Adaptive Gaussian representation for differentiable cryo-electron tomography reconstruction

Cryo-electron tomography (cryo-ET) enables 3D visualization of biological structures in their native state, but high-fidelity tomogram reconstruction remains challenging due to low signal-to-noise ratios and limited angular sampling. In this work, we present CryoETGS, a differentiable learning framework that reconstructs tomograms through adaptive 3D Gaussian representations of biological structures for cryo-ET. This representation enables efficient and interpretable reconstruction through a hardware-accelerated differentiable rendering pipeline aligned with the cryo-ET imaging geometry. CryoETGS incorporates hierarchical initialization, adaptive densification, and a tilt-weighted optimization strategy to enhance convergence and reconstruction fidelity. The framework further supports real-time projection synthesis and bidirectional conversion between voxel-based and Gaussian representations. Extensive experiments on both simulated and experimental datasets demonstrate that CryoETGS achieves state-of-the-art reconstruction performance, effectively mitigates missing wedge artifacts, and exhibits high computational efficiency. Source code is publicly available at https://github.com/JachyLikeCoding/ETGS.