Synapse Detection¶
What is synapse detection?¶
In connectome mapping, tracing out every neuron is not enough. We also need to know who talks to whom -- which neurons are pre-synaptic and which are post-synaptic. Synapses are the intricate junctions in neural space where information is exchanged between neurons.
In volume Electron Microscopy (vEM), with appropriate staining, we can visualise these synaptic junctions as darkly stained regions either between neuronal membranes (synaptic clefts) or attached to neuronal membranes. In the invertebrate datasets used in Catena, synapses typically appear as:
- T-bar -- like structures at pre-synaptic sites (also called active zones), and
- fuzzy, dark vesicle clouds at post-synaptic sites.
These ultrastructures are particularly well resolved in anisotropic vEM at resolutions around 4 x 4 x 40 nm in XYZ, and become noticeably fuzzier (especially on the post-synaptic side) in isotropic vEM.
Synapse detection is the task of automatically finding these pre- and post-synaptic sites -- and, ideally, linking them into directed edges between neurons.
Why do we care?¶
Accurate synapse detection is central to connectomics because it lets us:
- turn a neuron segmentation into a connectivity graph,
- quantify convergence/divergence patterns and motifs,
- and study how circuit structure might give rise to behaviour.
Without synapses, a beautifully segmented neuron volume is just geometry.
Cleft segmentation vs point-based methods¶
Several supervised machine learning methods have shown excellent performance in detecting synaptic ultrastructures. Historically, a common strategy has been cleft segmentation:
- Annotate synaptic clefts with voxel-wise masks.
- Train a network to segment cleft voxels.
- Post-process these masks into discrete synaptic objects and then into edges between neurons (often using another ML network for polarity assignment).
While effective, cleft masks are tedious to generate: they require dense, voxel-level annotation for every synapse, which does not scale well to large volumes or multiple datasets.
To reduce annotation burden, more recent approaches such as Synful, AdaSyn, and SimpSyn operate on point annotations:
- Human annotators mark synapses with one or a few points per site (e.g. pre- and post-synaptic centroids).
- Networks are trained to detect or refine these point locations, sometimes together with small local masks.
- Partner assignment is then done locally, using geometry, segmentation, or jointly learned directionality.
Point annotations are far more efficient to collect than full cleft masks, making it realistic to:
- label more synapses per hour,
- cover more brain regions or multiple animals,
- and iterate faster when the dataset or staining changes.
Catena's synapse module is designed with this point-based mindset: it aims to work with lightweight, human-friendly annotations while staying compatible with more traditional cleft-based workflows if you already have them.
Data Curation¶
Currently, synapse annotations exist multiple formats. They can be stored as in the WASPSYN challenge as .h5 or in CREMI format as .hdf or even as .jsons and .zarrs.
Data is also often stored a XYZ or ZYX or even ZXY. This causes a lot of discrepancies during model training.
Within Catena, we spent a lot of time (because it is a hard job) to curate these these datasets into one format. We chose the CREMI format since it is directly compatible with Synful and can be extended to support other datasets. CREMI format enables us store all data in ZYX, all annotations in nm space and we explicitly mention each RoIs resolution and offset within .hdf itself. Some examples are below in Figure1
We share our conversion code here.
Software¶
We use predominantly use CATMAID for annotating synapses.
For proofreading detections, we import them into CATMAID and then classify predictions into correct, incorrect and uncertain. More details on how we do it can be found in importing-detected-synapses.