Review: automatic particle detection in electron microscopy

Abstract

Advances in cryoEM and single-particle reconstruction have led to results at increasingly high resolutions. However, to sustain continuing improvements in resolution it will be necessary to increase the number of particles included in performing the reconstructions. Manual selection of particles, even when assisted by computer preselection, is a bottleneck that will become significant as single-particle reconstructions are scaled up to achieve near-atomic resolutions. This review describes various approaches that have been developed to address the problem of automatic particle selection. The principal conclusions that have been drawn from the results so far are: (1) cross-correlation with a reference image ("matched filtering") is an effective way to identify candidate particles, but it is inherently unable to avoid also selecting false particles; (2) false positives can be eliminated efficiently on the basis of estimates of particle size, density, and texture; (3) successful application of edge detection (or contouring) to particle identification may require improvements over currently available methods; and (4) neural network techniques, while computationally expensive, must also be investigated as a technology for eliminating false particles.