Self-supervised machine learning framework for high-throughput electron microscopy.

Area of Science:

• Materials Science
• Structural Biology
• Electron Microscopy

Background:

• Transmission electron microscopy (TEM) provides high spatiotemporal resolution for materials and biological structure analysis.
• Electron beams in EM are inherently damaging, limiting low-dose imaging applications.
• Current methods struggle with noise in low-dose EM, hindering detailed analysis.

Purpose of the Study:

• To introduce SHINE (Self-supervised High-throughput Image denoising Neural network for Electron microscopy) for accelerated, minimally invasive low-dose EM.
• To develop a method that overcomes the limitations of current high-resolution TEM techniques.
• To enable high-throughput structure analysis across diverse material systems.

Main Methods:

• SHINE utilizes a self-supervised, high-throughput image denoising neural network.
• The method employs a single raw image dataset with intrinsic noise for training.
• No expensive ground-truth training datasets are required.

Main Results:

• SHINE effectively reduces noise in low-dose EM images, improving clarity.
• The method overcomes information limits in high-resolution TEM, in situ liquid phase TEM, time-series scanning TEM, and cryo-TEM.
• Demonstrated quantitative improvements in structure analysis across various materials.

Conclusions:

• SHINE facilitates unambiguous, high-throughput structure analysis in low-dose EM.
• The self-supervised approach makes it suitable for limited datasets and eliminates the need for ground-truth data.
• SHINE accelerates minimally invasive EM, advancing materials science and structural biology.