Important
This repository is used to reproduce the results fron the paper, it leverages the version of CAREamics used at the time of the work. We are currently refactoring the code:
- Compatibility with the more recent CAREamics version
- Better usability, simpler configuration
- All in CAREamics, no extra repository
Stay tune for the CAREamics MicroSplit release!
MicroSplit is a deep learning-based computational multiplexing technique that enhances the imaging of multiple cellular structures within a single fluorescent channel, allowing to image more cellular structures, image them faster, and at reduced overall light exposure.
MicroSplit is based on a hierarchical variational auto-encoder (LVAE) using lateral context.
MicroSplit is implemented in the CAREamics library, and this repository contains example notebooks and utilities for reproducing the results of the MicroSplit paper.
Important
A GPU is necessary for training the models from scratch. For users interested in testing the examples from the paper, our notebooks allow loading pre-trained models and running the inference even without GPU access.
Depending in your system, the installatioin will take between 5 and 10 minutes, on systems that also require a fresh CONDA/MAMBA installation maybe even a few minutes longer.
In order to run the examples, you will need to install PyTorch, CAREamics and the utilities in this repository.
- Create a new environment with the package manager of your choice, we recommand mamba, but you can also use conda (in which case, substitute
mambaforcondain the following bash commands).mamba create -n microsplit python=3.10 mamba activate microsplit
Tip
If you are on a mac, and wish to make use of Mac silicon GPU (using M1, M2 and M3 chips), create the environment using the following commands:
CONDA_SUBDIR=osx-arm64 conda create -n microsplit python=3.9
conda activate microsplit
conda config --env --set subdir osx-arm64-
⚠️ Install PyTorch following the instructions on the official website. -
You can test that you have GPU access by running the following command:
python -c "import torch; print([torch.cuda.get_device_properties(i) for i in range(torch.cuda.device_count())])"To confirm that mac silicon is available do:
python -c "import torch; import platform; print(platform.processor() in ('arm', 'arm64') and torch.backends.mps.is_available())" -
Install MicroSplit utilities from this repository by cloning and navigating into it, then by installing all the necessary packages using
pip.git clone https://github.com/CAREamics/MicroSplit-reproducibility.git cd MicroSplit-reproducibility pip install .
Tip
If you are on a Windows machine and have trouble running unix-like commands, check out Git for Windows. This tool installs Git Bash, a terminal that you can use to run the commands above.
-
You can now open the notebooks in
jupyterby running the following command and navigating to the example folder:jupyter notebook
Note
The Jupyter notebooks in each example are numbered by their order in the MicroSplit pipeline:
- 00: Create the noise models for the dataset
- 01: Train the MicroSplit model
- 02: Apply MicroSplit to data
- 03: Calibrate the MicroSplit errors
The notebooks are designed to be run in order, but we designed them so that each notebook, except the calibration, has entry points using pre-trained models. Hence, depending on your interests, you can test MicroSplit within some minutes of browsing/running these Jupyter Notebooks. If you want to run each step, you will likely need 30-60 minutes, if you plan to fully train all networks, and depending on your GPU, this process can take several hours.
A list of problems that might be encountered and suggestions on how to solve them.
-
Problem: An error saying that your NVIDIA Drivers are too old.
Solution: Try downgrading your PyTorch version, for example:
pip3 install torch==2.2 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
-
Problem: Test for Mac silicon GPU above returns False.
Solution: Make sure you install PyTorch with pip, installing it with conda might not work. Make sure you installed the macOS-arm64 release of Anaconda or Mamba.
The notebooks were tested on multiple systems including Linux, Mac and Windows; more detailed system specifications are as follows:
- OS Version: Red Hat Enterprise Linux 8.10
- GPU: NVIDIA A40-16Q, 16GB
- CudaToolKit Version: 12.4
- OS Version: macOS 14.1
- GPU: Apple M3 GPU, 16GB
- OS Version: Windows 10 Enterprise
- GPU: Nvidia RTX A3000 Laptop, 6GB
- CudaToolKit Version: 12.3
(soon)
Ashesh, Ashesh, et al. "Micro Split: semantic unmixing of fluorescent microscopy data." Nature Methods (2026): 1-11. DOI:10.1038/s41592-026-03082-1
@article{ashesh2026micro,
title={Micro S plit: semantic unmixing of fluorescent microscopy data},
author={Ashesh, Ashesh and Carrara, Federico and Zubarev, Igor and Galinova, Vera and Croft, Melisande and Pezzotti, Melissa and Gong, Daozheng and Casagrande, Francesca and Colombo, Elisa and Giussani, Stefania and others},
journal={Nature Methods},
pages={1--11},
year={2026},
publisher={Nature Publishing Group US New York}
}
This project is licensed under BSD-3-Clause License - see the LICENSE for details.