Real-time in-vivo reverberation suppression of ultrasound channel signals using convolutional neural network
Abstract/Contents
- Abstract
- With the rise in obesity in the US, the population is becoming increasingly at risk of developing ailments such as Type II diabetes, fatty liver disease and non-alcoholic steatohepatitis (NASH). NASH can also develop into hepatic fibrosis and cirrhosis, and lead to an increased chance of hepatocellular carcinoma (HCC). As the outcomes of the obesity epidemic in the US continue to unfold, the ability to effectively survey obese patients for NASH and HCC is imperative. ultrasound is well suited for liver imaging due to the liver's size and location in the abdomen, however there are distinct challenges related to visualizing liver structures due to noise introduced by the increased layers of subcutaneous fat. These layers of fat and their corresponding connective tissues increase phase aberration and diffuse reverberation noise in ultrasound images. This work aims to reduce one of these noise sources, diffuse reverberation. Diffuse reverberation noise occurs when the acoustic wave transmitted from the ultrasound transducer reflects multiple times in the imaging medium before returning. This multi-path scattered signal overlays the signal from single reflections in the deeper tissue, causing a haze over the region that looks similar to speckle. Not only does it decrease the contrast of the ultrasound image, reverberation noise also introduces errors to techniques that rely on radiofrequency (RF) channel signals, such as phase aberration correction, Doppler imaging, sound speed estimation, shear-wave estimation and adaptive beamforming techniques. Though some techniques exist to address reverberation in ultrasound channel signals, each technique has shortcomings that make it impractical to use on the ultrasound scanner in real time, without altering the scanner's hardware. In the past years, publications using neural networks and machine learning have flooded almost every research field that works with quantitative data. These numerous articles show again and again that neural networks are a powerful emerging tool to solve previously difficult problems. They are especially making headway in creating approximation functions for non-linear processes. In the field of medical imaging, neural networks have shown great success in automated diagnosis, improved resolution and more. This dissertation explores the use of fully convolutional neural networks (CNNs) to reduce reverberation noise in abdominal ultrasound imaging while preserving its use as a real-time imaging modality. In order to be used for a pre-processing filter for other techniques outside of imaging, the reverberation noise is reduced on the RF channel signals. The initial stages of this work begin by training a 3D convolutional neural network on a simple training dataset simulated using Field II pro, a linear acoustics simulator. The resulting neural network produced qualitative and quantitative reduction of reverberation on an in-vivo test dataset of the carotid artery and the thyroid. When developed, results from this neural network were very promising on their reverberation reduction performance, especially in metrics that indicated the presence of reverberation. However, the neural network architecture was computationally heavy, and could only be run at 9fps in real-time on an ultrasound scanner. Additionally, this network was not optimized for abdominal imaging, which is the target task of this work. The next stages progressed to focus more on the application of abdominal imaging, and improve the real-time implementation speed. To improve implementation speed, a series of neural network architecture experiments were performed to give similar performance on a light-weight, 2D convolutional neural network with permutations. To improve reverberation reduction, a more physically-accurate fullwave simulation dataset was simulated for training. The resulting neural network was implemented in real time on an ultrasound scanner to give a frame rate of 16 fps. A clinical study was run on 15 volunteers imaging the liver and kidney for neural network evaluation, resulting in improved liver vessel visibility in all selected quantitative metrics across an 85 image test set of liver images. The preservation of aberration and the stability of the network were also assessed.
Description
Type of resource | text |
---|---|
Form | electronic resource; remote; computer; online resource |
Extent | 1 online resource. |
Place | California |
Place | [Stanford, California] |
Publisher | [Stanford University] |
Copyright date | 2022; ©2022 |
Publication date | 2022; 2022 |
Issuance | monographic |
Language | English |
Creators/Contributors
Author | Brickson, Leandra Lynn |
---|---|
Degree supervisor | Dahl, Jeremy J, 1976- |
Thesis advisor | Dahl, Jeremy J, 1976- |
Thesis advisor | Nishimura, Dwight George |
Thesis advisor | Pauly, John (John M.) |
Degree committee member | Nishimura, Dwight George |
Degree committee member | Pauly, John (John M.) |
Associated with | Stanford University, Department of Electrical Engineering |
Subjects
Genre | Theses |
---|---|
Genre | Text |
Bibliographic information
Statement of responsibility | Leandra Brickson. |
---|---|
Note | Submitted to the Department of Electrical Engineering. |
Thesis | Thesis Ph.D. Stanford University 2022. |
Location | https://purl.stanford.edu/jf338vn3085 |
Access conditions
- Copyright
- © 2022 by Leandra Lynn Brickson
- License
- This work is licensed under a Creative Commons Attribution 3.0 Unported license (CC BY).
Also listed in
Loading usage metrics...