MorphoLearn: A Morphology-driven Workflow to Decipher 3D Electron Microscopy segmentation in Diatoms

Three-dimensional electron microscopy (3D EM) enables the quantitative analysis of cellular ultrastructure. However, large-scale segmentation of whole-cell volumes poses a significant challenge, especially in biologically diverse systems. Unlike medical and animal cell imaging, which often benefit from temporal redundancy and relatively uniform morphology, studies of microbial and microalgal biodiversity must rely on static snapshots. These snapshots exhibit high variability in cell shape, organelle organisation, and image contrast. Consequently, robust AI-assisted segmentation in this context requires models that learn directly from morphological features and can adapt to heterogeneous sample preparation. In this paper, we present a systematic framework for AI-assisted segmentation of Focused Ion Beam-Scanning Electron Microscopy (FIB-SEM) datasets. This framework is specifically designed to address the challenges posed by morphological diversity and contrast variability while remaining within realistic computational constraints. We evaluate multiple lightweight 3D encoder-decoder architectures and identify VNet as the best option for balancing computational efficiency and volumetric accuracy in whole-cell segmentation. Using datasets from two strains of Phaeodactylum tricornutum and extending our analysis to cross-species comparisons, we demonstrate that training on specific regions of interest can lead to an overestimation of model performance. In contrast, performing whole-cell segmentation uncovers significant differences in architectural robustness. Moreover, we show that transfer learning and contrast-aware hybrid strategies allow for efficient adaptation to previously unseen datasets with minimal annotation. The incorporation of boundary-aware loss functions significantly enhances the delineation of closely associated organelles, such as chloroplasts and mitochondria, in multi-class segmentation tasks. Together, these findings establish a scalable, reproducible, and biologically informed AI framework for 3D FIB-SEM segmentation. This framework enables high-throughput analysis of cellular ultrastructure across diverse species and imaging conditions.