ProPicker: Promptable Segmentation for Particle Picking in Cryogenic Electron Tomography

Cryogenic electron tomography (cryo-ET) can produce detailed 3D images, called tomograms, of cellular environments. An essential step of cryo-ET data analysis is to find all instances of a particle in a set of tomograms. This particle picking task is a challenging 3D object detection problem due to strong noise and artifacts in the tomograms, as well as the diverse, crowded cellular environment. To enable a fast, flexible, and data-efficient workflow for particle picking, we propose ProPicker, a pretrained, promptable 3D segmentation model. By specifying a prompt, the model is conditioned to detect a specific particle of interest. The prompted model can be used for particle picking as is or can be fine-tuned to a particle-specific picker. Through experiments on simulated and real-world tomograms, we demonstrate that based on a single prompt, ProPicker can accurately detect a range of particles; in some cases even such that are not part of the training set. ProPicker can achieve performance close to or on par with state-of-the-art prompt-based pickers, while being up to an order of magnitude faster. We also show that fine-tuning ProPicker outperforms state-of-the-art particle-specific pickers if limited training data is available.