Transfer Learning for Efficient Segmentation of Subcellular Structures in 3-D Electron Microscopy

Abstract

Automating the segmentation of biological structures at the cell and organelle level is essential for making efficient use of 3-D scanning electron microscopy (SEM) technologies, such as serial block-face (SBF)-SEM and focused ion beam (FIB)-SEM. Currently, automated segmentation using convolutional neural networks (CNNs) typically addresses a specific biological system, and it is necessary to train a CNN for each new problem. However, addressing each problem individually imposes impractical labor requirements for machine learning developers, and requires large amounts of annotated training data, which are costly to produce. Here, we present our work on reducing this burden by applying transfer learning techniques to neural network architectures to solve instance segmentation and semantic segmentation problems for multiple biological datasets imaged via 3-D SEM. We are able to train instance segmentation networks to detect and localize objects across different cells and tissues. By sharing semantic segmentation network pathways between biological systems, we are able to reduce the amount of training data required for effective segmentation of new EM datasets.