Effective Super-Resolution X-ray Tomography using MSDnet for Nondestructive Testing of Metallic Lattices: Analysis of Training Dynamics and Strategies

Abstract

Deep learning-based super-resolution has shown significant potential for enhancing the resolution of low-resolution X-ray computed tomography (CT), a 3D nondestructive imaging technique. It could accelerate CT scanning by several orders of magnitude, opening new possibilities for high-throughput acquisition, in situ, and low-dose experiments. However, current assessments of the resulting image quality often rely on 2D image quality metrics such as PSNR and SSIM, which may not correlate directly with scientific measurements. In the present study, the relationship between training dynamics and image quality in deep learning super-resolution is investigated. A super-resolution method was applied to a stainless steel lattice structure featuring numerous quantifiable defects, imaged with a laboratory CT. The core contribution lies in employing a 2.5D Mixed-Scale Dense neural Network (MSDnet) on experimental data and evaluating its performance using scientific and task-based metrics–specifically related to porosity and surface roughness–while monitoring training dynamics. The results demonstrate that even a standard loss function can effectively reflect network performance dynamic for such material science applications. The best super-resolution accuracy with a magnification factor of 3 was achieved after 100 epochs, generating less than 2 % of missing pores and only around 15 % average error in pore volume. Additionally, practical considerations are proposed to assist in the design of tailored training strategies.