Practical method of cell segmentation in electron microscope image stack using deep convolutional neural network☆.

Area of Science:

• Neuroscience
• Computational Biology
• Microscopy

Background:

• Three-dimensional (3D) electron microscopy (EM) image segmentation is crucial for understanding neural structures but is labor-intensive.
• Deep convolutional neural networks (CNNs) offer automated segmentation but typically require extensive training datasets, posing a significant challenge for sparse data.

Discussion:

• This research addresses the impediment of large training datasets for CNN-based segmentation in 3D EM.
• The study successfully employed a CNN with a minimal training dataset for segmenting a Cerebellar Purkinje cell from mouse cerebellum EM data.

Key Insights:

• Sparse segmentation of 3D EM images can be achieved efficiently using CNNs with reduced training data.
• The methodology significantly decreased the time required for segmentation, completing the task in under two working days compared to conventional methods.
• Reducing the training dataset size is a viable strategy to minimize overall labor time in sparse segmentation.

Outlook:

• This approach holds promise for accelerating neuroanatomy research and large-scale neural circuit mapping.
• Further validation on diverse neural structures and EM datasets is warranted.
• Exploration of optimized CNN architectures for minimal data segmentation could enhance efficiency.