Automatic particle picking and multi-class classification in cryo-electron tomograms

Macromolecular structure determination using cryo-electron tomography requires large amount of subtomograms depicting the same molecule, which are averaged. In this paper, we propose a novel automatic particle picking and classification method for cryo-electron tomograms. The workflow comprises two stages: detection and classification. The detection method consists of a template-free picking procedure based on anisotropic diffusion filtering and connected component analysis. For classification, a novel 3D rotation invariant feature descriptor named Sphere Ring Haar and a hierarchical classification algorithm consisting of two machine learning models (DBSCAN and random forest) are proposed. The performance of our method is superior compared to template matching based methods and we achieved over 90% true positive rates for detection of proteasomes and ribosomes in experimental data.