Sample-Based Training Data for Effective 3D Cell Segmentation

Deep learning remains the leading choice for cell segmentation, a critical step in bioimaging analysis. Whilst deep learning models provide excellent segmentation accuracy, a large number of hand-annotated samples are required which are scarce in 3D. We propose a method for generating synthetic data for training deep learning models to augment or replace such datasets. By preserving key features in real data, we show that a standard UNet trained on synthetic data can segment single motile cells with branching filopodia with high accuracy. We also demonstrate how low-effort slice annotations can sufficiently replace volume annotations in our data generation pipeline. Overall, we provide a simple alternative to annotating large 3D datasets for training neural networks to segment cell imaging data.