Showing papers by "Misuzu Baba published in 2017"

Segmentation and Contour Extraction in Biological Transmission Electron Microscope Images with ‘Bag-of-Features’ Method in Machine Learning

Abstract: When observing minute objects and structures with a transmission electron microscope (TEM), it is necessary to search a very large specimen area for them and keep them inside the field of view. Because of the relatively small field of view of a TEM this manual work is very time consuming. Several research works to overcome the problem have been reported. For example, an automatic method picking up of particles in cryo-electron microscopy [1], and unique one that finds out localized fine crystalline samples such as graphitized carbon [2]. However, at present, researches of the general-purpose technique of automatically and mechanically searching for any biological object appearing in thin section images or in tomographic images with electron CT is seemed to be few.

A Particle Extraction Method with an Improved Auto-encoder in Neural Networks with the Aid of HOG Feature Analysis

Abstract: In the field of electron microscopic image analysis, machine learning such as neural networks [1] is remarkable method, which has abilities of automatic extraction of structural objects and automatic classification. Recently, deep leaning in the neural networks is actively applied because of high recognition rate [e.g. 2]. This learning method is able to choose effective feature patterns from a set of learned local images. We are developing a deep learning software for application to biological transmission electron microscope images (including electron tomographic images), especially yeast cell (Saccharomyces cerevisiae) images, in which automatic particle extraction is a purpose. We already applied a preliminary software to ultra-thin section images of the yeast cell for picking out some kinds of structural particles (e.g. Virus-like particle (VLP) in autophagy). The characteristics of the software is that it is made of auto-encoder in neural networks. The image analysis with the auto-encoder is essentially equivalent to that with the principal component analysis (PCA) [3]. The auto-encoder analyses a lot of particle images which are manually collected and a certain number of important image components are automatically extracted from the particle images like PCA. These components are visualized and the number of components is adjustable. We are improving the function of the autoencoder so as to raise the efficiency.