Generalization of deep learning image restoration method for compressed sensing in electron tomography with a limited number of projections

Abstract

Electron tomography (ET) is a technique for 3D nanoscale characterization whose practical application is often hampered by severe artifacts arising from an experimentally limited number of projections and a restricted tilt range. While methods like Compressed Sensing (CS) have been developed to address this data scarcity, their performance degrades significantly under highly constrained conditions. This paper introduces a novel deep learning methodology for image restoration that overcomes these limitations. We propose a supervised Convolutional Neural Network (CNN) architecture based on conditional GAN and RIDNet to eliminate artifacts from initial reconstructions. The central innovation lies in our training strategy: the network is trained exclusively on simple geometric primitives, such as circles and squares, thereby circumventing the need for large, complex, and sample-specific training datasets. We demonstrate that a network trained on this simple basis can remarkably generalize to restore complex, irregular nanomaterials that it has never seen. Quantitative and qualitative comparisons demonstrate that our method significantly outperforms traditional CS, producing high-fidelity 3D reconstructions free of common artifacts. This work establishes a broadly applicable and data-efficient restoration framework that presents a robust and accessible tool for improving the reliability of electron tomography in materials science.