Machine Learning Denoising Solution for Acoustoelectric Imaging

Abstract

Acoustoelectric imaging is a novel medical imaging technique that uses ultrasound to image the flow of current in a tissue. Acoustoelectric imaging has a wide range of high impact applications such as being able to study and diagnose diseases of the heart and brain as well as being a potential surgical assistive tool. The primary obstacle in bringing acoustoelectric imaging to the forefront of medical imaging is the inability to effectively perform the technique noninvasively. Performing the technique in-vivo requires cutting edge ultrasound and radiofrequency (RF) sensing technology, and along with the ultra-high sensitivity of the instruments comes a high susceptibility to background noise. The very small acoustoelectric signals end up buried in a high noise floor, thus the need for high performing denoising capabilities. This thesis explores the potential of pursuing machine learning as a denoising solution. An acoustoelectric data simulation pipeline is used to provide realistic acoustoelectric data along with the corresponding ground truth. A machine learning architecture called HINet is explored for its reputation as being high performing in image restoration tasks. Through a series of hyperparameter testing, the HINet model is optimized for the simulated acoustoelectric data and is found to outperform the traditional denoising methods by an average of 10dB by PSNR.