Quantitative Analysis System for Bacterial Cells in SEM Image using Deep Learning

Abstract

In this paper we propose a system to analyze bacteria from a given Scanning Electron Microscope (SEM) image of the bacterial sample. Thousands of bacteria lives in the human gut and recent studies have shown that the quantitative features of the microbiome, such as co-existence ratio of different bacteria, can be indicative of the health condition in humans. Conventional bacteria analysis methods using microscopic images, can only be used to examine a single bacteria colony. In contrast, we propose a novel system to morphologically analyze the bacteria from SEM images. By this, we expect to enable a rapid analysis of the human gut bacteria ratio, in which various type of bacteria are mixed. However, to achieve an automatic and accurate count of the bacteria in the SEM images, it is important to accurately identify the bacteria regions, separate the connected bacteria regions and classify them. To address this, we propose a system that includes a segmentation, separation and a classification module. Our system achieves more than 90% recall for all of original three datasets that we have created. Subsequently, we show the comparison results between another state-of-the-art segmentation method and our system, and we empirically report that our system has a better performance.