GSoC NIU Projects 2025: BrainGlobe#
BrainGlobe is a community-driven suite of open-source Python tools. The BrainGlobe tools are widely used to process, analyse and visualise images of brains (and other related data) in neuroscientific research. If you are interested in any of these projects, get in touch! Feel free to open a new topic on Zulip and tag the potential mentors.
Our working language is English, but our mentors for these projects also speak Italian, French, and German.
Improve cellfinder’s classification algorithm
The BrainGlobe cellfinder tool is used to detect cells in large whole-brain images. It uses traditional image processing to find possible cell candidates and passes them to a customisable classifier to split the candidates into cells and no-cells. cellfinder relies heavily on pytorch and keras.
cellfinder currently uses a residual neural network (ResNet) to classify cell candidates, and deep learning network architectures have progressed since. In this project, you will explore newer architectures for the classification network as alternatives to ResNet, and see whether they work better and/or faster than the existing implementation.
Deliverables
A Python implementation of at least one new Deep Learning network architecture in
cellfinderQuantitative comparison between the current and new architecture
Tests to cover any added functionality.
Documentation for the new functionality.
A blog explaining the new network and its advantages and disadvantages.
Duration
Medium (~175 hours) or Large (~350 hours).
Difficulty
This project is well-suited for an intermediate contributor to open source.
Required skills
At least initial understanding of machine learning algorithms
Nice-to-haves
Experience working with machine learning frameworks, in particular
kerasorpytorchExperience working with image data
Potential mentors
Further reading
cellfinder support for two-dimensional brain images
The BrainGlobe cellfinder tool is used to detect cells in large whole-brain images. It uses traditional image processing to find possible cell candidates and passes them to a customisable classifier to split the candidates into cells and no-cells. cellfinder relies heavily on pytorch and keras.
cellfinder currently uses a residual neural network (ResNet) to classify cell candidates. This network is designed for three-dimensional whole-brain images, but neuroscientists often take images of two-dimensional slices of the brain. In this project, you would adapt cellfinder to support two-dimensional data. This would involve implementing a new neural network suitable for two-dimensional images, as well as refactoring some of the image processing code.
Deliverables
A modified implementation of the existing image processing algorithm (blob detection) in
cellfinderthat detects cell candidates in both two- and three-dimensional imagesA Python implementation of a neural network in
cellfinderthat classifies cell candidates from two-dimensional images.Tests to cover any added functionality.
Documentation for the new functionality.
A blog showcasing the use of
cellfinderon two-dimensional images of brains.
Duration
Large (~350 hours).
Difficulty
This project is well-suited for an intermediate contributor to open source.
Required skills
At least initial understanding of machine learning algorithms
Nice-to-haves
Experience working with machine learning frameworks, in particular
kerasorpytorchExperience working with image data
Potential mentors
Further reading
cellfinder support for arbitrary number of channels
The BrainGlobe cellfinder tool is used to detect cells in large whole-brain images. It uses traditional image processing to find possible cell candidates and passes them to a customisable classifier to split the candidates into cells and no-cells. cellfinder relies heavily on pytorch and keras.
cellfinder was originally designed to work with two channels: the “background” channel and the “signal” channel. This is a limitation, because sometimes only the signal channel is acquired, or there are several signal channels containing useful information. In this project, you would explore ways to allow cellfinder to classify cells based on information from an arbitrary number of channels. This could allow cellfinder to not just detect cells, but classify their type based on their shape and other features.
Deliverables
A modified implementation of the existing image processing algorithm (blob detection) in
cellfinderthat can detect cell candidates from an arbitrary number of channels.A modified implementation of the neural network in
cellfinderthat can classify cell candidates from brain images with an arbitrary number of channels.Tests to cover any added functionality.
Documentation for the new functionality.
A blog showcasing the use of
cellfinderon images with 1 or 3 channels.
Duration
Large (~350 hours).
Difficulty
This project is well-suited for an intermediate contributor to open source.
Required skills
At least initial understanding of machine learning algorithms
Nice-to-haves
Experience working with machine learning frameworks, in particular
kerasorpytorchExperience working with image data
Potential mentors
Further reading
Add to BrainGlobe’s data visualisation tools
The BrainGlobe brainrender tool is widely used to visualise brain data in a common coordinate space defined by a “brain atlas” (We refer to this data as “atlas-registered” data). However, brainrender is inaccessible to people without programming skills. The brainrender-napari tool aims to make brainrender functionality available to more people through a plugin for the popular open-source graphical image viewer napari.
Although brainrender and brainrender-napari share some functionality, some publicly available atlas-registered data is not yet available in brainrender-napari. In this project, we would implement code to allow users to visualise publicly available atlas-registered data from mouse and fish brains in brainrender-napari
Deliverables
A Python implementation of a
napariwidget that allows users to download and visualise atlas-registered data.Tests to cover any added functionality.
Documentation for the new functionality.
Duration
Small (~90 hours) or Medium (~175 hours).
Difficulty
This project is well-suited for a student or a beginner contributor to open source.
Required skills
Experience with Python.
Nice-to-haves
Experience working with data visualisation
Experience working with image data
Potential mentors
Further reading
napari Plugin documentation, particularly the section on Building a plugin.
Update atlas packaging scripts to atlas API v2
Neuroanatomical atlases define a common coordinate space for the brain, and have been created for many species (mouse, fish, …) and imaging modalities (MRI, lightsheet microscopy, …). A central piece of BrainGlobe’s ecosystem is the BrainGlobe Atlas API. Importantly, the Atlas API allows the use of any BrainGlobe tool with many publicly available atlases, by exposing them through the same, consistent Python interface. This is possible thanks to packaging scripts contributed by the community, that convert public data to a standard format.
The standard format is somewhat limited, and we improve its utility by adhering to a new community standard for neuroanatomical atlases, “OpenMINDS SANDS”.
Deliverables
Modifying atlas packaging Python code to write to OpenMINDS SANDS
Adapting at least one packaging script to use the new functionality.
Tests to cover any added/changed functionality.
Documentation for the new functionality.
Duration
Medium (~175 hours) or Large (~350 hours).
Difficulty
This project is well-suited for a student or a beginner contributor to open source.
Required skills
Experience with Python.
Nice-to-haves
Experience working with data standards or data schema
Experience working with image data
Potential mentors
Further reading
brainglobe-registration
The BrainGlobe brainglobe-registration tool is used to register 2D and 3D images to a common coordinate space defined by an atlas (`brainglobe-atlasapi’). This is a crucial step in many neuroscientific analyses, as it allows data from different experiments to be compared and combined.
The current implementation relies on a manual selection of the specific 2D region, or 3D subvolume of the atlas to be used as the registration target. This is a time-consuming process, and can be error-prone. In this project, you would implement and compare methods to automatically select the region to be registered, based on the data itself. This could be done by using an adaptive grid search, ML based techniques, or Bayesian optimisation.
Deliverables
A new preprocessing step in
brainglobe-registrationthat automatically selects the region of the atlas to be used as the registration target.Quantitative comparison between at least two different methods for selecting the region.
Tests to cover any added functionality.
Documentation for the new functionality.
A blog showcasing the new functionality.
Duration
Medium (~175 hours) or Large (~350 hours) depending on experience.
Difficulty
This project is well-suited for an intermediate contributor to open source.
Required skills
Experience with Python, NumPy.
At least initial understanding of machine learning algorithms
Nice-to-haves
Experience with image registration
Experience with
ITKorelastixExperience working with image data
Potential mentors
Further reading