Google Health’s AI tool slashes abnormal chest X-ray turnaround times by 28%
Google recently unveiled a new deep learning algorithm that can distinguish normal from abnormal X-rays and dramatically reduce turnaround times.
The Palo Alto internet giant’s health division trained and tuned their AI system using chest X-rays from nearly 250,000 patients. During testing, the tool adapted well to patient populations and diseases it hadn’t seen before, including COVID-19.
Zaid Nabulsi, a software engineer at the company, and colleagues reported promising results, sharing them this month in Scientific Reports. AI notched high scores for separating normal from abnormal images across various abnormalities and healthcare settings.
What’s more, the AI expedited abnormal cases and reduced turnaround times for these urgent matters by 28%.
“This reprioritization setup could be used to divert complex abnormal cases to cardiothoracic specialist radiologists, enable rapid triage of cases that may need urgent decisions, and provide the opportunity to batch negative CXRs for streamlined review,” the authors wrote in the study.
The group utilized a deep learning system pre-trained on ImageNet, further tweaking it with de-identified data from 248,445 patients treated across a hospital network in India. Each CXR was labeled as “normal” or “abnormal” based on associated rad reports.
When tested on exams split between the Apollo Hospital datasets from India and publicly available ChestX-ray14, the system achieved areas under the receiver operating characteristic curve scores of 0.87 and 0.94, respectively.
In order to further determine its capabilities in settings with diseases never encountered before, the Google team fed its AI four datasets. Two included public tuberculosis data while the others contained COVID-19 exams. The AI scored AUCs between 0.95-0.97 for the former and 0.65-0.68 for the latter.
Finally, Nabulsi and the Google Health team simulated using AI to prioritize studies, reporting 28% shorter turnaround times for abnormal exams.
“These results represent an important step towards evaluating whether AI can be safely used to flag cases in a general setting where previously unseen abnormalities exist,” the team explained.
Matt joined Chicago’s TriMed team in 2018 covering all areas of health imaging after two years reporting on the hospital field. He holds a bachelor’s in English from UIC, and enjoys a good cup of coffee and an interesting documentary.