Deep Learning-Based Segmentation of Cryo-Electron Tomograms

Cryo-electron tomography (cryo-ET) allows researchers to image cells in their native, hydrated state at the highest resolution currently possible. The technique has several limitations, however, that make analyzing data it generates time-intensive and difficult. Hand segmenting a single tomogram can take from hours days, but microscope easily generate 50 or more tomograms day. Current deep learning segmentation programs for cryo-ET do exist, are limited one structure time. Here, multi-slice U-Net convolutional neural networks trained applied automatically segment multiple structures simultaneously within cryo-tomograms. With proper preprocessing, these be robustly inferred many without need training individual each tomogram. This workflow dramatically improves speed with which cryo-electron analyzed by cutting time down under 30 min most cases. Further, segmentations used improve accuracy of filament tracing cellular context rapidly extract coordinates subtomogram averaging.