Deep learning training data (JOVE)

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