SyConn
Automated synaptic connectivity inference
for volume electron microscopy
In short: Teravoxel volume electron microscopy datasets from neural tissue can now be acquired in weeks, but data analysis requires years of manual labor (see also Fig. 1). The SyConn framework uses deep convolutional neural networks (CNN) and random forest classifiers (RFC) to infer a richly-annotated synaptic connectivity matrix from manual neurite skeleton reconstructions by automatically identifying mitochondria, synapses and their types, axons, dendrites, spines, myelin, somata and cell types.
SyConn automates all steps shown in Fig. 1, but the neurite reconstruction, where it only requires manual skeleton reconstructions instead of volume segmentations (compare hatched bars and full bars in Fig. 1). Using several recursive 3D CNNs we first detect cell boundaries and all kinds of ultrastructures such as synaptic junctions (Fig. 2b). Using a ray cast approach we extract a hull for each skeleton based on the boundary prediction (Fig. 2a), which is then used to map mitochdondria, synaptic junctions and vesicle clouds to each cell (Fig. 2c). Leveraging this information SyConn uses RFCs to partition neurons into their subcellular parts (axon, dendrite, soma; Fig. 2d) and to assign them to one of four broad cell types (Fig. 2e). Finally, we combine this into a richly annotated synaptic connectivity matrix (Fig. 2f).
To test SyConn, we detected ultrastructural objects in three serial block-face electron microscopy (SBEM) data sets (zebra finch Area X, size: 97.9 × 95.6 × 115 μm³; zebrafish spinal cord, size: 81.8 × 89.9 × 210 μm³; mouse striatum, size: 17.9 × 15.0 × 69 μm³) and created a wiring diagram of the zebra finch basal-ganglia (Area X) from 612 skeleton reconstructions traced with KNOSSOS (knossostool.org).
Getting Started
Python 2.7 is required (we recommend Anaconda). SyConn has been tested on Linux distributions (CentOS and Arch Linux).
Quickstart
1 2 3 4 5 6
All requirements should be automatically installed, when following the instructions. Together with the provided sample data (SyConnDenseCube.zip) including our trained models one is able to run most parts of SyConn.
Ground Truth
The provided sample package contains trained CNNs, but no ground truth. Ground truth can be downloaded from the provided links below. The readme file in the zip container explains how the ground truth was used for training in SyConn.
Zebra finch:
Synaptic junctions, mitochondria and vesicle clouds: gt_syconn_bird_sj_vc_mi.zip (5GB)
Publication
SyConn was published in Nature Methods. If you use parts of this code base in your academic projects, please cite the corresponding publication.
BibTex
Contributors
Sven Dorkenwald
SyConn development and evaluation
Marius Killinger and Gregor Urban
ELEKTRONN development.
Philipp Schubert
SyConn development and evaluation
Fabian Svara and Shawn Mikula
Dataset contributions
Jörgen Kornfeld
SyConn development and evaluation
SyConn v1 was developed in the Denk department at the MPI for Medical Reserach in Heidelberg. The current version is being developed in the Kornfeld lab at the MPI of Biological Intelligence (in foundation) in Martinsried, Germany.
References
SyConn makes use of the ELEKTRONN neural network toolkit (elektronn.org) and the KNOSSOS Python tools (github). Neuroglancer and KNOSSOS (knossostool.org) are used for visualization and annotation of 3D EM data sets.