Developing deep neural networks for automated 3D electron micrograph segmentation
Thursday, September 14, 2017 — Poster Session III
- MD Guay
- AB Anderson
- ZA Emam
- RD Leapman
Modern electron microscopy (EM) is capable of producing large, high-resolution 3D images of biological structures, but the labor required to manually segment EM images into their semantic components hinders further data analysis. Currently, segmentation software incorporating deep neural networks offers state-of-the-art performance for automated EM segmentation. However, even state-of-the-art automated segmentation tools require extensive manual correction for many data sets of interest to the EM and systems biology communities, and are therefore impractical for image analysis. Our lab is designing novel neural networks and incorporating them into a segmentation software pipeline to improve automated segmentation performance for EM data sets taken from multiple biological systems. Our poster details our advances in neural network architecture design for biomedical image segmentation, presents an overview of the software framework we have developed for building segmentation neural networks, and displays the results of our work as applied to several data sets available to our lab.
Category: Biomedical Engineering and Biophysics