Deep Learning Based Bacteria Classification from SEM Images Using a Combination of Membrane and Internal Features

Scanning electron microscopes (SEM) take very high-magnification images that allow us to examine the mor-phological features of bacteria samples in great detail. However, there are two major problems associated with this task. First, typical SEM images of bacteria show many different bacteria touching each other. For accurate classification and quantitative analysis, the touching bacteria need to be distinguished correctly into individual bacteria regions. Second, in these images, different types of bacteria might share similar visual features and have only locally distinguishable features. Here, we propose a system to perform multi-class classification of bacteria from SEM images. The system incorporates distance map inference and feature classification of the membrane and internal bacterial region, in order to segment the bacteria regions accurately, distinguish bacteria in contact with one another and identify local bacterial features. Extensive experimentation on an original bacteria dataset that we prepared shows that our system outperforms other object detection and segmentation methods on this problem.