Fig. 5. Deep learning-driven resolution adjustment enables phenotype distinction while increasing imaging throughput.

S. aureus bacteria were treated with FtsZ inhibitor PC190723, which prevents cell division and results in larger bacteria phenotypes²³. S. aureus (wild type (WT) and treated) were then imaged using Structured Illumination Microscopy (SIM) and segmented with StarDist. a Representative images of S. aureus with ground truth (GT) segmentation and StarDist results at the original resolution (WT) and after 2³-factor downsampling (both WT and treated conditions). Scale bars: 100, 25, and 25 pixels, respectively. b (Boxplots) Distribution of Intersection over Union (IoU) values of StarDist models trained and tested only with WT bacteria images across multiple resolution factors (upsampling: 2⁻² to 2⁰; downsampling: 2¹ to 2⁴), showing the relationship between segmentation accuracy and the percentage of bacterial diameter covered by a single pixel. (Purple line in logarithmic scale) Theoretical throughput was calculated as the number of objects observable by a unit of time (τ) given a certain resolution. τ is specific to each microscopy acquisition modality and preserved across the different resolutions. c Diameter distributions across resolution scales estimated from GT and the StarDist results for WT bacteria images in (b). d Comparison of the diameter distributions across various resolution scales for WT and treated bacteria. StarDist was trained with a mixed dataset containing WT and PC190723-treated S. aureus images in this case.