Abstract
Cryo-electron tomography (cryo-ET) visualizes 3D cellular architecture in its near-native state. The various deep-learning methods have improved denoising and artifact correction, but remain challenged by a very low signal-to-noise ratio, a restricted tilt range (`missing wedge') and the lack of ground truth. Here, we present ICECREAM, which bridges earlier self-supervised methods with the recent equivariant imaging framework [Chen et al. (2021), IEEE/CVF International Conference on Computer Vision (ICCV), pp. 4359-4368]. Across diverse experimental datasets, ICECREAM achieves substantially better denoising and more reliable missing-wedge filling than existing methods. ICECREAM can be applied to any tomography problem that provides two statistically independent views of the volume; in cryo-ET these are obtained by dose splitting or angular partitioning of the tilt series. ICECREAM is openly available at https://github.com/swing-research/icecream.