New research shows that artificial intelligence can detect COVID-19 in lung ultrasound images similar to how facial recognition software can identify faces in a crowd.
This discovery is a significant advancement in AI-driven medical diagnostics, bringing healthcare professionals closer to quickly diagnosing patients with COVID-19 and other pulmonary diseases by using algorithms to analyze ultrasound images for signs of illness.
Published in Communications Medicine, these findings are the result of an initiative that began during the early stages of the pandemic when healthcare providers needed tools to rapidly evaluate large numbers of patients in overwhelmed emergency rooms.
Lead author Muyinatu Bell, a professor at Johns Hopkins University, stated, “We created this automated detection tool to assist doctors in emergency situations with a high volume of patients who require prompt and accurate diagnoses, especially during the initial phases of the pandemic. Our goal is to potentially provide wireless devices that patients can use at home to monitor the progression of COVID-19.”
This tool also has the potential to develop wearables that monitor conditions like congestive heart failure, which can cause fluid accumulation in the lungs similar to COVID-19, according to co-author Tiffany Fong from Johns Hopkins Medicine.
Fong added, “The use of AI tools in point-of-care settings represents the next frontier. An ideal application would be wearable ultrasound patches that monitor fluid buildup and alert patients when they need medication adjustments or medical attention.”
The AI technology analyzes ultrasound images of the lungs to identify features known as B-lines, which are bright, vertical abnormalities indicating inflammation in patients with pulmonary issues. It combines computer-generated images with real ultrasound scans from patients, including those treated at Johns Hopkins.
Bell explained, “We had to accurately model the physics of ultrasound and sound wave propagation to generate believable simulated images. We then trained our computer models to interpret real patient scans with affected lungs using this simulated data.”
Initially, scientists faced challenges using AI to analyze COVID-19 indicators in lung ultrasound images due to limited patient data and a lack of understanding of how the disease presents in the body. Bell’s team developed software that can learn from a combination of real and simulated data to detect abnormalities in ultrasound scans indicative of COVID-19 infection.
Lead author Lingyi Zhao, who developed the software while at Bell’s lab and now works at Novateur Research Solutions, stated, “During the early stages of the pandemic, we lacked sufficient ultrasound images of COVID-19 patients to train our algorithms, resulting in suboptimal performance. However, we have now demonstrated that using computer-generated datasets can achieve high accuracy in evaluating and detecting COVID-19 features.”
The team has made their code and data publicly available at: https://gitlab.com/pulselab/covid19
