By Kimberly Mann Bruch
A University of California San Diego team suggests AI could help diagnose and stage chronic obstructive pulmonary disease (COPD) using computed tomography (CT) scans, potentially simplifying the diagnostic process for millions of patients.
Albert Hsiao, M.D., Ph.D., a clinical radiologist at UC San Diego Health, professor of radiology at the UC San Diego School of Medicine and professor with the Halıcıoğlu Data Science Institute – part of UC San Diego’s School of Computing, Information and Data Sciences – worked with Kyle Hasenstab, a former post-doctoral fellow at UC San Diego and now assistant professor of mathematics and statistics at San Diego State University, and Amanda Lee, a UC San Diego graduate student, to study how convolutional neural networks (CNNs) can diagnose and predict COPD severity.
“COPD affects millions of Americans and is a leading cause of death worldwide,
and our research suggests that AI analysis of CT scans could complement or potentially streamline the diagnosis.” -Albert Hsiao, M.D., Ph.D.
CNNs, which are computer systems that learn to recognize patterns in images and other visual data, help AI understand and interpret sights much like humans do. The team found that CNNs can diagnose and predict COPD severity with moderate to good accuracy using only a single-phase CT scan, potentially improving the effectiveness of CT for diagnosing lung disease. They used data collected from the COPDGene® study, including 8,893 clinical participants with an average age of 59.6 years – with slightly more than half being male.
“Importantly, the addition of clinical data improved the AI’s performance across nearly all metrics, suggesting that combining imaging with patient information yields the best results,” Hsiao said. “COPD affects millions of Americans and is a leading cause of death worldwide, and our research suggests that AI analysis of CT scans could complement or potentially streamline the diagnosis.”
The team reported their findings in the Radiology: Cardiothoracic Imaging journal and concluded that the accuracy of single-phase CT scans matches that of dual phase (inspiratory-expiratory) scans when combined with clinical data. Hsiao said that the team’s AI models showed moderate to good agreement with traditional spirometry measurements, a traditional method of diagnosis for COPD that measures how well a person inhales and exhales. Accuracy for predicting COPD severity with the AI models generated during the COPDGene® study ranged from 59.8% to 88%, depending on the model and data used.
Hsiao said that this research builds on their prior work showing that the combination of AI and CT can be effectively used to diagnose and stage COPD.
In conjunction with their studies, Hsiao and his group published a parallel piece, also in the Radiology: Cardiothoracic Imaging journal, on the surveillance of lung transplant patients with inspiratory-expiratory CT for signs of lung disease and graft rejection. In this work, they found that AI measurements of air trapping were strongly predictive of functional decline of lung allografts. Hsiao explained that this technology may not only help to diagnose COPD, but also improve the ability to extend the longevity of patients who require lung transplantation to treat a variety of other lung diseases.
The COPDGene® U01 research was funded by the National Heart, Lung, and Blood Institute (award nos. U01-HL089897 and U01-HL089856).
