Neuron Segmentation

This project is for training a U-net model to segment the soma and neurites of a neuron.

This project is unique in that the imagery is from a simulated fluorescent confocal microscope and the neuron geometry is from a collection of digitally reconstructed neurons. This allows the model to be trained with physically based parameters, such as the Z step distance. The sister project for simulating microscopes can be found here. We plan on expanding the capabilities of the simulation to include phase contrast microscopy and more physically-based parameterization of confocal fluorescent microscopes. We source the neuron models from NeuroMorpho and support reading standardized SWC files.

Project Structure

| - notebooks/
|   | - 00_Data_Generation.ipynb  # Use this notebook first, to generate the training data.
|   | - 01_Model_Training.ipynb   # This trains a variant of U-net to segment the neurons.
| - docs/
|   | - README.md                 # This file.
| - code/                         # A directory containing the source code for the demo web page.
| - demo/                         # The pre-built demo web page

Setup with HP AI Studio

We use HP AI Studio to generate data and train models. This allows us to iterate quickly between tweaking physical parameters of the microscope and training the model. If you're familiar with how to setup and run Jupyter, MLFlow, and Tensorboard, you may also set those up and run the code with your own configuration. To get started with HP AI Studio:

Step	Instructions
1	Create a new project in HP AI Studio
2	Set the Git Repository URL to https://github.com/tay10r/neuron-segmentation.git
3	Choose the Git Local Folder so that the cloned repo has somewhere to live.
4	Proceed to Create a Workspace and select the Deep Learning GPU template.
5	Name the workspace (for example: Neuron Segmentation Workspace)
6	Open a new Terminal in Jupyter and navigate to the repo at neuron-segmentation
7	Install Embree (for the simulator): conda install conda-forge::embree
8	Install Python Dependencies: pip3 install -r requirements.txt

Running Project

The project is meant to be run using the notebooks. Assuming you have run the setup instructions listed above, you can start running the notebooks. The notebooks are numbered by the order in which they should be executed.

• Run 00_Data_Generation.ipynb in order to generate training and validation data.
• Next, you may run the training notebook, 01_Model_Training.ipynb in order to train a model.
• Finally, you may run the deployment notebook, 02_Deploy.ipynb in order to deploy the model to a local MLFlow server. If you are running this in HP AI Studio, you may also use the deployed model and demo web page after running the last notebook.

If you've deployed the model with the last notebook, you can deploy the model in HP AI Studio. The web page will be hosted on localhost with the port being listed in the HP AI Studio UI. With the hostname and port number, you can navigate to the web page for testing the network using the following format:

http://localhost:<PORT>/demo/index.html

Where <PORT> should be replaced with the port number you see listed in the HP AI Studio Deployments tab.