A deep framework for bacterial image segmentation and classification

Abstract

Bacterial image segmentation and classification is an important problem because bacterial appearance can vary dramatically based on environmental conditions. Further, newly isolated species may exhibit phenotypes previously unseen. Conventionally, biologists identify bacteria using colony morphology, biochemical properties, or molecular phylogenetic approaches. However, these phylogenetic classification approaches do not provide predictive information about the colony morphology that is expected to result from the growth of these bacteria on agar, or how these morphological phenotypes might vary in the presence of other bacterial species. In this paper, we propose a framework to automatically identify and classify regions of bacterial colony images and correspond them across different images from different contexts. Importantly, this approach does not require prior knowledge of species' appearances. Rather, our method assumes that images contain one or more bacteria from a pool of bacteria, and learns morphological features relevant to distinguishing between bacteria in this pool. Our method first segments the image into regions covering the bacterial colonies, agar, plate, and various border artefacts. To achieve this, we use an unsupervised deep learning technique, Convolutional Deep Belief Network (CDBN). This technique provides a deep representation of small image patches. Using this high-level representation instead of raw pixel intensities, we train a support vector machine (SVM). The trained SVM accurately classifies foreground and background patches. Once the foreground patches are identified, we train a supervised deep learning method, Convolutional Neural Network (CNN), that predicts which bacterial colonies from the pool occur in a query image. Experimental results demonstrate that our method outperforms other classical methods on segmentation and classification.