Artificial-intelligence-driven segmentation and analysis of microbial cells

Understanding microbial growth and morphology at the single-cell level is essential for studying microbial physiology, providing valuable insights for research and biotechnology. However, existing workflows often rely on manual cell annotation, which limits efficiency and scalability. Therefore, developing an automated workflow for quantitative analysis of cell growth and morphology is highly desirable.

We aim to develop an AI-driven analysis system that efficiently segments and indexes microbial cells, and quantitatively analyzes individual cellular features without requiring expensive annotations, for automated monitoring of cell counts and morphological characteristics.

The automated system consists of four modular components: denoising, zero-shot segmentation using the Segment Anything Model (SAM), structured post-processing, and a quantitative feature extraction. To evaluate the effectiveness of each component, we conducted ablation experiments and systematically studied their impact on the overall system performance.

Denoising and post-processing improved segmentation accuracy by 12.10% and 2.30%, respectively. Among the evaluated SAM variants, the SAM-H model achieved the best performance, with an average error rate of only 3.0% across 1162 manually annotated Escherichia coli cells. Using the optimized SAM-H pipeline, the system efficiently extracted morphometric and intensity features from Escherichia coli cells and nuclei of the yeast and cancer cell lines.

This framework automates quantitative analysis of microbial cells in high-resolution microscopy images. It will enable advanced research on microbial adaptations, with the potential to accelerate studies of extremophiles under harsh environments.