Loesch, Olivia; Leyba, Katie; Chan, Halyley; and Goergen, Craig, "Deep Learning Approach to Improved Image Quality for Medical Diagnostics" (2021). Discovery Undergraduate Interdisciplinary Research Internship. Paper 36.
Date of this Version
The United Nation’s health-related Sustainable Development Goals are difficult to achieve in low- and middle-income countries due to workforce shortages and inadequate health surveillance systems. However, with the growth of artificial intelligence (AI) and computer algorithms, it is possible to apply AI to healthcare technologies to improve progress towards these UN standards. This project aims at using and improving computer algorithms and deep learning to aid in the extraction of important structural and functional information from murine carotid artery ultrasound and photoacoustic images. First, we created a large database of simulated photoacoustic images to optimize the algorithms. These images were augmented binary masks of murine carotid arteries extracted from ultrasound images, which were then passed through MATLAB’s k-Wave toolbox. We used the toolbox’s time reversal reconstruction algorithm to simulate and reconstruct a photoacoustic image from the original ultrasound image. These simulated photoacoustic images were then passed through a convolutional neural network to improve image quality by increasing contrast-to-noise and signal-to-noise ratios, as well as reducing unwanted artifacts. Additionally, we acquired data from optimized phantoms and tested the algorithms on in vivo preclinical data to determine their effectiveness in improving image quality for cardiovascular applications. While preliminary, these initial results suggest improved image quality and allows us to use these algorithms for applications beyond carotid artery imaging. Using our deep learning approach, we will continue to add to the database of images available in order to perform more advanced tasks including pixel-wise classification and vessel segmentation.