Contour detection and segmentation method applicable to electron tomography images with auto-classification by machine learning

Automatization of object recognition from electron microscope images is important theme. Recently, this theme is essential not only to reduce complicatedness of various morphometry by manual means but also to segment objects in electron tomography and serial slice SEM method. For this theme, using auto-classification by machine learning, we are developing an automatic identification method of structural objects that a researcher wants to observe [1] and also an extraction method of object regions to input a computer as supervised learning images. In this time, we focus on the fact that various kinds of structural objects in biological electron microscope images have unique features on their contours or around the contours. As a result of the study, we propose an auto-classification method by machine learning using these characteristics, which automatically segments structural objects. The proposed method mainly consists of two steps. First, the collection of local contour images and local images around the contours as mentioned above is performed for the supervised leaning. This collection work is assisted by a developed semi-automatic method of tracking contours by using a unique ‘Gabor Wavelet’ analysis. Second, our machine leaning software automatically detects various contours of structural objects exposed in tomographic cross-section images and segments them into individual objects by the classifier based on the learned data obtained in the first step. In this presentation, results of 2D segmentation are reported. Further, the proposed method is planned to expand toward 3D segmentation. © 2016, Oxford University Press. All rights reserved.