Object Detection Frameworks for Fully Automated Particle Picking in Cryo-EM

Particle picking in cryo-EM is a form of object detection for noisy, low contrast, and out-of-focus microscopy images, taken of different (unknown) structures. This thesis presents a fully automated approach which, for the first time, explicitly considers training on multiple structures, while simultaneously learning both specialized models for each structure used for training and a generic model that can be applied to unseen structures. The presented architecture is fully convolutional and divided into two parts: (i) a portion which shares its weights across all structures and (ii) N+1 parallel sets of sub-architectures, N of which are specialized to the structures used for training and a generic model whose weights are tied to the layers for the specialized models. Experiments reveal improvements in multiple use cases over the-state-of-art and present additional possibilities to practitioners.