BPS2026 – Synthetic data-driven deep learning for robust membrane segmentation in cryo-electron tomography

The quantitative analysis of biological membranes from cryo-electron tomography (cryo-ET) data is hindered by imaging artifacts, noise, and the complexity of membrane morphologies. The missing wedge remains a critical barrier because it erases regions of membranes oriented outside of the tilt range, which prevents complete annotation. Current deep learning models can achieve strong results but rely on incomplete and error-prone manual annotations, leading to variable generalizability across biological systems and little recovery of the data lost from the missing wedge. Here, we introduce a nnUNet model trained exclusively on a novel synthetic data set generated by an extended version of the PolNet simulator that can generate considerably more realistic membranes. Importantly, this nnUNet trained only on synthetic data achieves competitive or superior accuracy compared to models trained on manually annotated real data. Moreover, our model demonstrates enhanced robustness across diverse biological contexts and successfully recovers membranes degraded by the missing wedge artifact. This work establishes a framework for the systematic design of synthetic training data using PolNet, enabling more accurate, generalizable, and artifact-aware deep learning models for segmentation of cryo-ET data.