Designing deep neural networks to automate segmentation for serial block-face electron microscopy

• Published in: 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018) pp. 405 - 408
• Main Authors: Guay, Matthew, Emam, Zeyad, Anderson, Adam, Leapman, Richard
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.04.2018
• Subjects: automated segmentation; Biology; Convolutional codes; Deep learning; electron microscopy; Image segmentation; Neural networks; Scanning electron microscopy; serial block-face imaging; Training data
• DOI: 10.1109/ISBI.2018.8363603

Today, serial block-face scanning electron microscopy (SBF-SEM) is capable of producing teravoxel-scale 3D images of biological structures at nanometer-scale resolutions. Image segmentation is fundamental to data analysis workflows in biological electron microscopy (EM), but SBF-SEM datasets can greatly exceed the manual segmentation capacity of a laboratory. Fast automated segmentation algorithms would alleviate this problem, but practical solutions remain unavailable for many biological problems of interest. Segmentation algorithms using deep neural networks have recently demonstrated significant performance gains, but designing high-performing networks that effectively solve targeted problems remains challenging. We are developing genenet, a Python package to rapidly discover, train, and deploy high-performing neural network architectures for SBF-SEM segmentation with little user intervention. Here, we demonstrate how to use genenet to train an ensemble of segmentation networks for a human platelet tissue sample. Initial results indicate this approach is viable for accelerating the segmentation process.