High Resolution Microwave Tomographic Imaging of Brain Strokes Using Low-Frequency Measurements and Deep Neural Networks

点击翻译按钮获取中文摘要

PROJECT SUMMARY / ABSTRACT According to the CDC, a stroke occurs in the United States every 40 seconds, with a fatality every 4 minutes and associated reduction in mobility in more than half of survivors of ages 65 and over. The ability to differentiate ischemic/hemorrhagic strokes in the pre-hospital setting and to monitor stroke evolution by the bedside has the great potential to improve outcomes and reduce mortality. Unfortunately, state-of-the-practice MRI and CT systems are bulky and pricy, restricting imaging to the clinical setting and sparse intervals. CT also uses ionizing radiation that poses safety risks and further prohibits frequent imaging. Microwave Tomographic Imaging (MTI) is a promising alternative/complementary option to MRI and CT, but has yet to be used in the clinical setting. This is mainly due to its poor spatial resolution as feature dimensions are comparable to the wavelength of the electromagnetic wave. Unfortunately, reducing the wavelength (i.e., increasing the measurement frequency) of MTI is not viable as high frequencies are prone to noise and severe attenuation inside tissues. Instead, our goal is to explore the feasibility of expanding the fundamental limits of MTI resolution via innovations in estimating high-frequency data from low-frequency measurements using Deep Neural Networks (DNNs). We target detection of strokes <1cm×1cm that meets clinical expectations for a much needed addition to the pre-hospital setting and throughout the stroke monitoring process. Hypothesis 1: A relationship exists between the low- and high-frequency data measured around a biological imaging domain that we can use to ‘artificially’ increase the highest usable frequency for any given low-frequency measurements. Hypothesis 2: An ‘artificial’ increase in frequency by N times will improve image resolution by N times, regardless of the MTI reconstruction method used. Here, N depends on the highest usable frequency (to be determined) and is expected to be at least equal to two. The study is significant because it reveals previously unknown knowledge for enhancing MTI resolution in biological media. In Aim 1, we will develop the DNN using 2D/3D solvers, canonical/anatomical head models, and a new class of into-body radiating antennas with unprecedented efficiency. Our study will validate Hypothesis 1. In Aim 2, we will validate the DNN numerically by using the estimated high-frequency data to reconstruct the image. Our study will validate Hypothesis 2. In Aim 3, we will validate the DNN experimentally using tissue-emulating phantoms. Successful reconstruction will entail improved (N times higher) image resolution vs. state-of-the-art MTI reconstruction at the same measurement frequency. A comparison of image reconstruction accuracy using actual vs. estimated high-frequency data will further reveal the method’s efficacy. Feasibility will form the basis of future studies on human subjects. We envision this technique to be a much needed breakthrough to overcoming the upper frequency limit of MTI algorithms for various diagnosis and/or pre-hospital assessment applications in brain stoke applications and beyond.