Ultrasound Brain Tomography: Comparison of Deep Learning and Deterministic Methods

The general purpose of this document is to develop a lightweight, portable ultrasound computer tomography (USCT) system that enables noninvasive imaging of the inside of the human head with high resolution. The goal is to analyze the benefits of using a deep neural network containing convolutional neural network (CNN) and long short-term memory (LSTM) layers compared to deterministic methods. In addition to the CNN + LSTM and LSTM networks, the following methods were used to create tomographic images of the inside of the human head: truncated singular value decomposition (TSVD), linear backprojection (LB), Gauss-Newton (GN) with regularization matrix, Tikhonov regularization (TR), and Levenberg-Marquardt (LM). A physical model of the human head was made. Based on synthetic and real measurements, images of the inside of the brain were reconstructed. On this basis, the CNN + LSTM and LSTM methods were compared with deterministic methods. Based on the comparison of images and quantitative indicators, it was found that the proposed neural network is much more tolerant of noisy and nonideal synthetic data measurements, which is manifested in the lack of the need to apply filters to the obtained images. An important finding confirmed by hard evidence is the confirmation of the greater usefulness of neural models in medical ultrasound tomography, which results from the generalization abilities of the deep hybrid neural network. At the same time, research has shown a deficit of these abilities in deterministic methods. Considering the human head's specificity, using hybrid neural networks containing both CNN and LSTM layers in clinical trials is a better choice than deterministic methods.