BACTERIA TRACKING AND DIVISION DETECTION USING GRAPH NEURAL NETWORKS

Tracking of bacteria is important to investigate dynamic biological processes at the single cell level. While deep learning methods are commonly used for tracking objects in natural images, such methods are rare for bacteria. We propose a novel graph neural network for bacteria tracking and division detection. Object correspondences and divisions are represented by the edges of a graph, and both tasks of tracking and division detection are performed simultaneously by graph edge classification. The method was evaluated using live-cell bright-field microscopy image sequences of spore germination and outgrowth of rod-shaped Bacillus subtilis. It turned out that our method outperforms previous methods. Particularly, we found that direct division detection by graph edge classification outperforms overlap-based division detection.