Optimizing and Benchmarking Machine Learning and Traditional Synaptic Event Detection Pipelines in Neurophysiology...

Area of Science:

• Neuroscience
• Computational Neuroscience
• Electrophysiology

Background:

• Accurate detection of synaptic currents is crucial for high-quality neuroscience experiments.
• Traditional manual event counting for synaptic currents is time-consuming and labor-intensive.
• Automated methods, including machine learning, offer potential for faster and more accurate event detection.

Purpose of the Study:

• Establish a practical ground truth for synaptic event detection using meticulous hand counting.
• Quantitatively compare the accuracy of various detection methods across different laboratories and cell types.
• Benchmark popular automated detection strategies, including a supervised deep learning approach, against manual analysis.

Main Methods:

• Conducted extensive synaptic physiology experiments.
• Performed meticulous hand-counting of synaptic events to establish a ground truth dataset.
• Benchmarked automated detection algorithms, including a supervised deep learning model, against the hand-counted ground truth.

Main Results:

• Significant variability exists in detection results across different laboratories and analysis techniques.
• A supervised deep learning approach demonstrated accuracy comparable to manual event counting by expert electrophysiologists.
• Automated methods showed varying degrees of accuracy, with deep learning outperforming other automated approaches.

Conclusions:

• Current automated synaptic event detection strategies exhibit considerable variance in performance.
• Supervised deep learning offers a promising alternative to manual analysis, rivaling expert performance.
• Establishing standardized ground truth datasets is essential for reliable benchmarking of synaptic event detection techniques.