Deep-Learning-Based Ultrasound Computed Tomography for Material Characterization

Abstract Full waveform inversion (FWI) is one of the advanced inversion-based ultrasound computed tomography (USCT) techniques that can identify the material properties and produce high-resolution inverted images of the specimen, making it promising for applications in nondestructive evaluation (NDE) and structural health monitoring (SHM). However, this process is computationally expensive and time-consuming due to its ill-posed nature and inherent nonlinearity. To mitigate these challenges and enhance the efficiency of FWI, this study explores an alternative approach using artificial intelligence (AI) as a data-driven regularization. However, one of the major challenges in such a data-driven approach is the inadequate amount of data to train the neural network. To address this bottleneck, 2D numerically generated samples are used to train the neural network in this study. In addition, two different source time functions (STF) were used for FWI simulations to understand the effect of using different waveforms on the deep-learning (DL) based solution. The first STF was numerically generated and the second was acquired by inverting experimental data using a multi-channel data acquisition system from the Verasonics. For evaluation, the structural similarity index measure (SSIM) and peak-signal noise ratio (PSNR) metrics are used to compare the predicted images with the ground truth images. The results demonstrate the efficacy of the DL-based approach, with improvements of 22.5% and 78% in SSIM and 47% and 73% in PSNR over the classical FWI simulations for the first and second STFs, respectively. The DL-based adjoint tomography architecture for USCT presented in this work has a strong potential for material characterization and detecting internal defects and variations.