Fast Improvement of TEM Images with Low-Dose Electrons by Deep Learning

Area of Science:

• Materials Science
• Microscopy
• Artificial Intelligence

Background:

• Low electron dose observation is crucial for transmission electron microscopy (TEM) to study sensitive samples.
• Traditional image processing methods improve TEM image quality but are often insufficient for dynamic, in situ observations.
• Electron beam-induced damage limits the observation of beam-sensitive specimens.

Purpose of the Study:

• To develop an advanced image processing technique for TEM using convolutional neural networks (CNNs).
• To enable high-quality in situ observations of electron-beam-sensitive specimens at low electron doses.
• To achieve real-time imaging capabilities for dynamic processes.

Main Methods:

• A convolutional neural network (CNN) was applied to transmission electron microscopy (TEM) imaging.
• An end-to-end training pipeline was developed using a dataset of short-exposure and long-exposure TEM images.
• The processed short-exposure images were evaluated for quality and conversion time.

Main Results:

• Images acquired with a low electron dose (approx. 5 e-/pixel) achieved quality comparable to high-dose images (approx. 1,000 e-/pixel).
• The image processing pipeline demonstrated a fast conversion time of approximately 8 ms.
• This enables in situ observation at a high frame rate of 125 frames per second (fps).

Conclusions:

• CNN-based image processing significantly enhances low-dose TEM imaging.
• The developed technique allows for high-speed, in situ observation of electron-beam-sensitive materials.
• This advancement opens new possibilities for studying dynamic phenomena in delicate specimens.