ProPicker: Promptable Segmentation for Particle Picking in CryogenicElectron Tomography

Abstract

Cryogenic electron tomography (cryo-ET) produces detailed 3D images (tomograms) of cellular environments. A key step in cryo-ET data analysis is detecting all instances of a specific particle across tomograms (particle picking). This is a challenging object detection task due to strong noise, artefacts, and the crowded cellular context. Here, we propose ProPicker, a pretrained, promptable 3D segmentation model that enables a flexible and data-efficient particle picking workflow. By specifying a prompt, ProPicker is conditioned to detect a particle of interest and can then be used directly or can be fine-tuned to a particle-specific picker for improved accuracy. Experiments on simulated and real-world tomograms show that, using a single prompt, ProPicker achieves performance close to or on par with state-of-the-art methods while being up to an order of magnitude faster. Moreover, ProPicker can detect particles not seen during training. Fine-tuning ProPicker outperforms state-of-the-art particle-specific pickers if limited training data is available.