A deep learning-based approach to improve reconstruction of ultrasound computed tomography with full waveform inversion

Enabled by multiple modalities of smart materials-based actuation and sensing, full waveform inversion (FWI) nowadays is an advanced ultrasound computed tomography technique that utilizes waveform data to generate high-resolution images of scanned regions. This technology offers promise for defect/damage detection and disease diagnosis, showing potential in nondestructive testing, structural health monitoring, and medical imaging. To reduce the lengthy computational time caused by time-domain FWI, modern AI-driven and data-driven approaches have been studied to accelerate the reconstruction process in recent years. However, most existing research focused on tuning specific neural networks for fixed-domain applications, leaving the relationship between model performance and characteristics of various training datasets underexplored. This paper presents a comprehensive investigation into the improvements achievable by integrating deep learning with the adjoint tomography theory, addressing the scientific questions of how amounts/distributions of the training data, augmentation, and loss functions influence the efficacy of this approach. The selected integration strategy involves training a U-Net neural network model using pairs of low-resolution inverted images and their corresponding high-resolution ground truth images. Once trained, the U-Net model instantaneously converts low-resolution inference images to high-resolution reconstructions. Generated using a proposed high-performance computing-based framework, multiple datasets were designed to offer a general representation of various applications while maintaining the shared characteristics across different use cases. The study also incorporates augmentation strategies to expand the size and complexity of the training dataset without significantly increasing the number of samples. Furthermore, a hyperparameter tuning framework was introduced to investigate the impact of multiple loss functions on the model performance. An additional challenge in data-driven approaches is the generalizability of the neural network model when exposed to out-of-distribution data. This study rigorously tests the model’s generalizability against several out-of-distribution datasets and finds that the U-Net model maintains a degree of generalizability when trained with unscaled datasets.