Segmentation in large-scale cellular electron microscopy with deep learning: A literature survey

Area of Science:

• Microscopy and Imaging Science
• Computational Biology
• Artificial Intelligence in Biology

Background:

• Manual segmentation of large-scale electron microscopy (EM) datasets is labor-intensive and time-consuming.
• Automated segmentation methods are essential for efficient analysis of high-resolution cellular and sub-cellular structures.
• Deep learning has emerged as a powerful tool for addressing these challenges in recent years.

Purpose of the Study:

• To review the progress of deep learning-based segmentation techniques in large-scale cellular EM over the past six years.
• To detail key datasets, methodologies (supervised, unsupervised, self-supervised), and challenges in EM image segmentation.
• To provide an outlook on future trends, including large-scale models and learning from unlabeled data.

Main Methods:

• Comprehensive review of deep learning algorithms applied to 2D and 3D EM image segmentation.
• Analysis of supervised, unsupervised, and self-supervised learning approaches.
• Examination of network architectures designed to handle EM image complexities like heterogeneity and spatial complexity.

Main Results:

• Significant advancements in both semantic and instance segmentation of cellular and sub-cellular structures using deep learning.
• Identification of key datasets that have driven progress in EM segmentation.
• Description of methods that address challenges such as image heterogeneity and spatial complexity.

Conclusions:

• Deep learning has revolutionized automated segmentation in large-scale EM, enabling more efficient and accurate analysis.
• Future directions include leveraging large-scale models and unlabeled data for more generalized feature learning across diverse EM datasets.
• Continued development in deep learning promises further breakthroughs in understanding cellular and sub-cellular organization.