Fully-Automatic Synapse Prediction and Validation on a Large Data Set.

Area of Science:

• Neuroscience
• Computational Biology
• Machine Learning

Background:

• Extracting neuronal connections (connectomes) from electron microscopy (EM) data is crucial but time-consuming.
• Manual tracing and synapse annotation represent significant bottlenecks in connectome reconstruction.
• Previous methods relied on neuronal shape and contact area, which are insufficient for accurate synapse prediction.

Purpose of the Study:

• To develop and evaluate algorithms for automatic detection of presynaptic and postsynaptic partners in EM data.
• To address the limitations of manual annotation and improve the efficiency of connectome reconstruction.
• To provide a complete solution for polyadic synapse detection and introduce novel evaluation metrics.

Main Methods:

• Utilized a U-Net convolutional neural network (CNN) for presynaptic structure detection.
• Employed a multilayer perceptron (MLP) with local segmentation features for postsynaptic partner detection.
• Developed novel metrics for evaluating algorithm performance on large-scale connectomes.

Main Results:

• The developed algorithms accurately detect presynaptic and postsynaptic partners, enabling automated synapse detection.
• The approach requires minimal training data and leverages advances in image segmentation.
• Evaluation on *Drosophila* EM data demonstrated effective characterization of neuronal connectivity.

Conclusions:

• Automated synapse detection significantly accelerates connectome reconstruction.
• The proposed CNN and MLP-based approach offers a complete solution for polyadic synapse detection.
• The novel metrics provide a robust framework for evaluating connectome reconstruction algorithms.