A Latent Back-Projection Network for Novel Projection Synthesis for Improved Cryo-ET

Abstract

Cryo-electron tomography (Cryo-ET) is hindered by the missing wedge, a gap in Fourier-space information caused by limited tilt-series angular coverage, leading to anisotropic resolution loss and artifacts. Current methods, such as IsoNet, attempt to "inpaint" missing frequencies in reconstructed tomograms but are constrained by their reliance on pre-degraded data, often producing non-physical features. We present our proof-of-concept model, a generative latent back-projection autoencoder that bypasses traditional tomogram reconstruction and directly synthesizes novel projections from tilt-series data in the frequency domain. Our latent back-projection network encoder-decoder architecture maps raw projections to a 3D Fourier volume, leveraging the Fourier slice theorem to generate high-fidelity projections beyond the experimental tilt range. Evaluated on \textit{E. coli} mini-cells, our model achieves a lower MSE and higher correlation with ground-truth data. Crucially, our model robustly recovers withheld tilts (0° or ±15°) without retraining, outperforming IsoNet in MSE and correlation metrics. By mitigating the missing wedge through generating tilts at new angles, our proof-of-concept can potentially advances high-resolution in situ structural biology for radiation-sensitive specimens.