About

Magneton is a neuron segmentation pipeline developed by Kuan Lab. This pipeline employs a chunked based mode for processing large-scale 3D EM data in neuron segmentation tasks. It is currently available as a CLI tool, supporting both local jobs and HPC jobs.

Based on Linux machines with NVIDIA GPUs

Magneton consists of 3 main modules:

  1. Data pre- and post-processing toolkit

  • Resize a tif data

  • Split a big tif data to blocks

  • Merge h5 files as a tif volume

  • Convert a tif/h5 file as a precomputed data

  • Downsample a precomputed data

  • Generate a mask for affinity map

  • Mask a tif/precomputed data

  1. Deep learning based affinity maps inference

Based on pytorch connectomics.This is a deep learning framework for automatic and semi-automatic annotation of connectomics datasets, powered by pytorch.

  • Pre-tain / fine-tune a DL model

  • Inference affinity map by blocks

  1. Instance segmentation for affinity maps

Support global segmentation and large-volume chunking parallel processing, 2D supervoxel and 3D supervoxel

  • Affinity map blocks instance segmentation

  • Aggregation of each blocked segmentation result