SmartEM: machine-learning guided electron microscopy

Area of Science:

• Neuroscience
• Microscopy
• Artificial Intelligence

Background:

• Connectomics requires high-resolution neural circuit maps for understanding brain function.
• Generating these maps via electron microscopy (EM) is time-consuming and data-intensive.
• Current machine learning aids post-imaging analysis, making image acquisition the bottleneck.

Purpose of the Study:

• To accelerate EM image acquisition for connectomics by integrating machine learning into the imaging process.
• To develop an intelligent, data-aware imaging system for electron microscopes.

Main Methods:

• Integration of machine learning into real-time image acquisition in a single-beam scanning electron microscope (SEM).
• SmartEM intelligently allocates imaging time, prioritizing regions needing higher signal quality for accurate segmentation.
• The system performs rapid initial scans followed by slower re-scans of critical subareas.

Main Results:

• Achieved a 7-fold acceleration in EM image acquisition time for connectomics.
• Demonstrated the ability to reconstruct a portion of mouse cortex with accuracy comparable to traditional methods.
• Significantly reduced overall time required for generating neural circuit maps.

Conclusions:

• SmartEM offers a significant advancement in accelerating EM-based connectomics.
• Intelligent, real-time data acquisition can overcome the bottleneck in generating large-scale neural circuit maps.
• This approach makes high-resolution brain mapping more accessible and efficient for researchers.