Segmentation in Stacks of Electron Microscopy Images Using Deep Learning

Abstract

In the research field of connectomics, serial section electron microscopy images are reconstructed in three dimensions for the purpose of observing neural structure. Automatic segmentation of neuronal cells is important to reconstruct the images, and in recent years a method using deep learning has got a lot of attention. However, in the application of deep learning, there are many optional methods to improve segmentation accuracy. In this paper, we focus on the tuning method of hyperparameters. Deep Contextual Network (DCN), which is expected to be processed with particularly high precision among existing methods, has some feature maps with different spatial frequencies. We have an eye to this structure and propose hyperparameter tuning method based on visualization of feature maps. Experiments using two kinds of datasets showed that DCN accuracy can be improved by changing the kernel size of the deconvolution layer by the proposed method.