SynReEM: Synapse Reconstruction via Instance Structure Encoding in Anisotropic vEM Images

We develop SynReEM, an end-to-end synapse reconstruction framework for anisotropic volume electron microscopy scenarios.

Table of Contents

• Overview
• Installation
• Getting Started
  • AEMC Data Preprocessing
  • Segmentation process
  • Graph aggregation reconstruction
• Acknowledgments

Overview

The basic idea of SynReEM

The basic architecture of SynReEM

The Reconstruction Process of SynReEM

Synapse reconstruction results display

MICrONS

MICrONS.340p.mp4

Kasthuri

Kasthuri.340p.mp4

High-definition (HD) versions of the video can be found here.

Video download: google drive

Dataset

Synapse178: google drive

SynWTAD: google drive

Installation

1. Clone the repository:

git clone https://github.com/fenglingbai/SynReEM.git
cd ~/SynReEM

2. Create and activate the conda environment:

conda create --name SynReEM --file environment.txt -y
conda activate SynReEM

3. Verify PyTorch installation:

python -c 'import torch;print(torch.backends.cudnn.version())'
python -c 'import torch;print(torch.__version__)'

4. Install the nnUNet environment (integrated with this framework):

pip install -e .

Getting Started

AEMC Data Preprocessing

SynReEM first requires converting instance labels to AEMC labels to facilitate model learning.

Refer to the demo script for implementation details:

SynReEM\scripts\data_encode_demo.py

Segmentation process

Data Preparation

Convert original tif data to nii.gz format compatible with the nnUNet framework:

python SynReEM\nnunet\dataset_conversion\Task603_synapse178synins.py

Set up the experimental plan:

python SynReEM\nnunet\experiment_planning\nUNet_plan_and_preprocess.py -t XXX --verify_dataset_integrity

Training

python SynReEM\nnunet\run\run_training_synreem.py 3d_fullres SynReEMTrainer TaskXXX_MYTASK FOLD --npz

Example:

python SynReEM\nnunet\run\run_training_synreem.py 3d_fullres SynReEMTrainer Task603_synapse178synins 4 --npz

Inference

python SynReEM\nnunet\inference\predict_synreem.py -i INPUT_FOLDER -o OUTPUT_FOLDER -t TASK_NAME_OR_ID -m CONFIGURATION --save_npz

Graph aggregation reconstruction

Convert AEMC labels back to instance labels for final reconstruction results.

Refer to the demo script for implementation details:

SynReEM\scripts\data_decode_demo.py

Acknowledgments

This project builds upon the following open-source frameworks:

• nnUNet (https://github.com/MIC-DKFZ/nnUNet/tree/nnunetv1)
• EmbedSeg (https://github.com/juglab/EmbedSeg)