SmartEM: machine learning-guided electron microscopy.

Area of Science:

• Neuroscience
• Electron Microscopy
• Machine Learning

Background:

• Connectomics aims to map neural circuits at synapse resolution for understanding brain function.
• High-throughput electron microscopy is essential but limited by data acquisition speed.
• Current machine learning methods analyze images post-acquisition, making imaging the bottleneck.

Purpose of the Study:

• To accelerate electron microscopy (EM) image acquisition for connectomics.
• To integrate machine learning into the real-time imaging process.
• To overcome the data acquisition bottleneck in automated connectomics.

Main Methods:

• Developed SmartEM, a system integrating machine learning into real-time EM image acquisition.
• Implemented a data-aware imaging strategy: rapid initial scan followed by focused rescanning of regions of interest.
• Utilized a commercial single-beam scanning electron microscope (SEM) for demonstrations.

Main Results:

• Achieved up to a ~7-fold acceleration in image acquisition time for connectomic samples.
• Demonstrated SmartEM's effectiveness across various sample types (nematodes, mice, human brain).
• Successfully reconstructed a portion of mouse cerebral cortex with accuracy comparable to traditional EM.

Conclusions:

• SmartEM significantly speeds up EM image acquisition for connectomics.
• This approach makes high-resolution neural circuit mapping more accessible and efficient.
• SmartEM addresses the critical rate-limiting step in automated connectomics research.