Skip to main content

load bioimages for machine learning

Project description


Load bioimages for machine learning applications

Python version PyPI version PyPI download per month License

bioimageloader is a python library to make it easy to load bioimage datasets for machine learning and deep learning. Bioimages come in numerous and inhomogeneous forms. bioimageloader attempts to wrap them in unified interfaces, so that you can easily concatenate, perform image augmentation, and batch-load them.

bioimageloader provides

  1. collections of interfaces for popular and public bioimage datasets
  2. image augmentation using albumentations, which is popular and powerful image augmentation library (for 2D images)
  3. compatibility with pytorch
  4. and with others such as scikit-learn and tensorflow

Table of Contents

Quick overview

Find full guides at bioimageloader-docs:User Guides

  1. Load a single dataset

    Load and iterate 2018 Data Science Bowl

    from bioimageloader.collections import DSB2018
    import albumentations as A
    transforms = A.Compose([
        A.RandomCrop(width=256, height=256),
    dsb2018 = DSB2018('path/to/root_dir', transforms=transforms)
    for data in dsb2018:
        image = data['image']
        mask = data['mask']
  2. Load multiple datasets

    Load DSB2018 and Triple Negative Breast Cancer (TNBC)

    from bioimageloader import Config, ConcatDataset
    from bioimageloader.collections import DSB2018, TNBC
    import albumentations as A
    transforms = A.Compose([
        A.RandomCrop(width=256, height=256),
    cfg = {
        'DSB2018': { 'root_dir': 'path/to/root_dir' },
        'TNBC'   : { 'root_dir': 'path/to/root_dir' },
    config = Config.from_dict(cfg)
    datasets = config.load_datasets(transforms=transforms)
    cat = ConcatDataset(datasets)
    for meow in cat:
        image = meow['image']
        mask = meow['mask']
  3. Batch-load dataset

    from bioimageloader import BatchDataloader
    call_cat = BatchDataloader(cat,
    for meow in call_cat:
        batch_image = meow['image']
        batch_mask = meow['mask']

    or directly use pytorch's DataLoader

    from import DataLoader
    call_cat = Dataloader(cat,
    for meow in call_cat:
        batch_image = meow['image']
        batch_mask = meow['mask']

bioimageloader is not/does not

  • not a full pipeline for ML/DL
  • not a hub to bioimage datasets (if it ever becomes one, it would be awesome though)
  • does not host data (only interfaces)
  • does not provide one-click links for downloading data
  • does not overwrite the source data

Why bioimageloader

bioimagesloader is a by-product of my thesis. This library collected bioimage datasets for machine learning and deep learning. I needed a lot of diverse bioimages for self-supervised neural networks for my thesis. While I managed to find many great datasets, they all came with different folder structures and formats. In addition, I encountered many issues to load and process them, which were sometimes technical or just rooted from the nature of bioimages.

For instances of technical issues, some datasets were missing one or two pairs of image and annotation, had broken files, had very specific file formats that cannot be easily read in python, or provided mask annotation not in image format but in .xml format. Some filenames have typos, so sometimes I failed to iterate them.

For an example of intrinsic issues of bioimages, selecting a certain channel was an important functionality that I needed, and it was not easy for bioimage datasets. When a dataset provided separate files for each channel image, it was easy to select one. But in many cases, they just put all channels together in one image file. And even worse for 2 channel images (which are quite common), if they chose to use RGB(A) image formats such as JPEG or PNG other than TIFF, I needed to figure out manually which channel refers to what and which channel is the empty one.

There were other issues not mentioned above of course. It was rather painful to deal with all these edge cases one by one. But anyway I did it and I thought it would be valuable to package and share it with community so that others do not have to suffer, even though the number of implemented datasets is small for the moment,


Install the latest version from PyPI. bioimageloader requires Python 3.8 or higher. Find more options at bioimageloader-docs:Installation

pip install bioimageloader


Full documentation is available at bioimageloader-docs

Available collections

Go to bioimageloader-docs:Catalogue


Why no direct download link to each dataset?

bioimageloader provides only codes (interfaces) to load data but not data itself. We believe that it is important for you to go there, read papers, understand terms and licenses to appreciate their works, because bioimages themselves are sciences and results of time, efforts, and resources. You still can find links to their project pages or papers at bioimageloader-docs:Catalogue, and you need to follow their instruction to get data. Once you downloaded a dataset and unzipped it, (if it is supported by bioimageloader) you simply pass its root directory as the first argument to corresponding class from collections bioimageloader.collections.

Dataset that I want is not in the bioimageloader-docs:Catalogue

First of all, I named each dataset class rather arbitrary. Try to find the dataset you want with authors' names or with other keywords (if it has any), and you may find it having an unexpected name. If it is the case, I apologize for bad naming.

If you still cannot find it, then you have two options: either you do it yourself (see below question and please consider contributing!), or you can file an issue so that the community can help.

Don't know how to write my own dataloader.

Writing a dataloader requires a bit of Python skills. No easy way. Please read templates carefully and see how others are implemented. File an issue, and I am willing to help.

How to run a ML/DL model?

bioimageloader only helps loading images/annotations, not running ML/DL models. Still, you may find some useful examples at bioimageloader-docs:User Guides. Also check out ZeroCostDL4Mic.

I want more granular control over datasets individually

Each bioimage dataset is very unique and it is natural that users want more controls and it was true for my work as well. Good news is that bioimageloader suggests a template that you can extend from and make a subclass in your liking. Bad news is that you need to know how to make a subclass in Python (not so bad I hope. I suppose that you may have knowledge of Python, if you want to develop ML/DL in Python anyway). This guide Modifying existing collections covers it.


Find guide at bioimageloader-docs:Contributing

Also check out TODO list.


I am open to any feedbacks, suggestions, and discussions. Reach out to me by github or email.

Seongbin Lim

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

bioimageloader-0.1.1.tar.gz (60.4 kB view hashes)

Uploaded Source

Built Distribution

bioimageloader-0.1.1-py3-none-any.whl (101.6 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page