Application of Gaussian Mixture Models with Expectation Maximization in Bacterial Colonies Image Segmentation for Automated Counting and Identification

Abstract

This paper presents an approach to identification of homogenous bacterial colonies grown on agar in Petri dishes. The aim of the work was to recognize regions in digital images of Petri dishes were homogenous bacterial colonies were developed as well as to estimate their size. Isolation of bacterial cultures’ region from the dish was achieved by image segmentation based on histogram analysis. The histogram was parameterized using Gaussian Mixture Model with Expectation Maximization. This algorithm gave a good estimation of the actual gray level distribution and was able to separate merging distribution of two different objects. However, it performed poorly with the presence of outliers. The algorithm performance was also dependent on initial model and the number of Gaussians chosen. Overall, for images taken under controlled conditions the application of Gaussian Mixture model with Expectation Maximization proved to be successful and efficient approach to image segmentation of bacterial colonies. Final step was separation of circular-like colonies from non-circular ones. This was achieved using appropriate shape identification techniques. Validation of the proposed segmentation process was demonstrated using images from the microbiology laboratory, some artificially generated images and images downloaded from the Internet.KeywordsImage segmentationObject recognitionColony countingGaussian Mixture ModelExpectation-maximization algorithm