Evaluating the Quality of Serial EM Sections with Deep Learning.

Area of Science:

• Microscopy
• Image Analysis
• Machine Learning

Background:

• Automated image acquisition enhances serial section scanning electron microscopy (ssSEM) throughput.
• Image quality in ssSEM can fluctuate due to autofocusing and beam stigmation.
• Automated quality evaluation is crucial for generating high-quality ssSEM datasets.

Purpose of the Study:

• To develop and validate an automated method for assessing ssSEM image quality.
• To determine if convolutional neural networks can replicate human quality evaluations.
• To enable real-time identification of imaging issues during ssSEM acquisition.

Main Methods:

• Tested multiple convolutional neural networks for ssSEM image quality evaluation.
• Developed a modified ResNet-50, named Quality Evaluation Network (QEN).
• Trained and validated QEN against user-generated quality scores.

Main Results:

• QEN reliably predicted user-generated quality scores for ssSEM images.
• The Quality Evaluation Network demonstrated high accuracy in assessing image quality.
• QEN can be run in parallel with ssSEM acquisition for immediate feedback.

Conclusions:

• A modified ResNet-50 (QEN) provides reliable automated quality assessment for ssSEM images.
• QEN facilitates rapid identification of imaging problems, enabling image retakes.
• Publicly shared code and dataset support the use and further development of QEN.