Session: Research Posters

Paper Number: 119747

119747 - A Deep Learning Based Approach to Improve Reconstruction of Time-Domain Full Waveform Inversion 

Ultrasound computed tomography (USCT) has emerged as a prominent technique in the fields of nondestructive evaluation (NDE), structural health monitoring (SHM), and medical imaging, primarily due to its cost-effectiveness and rapid data acquisition capabilities. Full waveform inversion (FWI), a partial differential equation (PDE)-constrained optimization approach, has shown great promise in USCT. It is a cutting-edge inversion technique that utilizes the information contained in ultrasound scanning data to accurately determine material properties, such as wave speeds of different materials. However, time-domain FWI (TDFWI) can be computationally demanding and time-consuming despite its scalability. It aims to reconstruct the material distribution by iteratively calculating the gradient of waveform differences between measured and generated signals. Recent studies have shown that this iterative process can be considerably accelerated by incorporating AI-based methods, like convolutional neural networks (CNNs). Nonetheless, the current CNN-based approaches face challenges in maintaining reconstruction quality as the complexity of material distributions increases. Another challenge is the availability of sufficient experimental data, and in some cases, even synthetic surrogate data.

To address these challenges, this study systematically investigates the potential enhancement of a 2D CNN (U-Net) to improve the quality of reconstruction using limited training data. Various augmentation strategies are applied, such as horizontal and vertical flipping and data mixing. The mixing approach generates new synthetic samples by combing pre-existing training samples together. This was done to expand the training data's volume without the need to create an excessive number of samples. The objective is to evaluate the impact of these augmentation techniques on the performance of the U-Net model. Multiple datasets generated from 1-10 iterations of the reconstructed model have been developed to train the U-Net model. The reconstructed models are then compared to ground truth images using evaluation metrics such as the structural similarity index measure (SSIM) and average mean square error (MSE). A thousand numerically generated samples with acoustic material properties are employed to construct multiple datasets from different FWI iterations. Generating multiple datasets with the TDFWI setup can be still very computationally expensive and time-consuming. A parallelized, high-performance computing (HPC)-based framework has been developed to expedite the generation of training data.

The results demonstrate that the augmentation strategy does not provide significant improvements in prediction results when using higher iterations of the FWI training dataset. This can be attributed to the already improved performance of the FWI reconstruction at higher iterations, prior to being processed by the U-Net model. However, the inclusion of additional samples through augmentations does enhance the imaging of complex regions, particularly when using a low-iteration FWI training dataset.

Presenting Author: Shoaib Anwar The University of Alabama

Presenting Author Biography: Shoaib Anwar is a 2nd year Ph.D. student in the Aerospace Engineering and Mechanics department at The University of Alabama. He works in the Computational Imaging of Smart Structure lab with Dr. Jiaze Ha. His research includes novel ultrasonic imaging methods for material/structures/medical imaging, and machine learning. He completed his undergrad in mechanical engineering at Bangladesh University of Engineering and Technology. He also has three years of working experience in an automotive company in Bangladesh.

Authors:

Shoaib Anwar The University of Alabama
Austin Yunker Argonne National Laboratory
Rajkumar Kettimuthu Argonne National Laboratory
Mark Anastasio University of Illinois Urbana-Champaign
Umberto Villa The University of Texas at Austin
Jiaze He The University of Alabama