Convolutional Neural Network-Based Reconstruction of Binarized Ultrasound Data for Non-Destructive Testing

This article introduces an exploratory project focused on reconstructing original ultrasound data from binarized observations as used in the Phase Coherence Imaging (PCI) method proposed by [1]. The PCI method, gaining popularity in the field of non-destructive testing, involves extracting only the phase information from the data while discarding amplitude details. The challenge of reconstructing the original data from its binarized form shares similarities with well-known machine learning tasks, such as image colorization, which is fundamentally an ill-posed problem. In this project, we investigate techniques inspired by approaches commonly used for addressing these tasks. Using the popular U-Net architecture, we show that the complete ultrasound signal can be reconstructed from the phase information that remains in the binarized signal used for the PCI representation, which is easy to store and process in a portable device.