Realistic-Shape Bacterial Biofilm Simulator for Deep Learning-Based 3D Single-Cell Segmentation

Bacterial biofilms are complex biological systems that play significant roles in areas such as infectious diseases and fuel cells. While biofilms are important, our understanding of biofilm communities is limited due to barriers in live cell imaging and analysis. Development of effective automatic image analysis techniques for bacterial biofilms can give us new insights into many unknown characteristics of the individual bacteria cells such as structure and cell-to-cell interaction. Single-cell segmentation is the initial step for automatic analyses of the individual cells within a biofilm. While traditional model-based segmentation approaches often fail to accurately segment individual cells within a biofilm due to low image resolution, data-driven deep learning techniques can offer superior solutions by learning the segmentation task from an effective training dataset. Here we propose a framework to simulate 3D synthetic biofilms comprised of realistic-shaped bacteria. Such a synthetic dataset is used to train a deep segmentation network for single-cell segmentation of real biofilm images. We demonstrate that training the network with the synthetic biofilms generated by the proposed framework achieves significantly higher single-cell segmentation accuracy on real biofilm data in comparison to training the network with existing rod-shaped synthetic biofilms (with improvements in the range of 20% to 40% in F1 scores for single-cell segmentation).