Synthesizing Unpaired Images of Bacterial Biofilms with Auxiliary Diffusion Models

Bacterial biofilms are an essential part of the ecosystem. These biofilms reveal information about bacterial resistance, infection, and biological energy sources. Data-driven computation approaches require large volumes of annotated training data of the densely packed bacterial biofilms. However, the cells in these biofilms often require time-consuming and expensive superresolution imaging techniques as their dimensions fall well below the diffraction limit. To this end, we propose an end-to-end framework for generating realistic biofilm images from unpaired real biofilms. Unlike recent image-to-image translation methods, the proposed model involves a single generative network. The framework comprises two domain translation networks which are trained with an auxiliary diffusion model with a multiplicative noise model. We also use a loss term for histogram equalization between the real and generated images which further boosts the model performance. The experiments validate the efficiency of the proposed model for generating realistic biofilm images for different densities of bacterial cells.