EM-R50 (Synaptic Polarity Detection)¶

This tutorial covers synaptic polarity detection — predicting separated pre-synaptic and post-synaptic masks so the signal flow between neurons can be traced. The dataset was released by Lin et al. (2020) from Layer II/III of the primary visual cortex of an adult rat.

Unlike the cleft-detection task, polarity detection requires distinguishing individual synapses and assigning a side (pre / post) to each. The model outputs three channels — pre-synaptic, post-synaptic, and synaptic (union) — and a connected-components post-processor groups them into per-synapse instances.

Note

A dedicated EM-R50 tutorial config does not yet ship under tutorials/ in the modern (v2/v3) PyTC layout. The decoder and target-generation building blocks are present (polarity2instance decoder; seg_to_polarity target), but a canonical end-to-end YAML is a follow-up. The recipe below shows how to assemble one from the existing CREMI cleft-detection config (tutorials/syn_cremi.yaml) plus the polarity-specific pieces.

Goal¶

The intended pipeline:

Model output 3 channels — pre-synaptic, post-synaptic, and synaptic-union.
Target generation seg_to_polarity over instance labels; set exclusive=true if pre / post should never overlap (uses softmax) or false for the standard non-exclusive BCE setup (channels predicted independently with sigmoid).
Loss WeightedBCEWithLogitsLoss with rejection sampling on the foreground (synapses are sparse), or WeightedCE for the exclusive variant.
Decoder polarity2instance — connected-components on the synaptic-union channel, then voting on pre / post per instance.
Metric an IoU-based F1 score on instance-matched pairs.

1 - Get the data¶

mkdir -p datasets/jwr15 && cd datasets/jwr15
wget http://rhoana.rc.fas.harvard.edu/dataset/jwr15_synapse.zip
unzip jwr15_synapse.zip

2 - Building a v2 config (template)¶

Start from tutorials/syn_cremi.yaml and apply three changes:

(a) Switch the model output to 3 channels and use a non-exclusive BCE loss (or WeightedCE if you prefer exclusive masks):

default:
  model:
    out_channels: 3
    loss:
      losses:
      - function: WeightedBCEWithLogitsLoss
        weight: 1.0
        kwargs: {reduction: mean}
        pred_slice: "0:3"
        target_slice: "0:3"

(b) Drive target generation through the polarity helper. In your data.label_transform block, generate the 3-channel polarity target from instance labels by post-processing your label volume through connectomics.data.processing.target.seg_to_polarity (invoke from a small data-preparation script ahead of training, or add it as a custom processing step). Disable the cleft binary path that ships in syn_cremi.yaml.

Add the polarity decoder under decoding.steps:

decoding:
  steps:
  - name: polarity2instance
    kwargs:
      # default: non-exclusive (independent BCE channels);
      # set true if you trained with WeightedCE / softmax.
      exclusive: false

The rest of the CREMI config (RSUNet, anisotropic 18 × 256 × 256 patches, sliding-window inference, EMA, augmentation profile) carries over.

Tip

Synapses are sparse, so add rejection sampling to focus training on patches that contain foreground:
data:
  dataloader:
    reject_sampling:
      size_thres: 1000
      p: 0.95
This is what syn_cremi.yaml already does for cleft detection; keep it for polarity training.

3 - Run training¶

Once your config is assembled (e.g. saved as tutorials/syn_em_r50.yaml):

conda activate pytc
python scripts/main.py --config tutorials/syn_em_r50.yaml

4 - Inference, decoding, evaluation¶

python scripts/main.py --config tutorials/syn_em_r50.yaml \
    --mode test \
    --checkpoint outputs/<exp>/<timestamp>/checkpoints/last.ckpt

The decoding stage runs polarity2instance, which converts the 3-channel probability map into per-synapse instance masks with a recorded side (pre vs. post). For programmatic use:

from connectomics.decoding import polarity2instance
# volume: (3, D, H, W) prediction, channels = [pre, post, syn]
instances = polarity2instance(volume)

The exclusive=True variant interprets the channels as a softmax output over {background, pre, post} and uses a different connected- component path; pass exclusive=True to polarity2instance if your model was trained with WeightedCE.

5 - Reference behavior¶

Training loss is dominated by the synaptic-union channel early on; the pre / post channels improve as the model learns spatial directionality, typically after the first cosine decay phase. Rejection sampling plus higher foreground weights is essential — without it, pre and post collapse to all-zeros.
Inference is identical to the CREMI flow (same RSUNet, same sliding window). The decoder runs in CPU and is fast.
Reporting the published EM-R50 F1 is computed against the Lin et al. 2020 evaluation script. Comparable code lives in the upstream paper repo; contact the maintainers if you intend to benchmark formally.