Deep Convolutional Encoder-Decoder Architecture for Neuronal Structure Segmentation

Abstract

Electro microscopic connectomics is a practical application of research direction. It determines whether a nerve is damaged by judging the connectivity of nerves. However, the formidable size of Electro Microscopic (EM) image data generated by serial-section Transmitted Electron Microscopy (ssTEM) severely depends on human annotation, which is impractical. One of the main challenges in connectomics research is to take minimal user intervention into account during neuronal structures automatic segmentation. To address this problem, a network is constructed to segment neuronal structures automatically, which expands receptive field of feature maps. Besides, we also introduce data augmentation method to use the available training data more efficiently. Our model is proposed based on a context network, and its architecture consists of an encoding path that enables feature extraction. The novel introduction of summation-based skip connection is aimed to connect decoding path with encoding path. Finally, real experiments with ISBI EM dataset validate the approach.