TomoSegNet: Augmented membrane segmentation for cryo-electron tomography by simulating the cellular context

Abstract Membrane segmentation is an essential task in the workflow for processing cryo-electron tomography data. Recently, machine learning algorithms have successfully been adopted to perform membrane segmentation. However, the performance of these approaches is limited by the training dataset, as models are trained from manual annotations, thereby hindering the models’ generalization and preventing the recovery of membranes that have vanished due to distortions. Here, we address these limitations by generating training data with a simulator. To provide a representative and realistic dataset, we have extended the current state-of-the-art in simulators for cryo-electron tomography by incorporating a biophysical model for membranes. We demonstrate that our machine learning model, trained solely from synthetic data, and thanks to the physical knowledge learned from the simulator, outperforms the current state-of-the-art for membrane segmentation in a diverse set of experimental data. This performance is particularly noteworthy in terms of recovering membranes lost due to imaging distortions.