Automated Particle Picking in Cryo-Electron Micrographs using Deep Regression

Selection of good particles in cryo-electron micrographs is an important step in the reconstruction of high resolution 3D structures. In this study, we constructed a deep learning-based method to automatically detect particle centers from micrographs. This is a challenging task because of the low signal-to-noise ratio of cryo-EM micrographs, and the size, shape, and grayscale-level differences of particles. We proposed a Fully Convolutional Regression Network (FCRN) that maps the particle image to a continuous distance map that acts like a probability density function of particle centers. This approach is simple, but very effective in recognizing different grayscale patterns corresponding to 2D views of 3D particles. Our experimental results on dataset β-galactosidase (EMPIAR-10017) [1] showed that FCRN outperfomed Faster-RCNN, Apple picker, and RELION's particle picker. Compared to the ground truth of this dataset, FCRN achieved better picking performance, and 3D structure of those picked particles also had higher resolution.