A morphology-driven workflow to decipher 3D electron microscopy segmentation in diatoms

Cells display diverse shapes, sizes, and internal structures, particularly among microbial species. These differences reflect how cells adapt to their environments and perform essential functions. Modern three-dimensional (3D) electron microscopy captures this diversity at the nanometer scale, but analysis is slow because structures must be outlined manually by experts. Artificial intelligence (AI) has improved image analysis in medical and cell biology research by learning from repeated observations. However, studies of microbial and microalgal biodiversity often rely on single snapshots, complicating automated analysis because AI must learn from static morphology. In this study, we developed an AI-assisted segmentation framework for whole-cell 3D electron microscopy data. By comparing neural network architectures and using transfer learning and contrast-aware strategies, we showed that accurate segmentation is possible with limited training data and standard computing resources. Our method enables faster, reproducible analysis of cellular ultrastructure, supporting large-scale investigations into cell morphology and diversity.