MorphCell disentangles shape and scale for 3D cellular morphology representations

Abstract

Three-dimensional (3D) cell morphology provides a measurable phenotype of cellular state and function, yet learning transferable and interpretable morphological representations across biological systems and imaging modalities remains challenging. Here we introduce MorphCell, a self-supervised framework that learns shape-driven representations from cell-surface point clouds through cross-view reconstruction and spherical self-reconstruction. MorphCell captures both global and local features of 3D cell morphology while retaining physical scale as a separate measurement. Pretrained on non-biological 3D objects, MorphCell transfers without biological task-specific fine-tuning to cellular datasets acquired by confocal microscopy, X-ray microscopy and volume electron microscopy. Its representations outperformed the evaluated baseline representations in red blood cell morphotype and wheat root cell classification. The relative contributions of shape and physical scale differed across biological tasks, with shape alone outperforming shape-scale fusion for red blood cell morphotype classification. In colon epithelial cells, combining shape-driven representations with the nuclear-to-cytoplasmic volume ratio increased balanced accuracy for cancerous versus non-cancerous cell classification from 62.7% to 92.8%. Together, these results establish MorphCell as a general framework for representing, reconstructing and interpreting 3D cellular morphology across imaging modalities and biological contexts.