ExLLSM Synapse Detector

This repository contains code to train and evaluate the performance of a 3D U-Net shaped model for synapse detection. It was designed to classify synaptic sites in large scale volumetric images generated via expansion microscopy and lattice light-sheet imaging (ExLLSM). The inference directory contains the detector that uses the trained model to make synapse predictions. This code is integrated into the ExLLSM Circuit Reconstruction Pipeline and currently meant to be run there, inside of a Docker container.

3D U-Net shaped network structure

Quick Start

First setup a conda environment using the following command:

conda create -n exllsm-synapse
conda env update -n exllsm-synapse -f environment.yml

then activate it: conda activate exllsm-synapse

Training

To train a new model, some parameters in the python script ExLLSM_unet_v2_train.py need to be adjusted:

• modify line 10 to point to the directory containing raw data and corresponding ground truth image crops (500x500x500 and 100x100x100 crops have been used for testing)
• modify line 11-13 (and add additional lines for additional data crops) to indicate the name of each raw data crop and each corresponding ground truth data crop (both in nrrd format)
• modify line 24 to point to the save directory
• modify line 25 to change training parameters if desired
• modify line 27 to change the pkl name if desired
• modify line 30 to name the final trained model

Usage:

 python ExLLSM_unet_v2_train.py

Plot training progress

To plot the training progress over time, modify lines 5 and 6 of the python script ExLLSM_unet_training_result_plot.py as indicated.

Usage:

 python ExLLSM_unet_training_result_plot.py

Model Evaluation

To appropriately evaluate the model performance it should be compared to ground truth data from independent samples (brains/animals) not involved in training.

It is recommended that the trained model is used via the ExLLSM Circuit Reconstruction Pipeline where additional recommended post-U-Net processing steps not included here are found.

To calculate the precision and recall scores of a model run via the ExLLSM Circuit Reconstruction Pipeline to ground truth data from independent samples, use the ExLLSM_unet_performance_evaluate.py script. Modify lines 50-52 as indicated and run the updated python script.

Usage:

 python ExLLSM_unet_performance_evaluate.py

Inference

The inference directory contains the detector that uses the trained model to make synapse predictions. This code is integrated into the Expansion Microscopy Pipeline and currently meant to be run there, inside of a Docker container.