Improved deep learning based macromolecules structure classification from electron cryo tomograms

摘要: Cellular processes are governed by macromolecular complexes inside the cell. Study of native structures has been extremely difficult due to lack data. With recent breakthroughs in electron cryo tomography (CECT) 3D imaging technology, it is now possible for researchers gain accesses fully study and understand single cells. However, systematic recovery from CECT very high degree structural complexity practical limitations. Specifically, we proposed a deep learning based image classification approach large-scale structure separation our previous work was only initial step towards exploration full potential macromolecule separation. In this paper, focus on improving performance proposing three newly designed individual CNN models: an extended version (Deep Small Receptive Field) DSRF3D, donated as DSRF3D-v2, residual block neural network, named RB3D convolutional 3D(C3D) model, CB3D. We compare them with previously developed model (DSRF3D) 12 datasets different SNRs tilt angle ranges. The experiments show that new models achieved significantly higher accuracies. accuracies not than 0.9 normal datasets, but also demonstrate potentials operate levels noises missing wedge effects presented.