Deep-learning deconvolution and segmentation of fluorescent membranes for high-precision bacterial cell-size profiling

Abstract

Evolutionary studies in bacteria have emphasized genetic and metabolic diversity, while cell-size variation has received less attention. Here we introduce MEDUSSA, a high-throughput method for precise bacterial cell-size profiling based on automatic segmentation of fluorescent membrane images, well suited to studying cell-size diversity. The approach uses deep-learning-based membrane deconvolution, segmentation models fine-tuned for fluorescent-membrane images, and error-corrected cell measurement to extract accurate sizes from individual bacterial cells regardless of shape, size, chaining, or clustering. Our method overcomes limitations of phase-contrast segmentation, yielding reliable single-cell dimensions. We applied MEDUSSA to six strains of Priestia megaterium and found over twofold differences in cell volume across strains, largely driven by differences in cell width. We identified a partially-functional PBP1 allele that underlies the reduced width of one strain. Together, these results demonstrate the power of comparative analyses in bacterial cell biology and expand the toolkit to investigate the evolution of bacterial cell size.