A comparison of 3D reconstruction algorithms for high-resolution radiographic imaging and tomography

Radiographic imaging and tomography (RadIT) using X-rays and other forms of ionizing radiation, including neutrons and energetic protons, have found wide applications in off-line and real-time interrogation of complex and even dynamic structures in materials science, medical imaging, non-destructive testing, plasma science and engineering. Three-dimensional (3D) reconstruction algorithms from 2D projections play a crucial role in high-resolution RadIT. The first part of this work gives a summary overview of different methods: traditional methods, such as Filtered Backprojection (FBP), Principal Component Analysis (PCA)-based backprojection, iterative reconstruction techniques (IRT), Model-Based Iterative Reconstruction (MBIR), Algebraic Reconstruction Technique (ART) and Simultaneous Algebraic Reconstruction Technique (SART), emerging data-driven methods, such as Convolutional Neural Networks (CNNs), Generative Adversarial Networks (GANs), and hybrid methods, based on a combination of statistical approaches (e.g. Bayesian inference), computation and novel measurement techniques (e.g. compressed sensing). Traditional FBP remains valuable for its speed and simplicity, while newer algorithms offer improved quality at the cost of computational time and modern hardware such as GPUs. Deep learning approaches show great potential for automated high-quality reconstructions but require extensive high-quality training data and may struggle with generalization. Generative methods can perform well with limited data but may introduce bias based on priors and underlying assumptions. The second part of work highlights results of using several methods to process various datasets from 3D printed structures and plasmas. LANL report number LA-UR-25-25951.