|---|---|---|---|---|
| AffineTransform | 2 ± 0 ms | 23 ± 2 ms | 111 ± 11 ms | 988 ± 33 ms |
| RandomAffineTransform | 3 ± 1 ms | 24 ± 1 ms | 110 ± 7 ms | 1005 ± 28 ms |
| Scale | 3 ± 1 ms | 25 ± 2 ms | 109 ± 9 ms | 990 ± 21 ms |
| RandomScale | 3 ± 1 ms | 25 ± 2 ms | 111 ± 9 ms | 1000 ± 30 ms |
| Flip | < 1 ± 0 ms | 2 ± 0 ms | 14 ± 1 ms | 107 ± 8 ms |
| RandomFlip | < 1 ± 0 ms | 1 ± 1 ms | 13 ± 5 ms | 90 ± 31 ms |
| RandomRotate90 | < 1 ± 0 ms | 2 ± 1 ms | 18 ± 5 ms | 150 ± 44 ms |
| GaussianBlur | 1 ± 0 ms | 13 ± 1 ms | 150 ± 3 ms | 5501 ± 148 ms |
| RandomGaussianBlur | 1 ± 0 ms | 11 ± 1 ms | 151 ± 18 ms | 5623 ± 176 ms |
| GaussianNoise | < 1 ± 0 ms | 10 ± 1 ms | 83 ± 3 ms | 687 ± 8 ms |
| PoissonNoise | 1 ± 0 ms | 20 ± 1 ms | 160 ± 1 ms | 1285 ± 18 ms |
| HistogramEqualization | 2 ± 0 ms | 35 ± 1 ms | 266 ± 5 ms | 2120 ± 18 ms |
| Normalize | < 1 ± 0 ms | 2 ± 1 ms | 32 ± 5 ms | 346 ± 24 ms |
| NormalizeMeanStd | < 1 ± 0 ms | 1 ± 0 ms | 9 ± 1 ms | 69 ± 5 ms |
| RandomBrightnessContrast | < 1 ± 0 ms | 1 ± 0 ms | 6 ± 1 ms | 49 ± 8 ms |
| RandomGamma | < 1 ± 0 ms | 7 ± 0 ms | 59 ± 1 ms | 469 ± 6 ms |
We also present the execution times (mean ± standard deviation) of eight commonly used transformations,
comparing the performance of our Bio-Volumentations to other libraries capable of processing volumetric
image data: TorchIO [3], Volumentations [4, 5], and Gunpowder [6].
Asterisks (*) denote transformations that only partially correspond to the desired functionality. Dashes (-) denote transformations that are missing from the respective library. The fastest implementation of each transformation is highlighted in bold. The runtimes, presented in milliseconds, were averaged over 100 runs. All measurements were done with a single-channel volumetric input image of size (256, 256, 256) on a single workstation with a Ryzen 7-3700X CPU @ 3.60GHz.
| Transformation | TorchIO |
Volumentations |
Gunpowder |
Bio-Volumentations |
|---|---|---|---|---|
| Cropping | *27 ± 3 ms | 20 ± 3 ms | 6 ± 1 ms | 21 ± 3 ms |
| Padding | 64 ± 10 ms | 32 ± 6 ms | 64 ± 4 ms | 43 ± 2 ms |
| Flipping | 50 ± 13 ms | 40 ± 10 ms | 36 ± 3 ms | 31 ± 11 ms |
| Affine transform | 860 ± 24 ms | *3964 ± 816 ms | - | 2595 ± 39 ms |
| Affine transform (anisotropic) | - | - | - | 2608 ± 36 ms |
| Gaussian blur | 4417 ± 103 ms | - | - | 3017 ± 84 ms |
| Gaussian noise | 183 ± 7 ms | 393 ± 5 ms | *425 ± 14 ms | 266 ± 8 ms |
| Brightness and contrast change | - | 73 ± 2 ms | 197 ± 6 ms | 29 ± 4 ms |
| Z-normalization | 219 ± 11 ms | 124 ± 19 ms | - | 136 ± 13 ms |
[3] Pérez-García F, Sparks R, Ourselin S. TorchIO: A Python library for efficient loading, preprocessing, augmentation and patch-based sampling of medical images in deep learning. Comput Meth Prog Bio 2021;208:106236. https://www.sciencedirect.com/science/article/pii/S0169260721003102
[4] Volumentations maintainers and contributors. Volumentations 3D. Version 1.0.4 [software]. GitHub, 2020 [cited 2024 Dec 16]. https://github.com/ZFTurbo/volumentations
[5] Solovyev R, Kalinin AA, Gabruseva T. 3D convolutional neural networks for stalled brain capillary detection. Comput Biol Med 2022;141:105089. https://doi.org/10.1016/j.compbiomed.2021.105089
[6] Gunpowder maintainers and contributors. Gunpowder. Version 1.4.0 [software]. GitHub, 2024 [cited 2024 Dec 16]. https://github.com/funkelab/gunpowder
Authors of Bio-Volumentations: Samuel Šuľan, Lucia Hradecká, Filip Lux.
- Lucia Hradecká: [email protected]
- Filip Lux: [email protected]
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.
If you find our library useful, please cite its Zenodo record as:
Hradecká, L., & Lux, F. (2025). Bio-Volumentations (1.3.2). Zenodo. https://doi.org/10.5281/zenodo.15024087
and its journal publication as:
Lucia Hradecká, Filip Lux, Samuel Šuľan, Petr Matula. Bio-Volumentations: A Python library for augmentation
of volumetric image sequences. SoftwareX, 2025, vol. 30, 102151. DOI: 10.1016/j.softx.2025.102151.
Thank you! :)