With the world experiencing another wave of the coronavirus pandemic that threatens to overwhelm hospitals and testing capacity, the ability to quickly diagnose COVID-19 based on alternative methodologies has become increasingly important. For patients with respiratory symptoms, review of CT scans emerged as a relatively reliable indicator of infection with SARS-CoV-2 from the first days of its emergence, but the need for more accurate readings remains.

“With the U.S. in the midst of an unprecedented rise in COVID-19 infections, with current hospitalizations at an all-time record of more than 90,000 patients, there is an increasing need for AI solutions in medical imaging,” said Moshe Becker, CEO and co-founder of Radlogics Inc. “Coronavirus-related infection rates are experiencing a sharp increase in most states – from rural communities to urban areas – that have the potential to overwhelm ER, ICU and radiology teams with a surge of patients, and AI-powered medical imaging analysis solutions are poised to reduce this pressure through improved patient triage, monitoring and management.”

Radlogics offers an artificial intelligence-powered image analysis solution that powerfully augments the evaluation of a trained radiologist and improves results, especially with so many of the specialists suffering exhaustion and burnout from the relentless demand for imaging and interpretation created by the pandemic. The results can be used to support or replace RT-PCR testing for COVID-19, helping physicians make appropriate care decisions within hours rather than waiting for testing results.

Instead of relying on radiologists’ familiarity with the signature appearance of coronavirus-infected lungs and their ability to distinguish early stages from other indications under extraordinary pressure, “the solution provides quantitative measurement for patients with suspected COVID-19 disease including a score and a severity measurement to monitor findings over time,” Becker told BioWorld.

Other enhancements to the Boston and Tel Aviv-based company include three fully automated analyses on raw CT scans – three-dimensional lobe segmentation, identification of regions-of-interest, and detection of ground glass opacities.

These analyses provide key information to facilitate differentiation of COVID-19 from other conditions. If, for instance, a patient has lung abnormalities in more than a predetermined number of image slices, they undergo segmentation of the disease manifestations and texture analysis that looks for ground glass opacities and consolidation. In the end, the solution produces a 3D visualization of both lungs superimposed with color-coded findings, Becker explained.

The visualizations call out critical factors, simplifying interpretation. “Various key images are produced, with the findings from every corresponding case displayed with contours, bounding boxes, heatmaps and texture images with suspected COVID-19 disease,” Becker said.

Radlogic’s U.S. FDA-cleared CT and X-ray medical image analysis solutions are available to hospitals and health care systems across the country and work with any reporting system. As an on-premises SaaS solution, the system’s findings, measurements, alerts, and key images are delivered within minutes to the radiologist’s screen, imaging and reporting system.

The system may be installed onsite or accessed remotely for rapid deployment across multiple facilities and integration with existing workflows. The platform can process up to 1 million CT studies daily.

Measuring the algorithm’s performance

The enhancements draw on the solution’s experience processing and analyzing hundreds of thousands of suspected cases of coronavirus worldwide over the course of the pandemic. The algorithm’s ability to accurately distinguish and triage COVID-19 cases was recently validated in a research paper on the arXiv preprint server.

The researchers used multiple descriptive features such as lung and infections statistics as well as texture, shape, and location of abnormalities to train the AI then evaluated it on a dataset of 2,191 CT scans of individuals with a wide range of lung pathologies, including community acquired pneumonia.

The system demonstrated a 90.8% sensitivity and 85.4% specificity with 94.0% area under the curve in this study.

A previous evaluation found the system was able to distinguish between 157 patients with and without COVID-19 with a 0.996 area under the curve and 98.2% sensitivity and 92.2% specificity.