Deep Learning Approach for High-accuracy Electron Counting of Monolithic Active Pixel Sensor-type Direct Electron...

Area of Science:

• Materials Science
• Computer Science
• Image Processing

Background:

• Electron counting in direct electron detectors aims to reduce readout and Landau noise.
• Existing algorithms struggle with multielectron events, positional accuracy, and dark noise.

Purpose of the Study:

• To develop a supervised deep learning (DL) approach for accurate single electron event recognition.
• To improve electron counting in monolithic active pixel sensor direct electron detectors.

Main Methods:

• Utilized a Faster region-based convolutional neural network (R-CNN) for supervised learning.
• Trained the DL model to recognize single electron events under varying electron doses and voltages.
• Evaluated performance using modulation transfer function (MTF) and detector quantum efficiency.

Main Results:

• The DL approach achieved high accuracy, with near-ideal MTF and detector quantum efficiency.
• Demonstrated superior positional accuracy (0.47 pixel deviation) compared to conventional methods (0.59 pixel).
• Showed enhanced robustness against coincidence loss at higher electron doses, maintaining MTF above 0.83.

Conclusions:

• The supervised DL Faster R-CNN model accurately identifies single electron events.
• This DL method offers improved positional accuracy and robustness at higher dose rates.
• The DL approach extends the benefits of electron counting to a wider range of experimental conditions.