Icecream: High-Fidelity Equivariant Cryo-Electron Tomography

Abstract

Cryo-electron tomography (cryo-ET) visualizes 3D cellular architecture in near-native states. Recent deep-learning methods (CryoCARE, IsoNet, DeepDeWedge, CryoLithe) improve denoising and artifact correction, but performance remains limited by very low signal-to-noise ratio, a restricted angular range ( missing wedge), and the lack of ground truth. Here, we present Icecream, which follows the broad template of earlier self-supervised approaches, but treats symmetry in a way consistent with the recent equivariant imaging framework (Chen et al., 2021). Coupled with several engineering refinements, including mixed-precision arithmetic, Icecream achieves substantially better denoising and more reliable missing-wedge recovery, while reducing training and inference time relative to comparable baselines. Across diverse experimental datasets, we observe consistent gains in reconstruction quality, both visually and as quantified by Fourier shell correlation (FSC). Our framework extends to any tomography problem that provides two statistically independent reconstructions of the same volume; in cryo-ET these are obtained by dose splitting or angular partitioning of the tilt series.