Accurate segmentation of plant epidermal cell contours in SEM micrographs using fine-tuned Cellpose models

Abstract Background Quantitative analysis of plant epidermal cells requires accurate segmentation of cell boundaries, a task particularly challenging for jigsaw-puzzle-shaped cells captured via scanning electron microscopy (SEM). Most existing Arabidopsis thaliana epidermal cell segmentation pipelines focus on light microscopy images, and current SEM segmentation workflows are largely built upon backscattered electron (BSE) micrographs. Building on these efforts, we developed Cellpose cyto3 models derived via fine-tuning for epidermal cell segmentation across plant species and imaging modalities. Results Here, we established a secondary electron (SE) micrograph dataset using a Coolstage system and fine-tuned the pre-trained Cellpose model cyto3 to construct specialized segmentation models. Benchmark evaluations demonstrated that our derived model achieved segmentation accuracy comparable to manual annotations, significantly outperforming the baseline cyto3 model. For Arabidopsis and Chinese yam, gold sputtering provided SE images with boundary contrast and segmentation accuracy that matched or surpassed those from gold-sputtered BSE imaging. The model demonstrated strong robustness to topography and image contrast variations. Cross-modality and cross-species validations confirmed that our derived model delivered reliable segmentation of epidermal cells morphologically similar to Arabidopsis across diverse plant species and SEM imaging modalities. Representative segmentation examples show promising visual performance from bright field, differential interference contrast and fluorescence imaging of Arabidopsis , as well as from cucumber and Chinese yam (abaxial epidermis), tomato and tobacco via selected modalities. Additionally, fine-tuning cyto3 on small, domain-specific target datasets enabled accurate identification of stomatal complexes, stomatal pores, and diverse epidermal cell morphologies in monocots including maize. Conclusions This pipeline reduces the requirement for high contrast between cell boundaries and intracellular regions, enabling segmentation from images with relatively weak boundary contrast. Our derived cyto3 model, trained on SE-SEM topographic images, the most commonly acquired signal in SEM imaging, enables epidermal segmentation of Arabidopsis and generalizes to morphologically similar pavement cells. This workflow delivers a robust framework for automated leaf epidermal cell phenotyping across diverse plant species and multiple imaging modalities.