Maskflow: Object Detection and Segmentation for Cell Biology

WARNING: this project is in development. Come back later for an alpha version you can play with.

Maskflow is a Python framework for instance object detetion and segmentation using neural networks. It is focused on processing biological images but aims to be general enough to process any kind of dataset.

The framework aims to be:

• easy-to-use: Any person with a minimal Python experience should be able to train and deploy a model.
• flexible: While we implement a default model that works well we want to allow the creation of new models easily by just swapping sub-models (replace the ResNet backbone by a DenseNet backbone for example).
• universal: You should be able to predict labels, bounding boxes, masks. Moreover why do we have to restrict to those when you can predict any kind of polygon describing your objects (polygon detections is not yet implemented)?
• easy-to-deploy: Deploy your model using Docker and run inference by sending REST calls to it.
• test-driven: Maskflow development is test-driven in order to provide a robust and stable API.
• API first: Maskflow is library providing an API and not a software. A software n be built on top of Maskflow (see Maskflow Fiji plugin).

As Maskflow is originally designed for biological images, it comes with a Fiji plugin that can run a model previously trained (the plugin will be available very soon).

We use Python >= 3.6 an TensorFlow >= 2.0.

Trained Models

Check the notebooks to train your own dataset.

In addition to this, Maskflow comes with multiples ready-to-build-and-train datasets. The notebooks provide push-and-start code to build and train a model.

• Shapes: A "toy" dataset that can be trained very quickly for debugging and educational purpose.
• Nucleus: A dataset composed of cell nucleus with various image acquisition modalities.
• Synthetic_Cells: Virtual cell image generated in silico.
• Microtubule: Dataset consists of various simulated in vitro microtubule images.

Usage

We suggest you to use Anaconda as a Python distribution.

# Create a Python environement.
conda create -n my_env python

# Activate it
conda activate my_env

# Install maskflow dependencies
conda env create -f environment.yml

Tensorflow needs to be installed with Pip at the moment. Check the environment.yml file for the pip commands to install Tensorflow.

Then you can navigate to the notebooks in order to train and pack your model. Maskflow is also a perfectly valid Python package that you can install via pip:

pip install maskflow
# or
pip install https://github.com/hadim/maskflow/archive/master.zip

License

Under BSD license. See LICENSE.

Authors

• Hadrien Mary hadrien.mary@gmail.com