Google Cloud supports radiology research to improve COVID diagnoses on CT

A Colorado researcher confirms that chest CT scans of COVID patients are over ten times more likely to show patterns of lung injury than other pneumonias requiring hospitalization. By using AI to interpret the datasets stored on Google Cloud, she hopes to use these preliminary findings to improve the diagnostic criteria for COVID, hone the existing scoring criteria, and potentially predict patient outcomes.

Like many of her colleagues, Kathryn Olsen, M.D., cardiothoracic radiologist at Banner Health, spent the early months of the COVID-19 pandemic interpreting computerized tomography (CT) chest scans of COVID patients. She observed that the patient’s initial chest CT displayed patterns that reflected lung injury rather than a typical pattern of viral infection. What was even more unique was that this lung injury pattern on CT was rarely seen in the emergency room setting. As she shared these insights with fellow front-line clinicians, she felt more sure that this pattern on CT imaging was significant, even though the evidence was anecdotal. What did the pattern mean and could it help diagnose COVID infections or predict outcomes for patients? Could those answers improve clinical decision-making during the global health crisis?

To answer these questions, Dr. Olsen set out to analyze around 300 CT chest scans acquired over the nine weeks corresponding to the Denver metro’s 2020 Spring COVID-19 surge. As a control set, she compared them to CT scans from the same institutions over the previous three years before COVID emerged. Now that the first arm of her experiment is completed, she has the data to support her hunch: she reports that “patients with COVID-19 who are admitted to a hospital are over ten times more likely to present with a lung injury pattern, rather than what we normally think of as a classic viral pneumonia pattern, validating what we were seeing as radiologists.”

With the right tools and data, one doctor can make a difference

Dr. Olsen’s results emerged from a broad collaboration between volunteer radiologists and ancillary staff, several Colorado healthcare institutions, and Flywheel, a data management and collaboration platform for medical research. Flywheel, which deploys its technology on Google Cloud, helped Dr. Olsen create a workflow for the CT imaging and associated data analysis and storage using Google Compute and Google Cloud Storage. Marco Comianos, VP of Sales, Academic, Clinical, and AI at Flywheel, says, “with a project of this scale, Principal Investigators typically require substantial resources to build IT infrastructure for managing data workflows and processing pipelines. Since the research was time-sensitive, Dr. Olsen was able to use Flywheel to quickly set up and manage this multi-center project from home. She worked closely with our team to remotely establish complex imaging workflows and integrations, such as blind reader studies, REDCap Data Capture software, and Imbio’s imaging biomarkers. This project is a true testament to how a single researcher can manage a complex research project using the Flywheel platform.”

Dr. Olsen explains that they had legal and privacy agreements with four Colorado medical institutions providing datasets: SCL Health, Centura Health, UC Health, and Banner Health. She says that “this is an unprecedented collaboration among these healthcare institutions. Flywheel provided not only the infrastructure for the project, but also deidentified and uploaded the massive datasets to their central repository on Google Cloud where our readers could then interpret all the CT scans in a blinded fashion.” Using the CO-RADS scoring criteria, five radiologists assigned each exam a score of 1 to 5 based on how unlikely (1) or very likely (5) the CT pattern and other imaging findings suggested a COVID diagnosis. Each CT exam was scored by three readers and Olsen adjudicated any ambiguous results.

Dr. Olsen and her team have recently partnered with the AI company Imbio Inc. Their early collaboration has shown that increased air space density (i.e. “consolidation”) in patients with COVID is predictive of poor clinical outcomes. “AI has the potential to identify various patterns on chest CT scans,” Dr. Olsen says. This may be even more important if the now-standard PCR tests turn out to be less reliable on future COVID variants.

With secure data storage and management, the cloud enables more collaboration

The collaboration has been as inspiring as the preliminary results: “it’s amazing how this brought so many people together, and so many people volunteered their time,” Dr. Olsen says. “The radiologists put over 100 hours each into this project because they believe in it. Flywheel supported us, even though we weren't funded, because they believe in it. We collaborated together to make a difference.”

It’s amazing how this brought so many people together….We collaborated together to make a difference.

Dr. Kathryn Olsen, Cardiothoracic Radiologist, Banner Health

Comianos adds, “In the past, data was shared on thumb drives, and it would be very difficult to manage blind reader studies and correlate data from multiple sources. With Flywheel being hosted on the cloud, we can provide a secure, centralized platform where data is de-identified and managed with appropriate data privacy and regulatory compliance. Cloud-based research has opened up our world to new data sharing and collaboration possibilities.”

Cloud-based research has opened up our world to new data sharing and collaboration possibilities.

Marco Comianos, VP of Sales, Academic, Clinical, and AI, Flywheel

Dr. Olsen participated in a webinar on her research with Flywheel and Imbio on June 29. The recording is available online. To get started with Google Cloud, apply for free credits towards your research.