A group of campus and UCSF researchers created a computer algorithm to detect acute hemorrhages in CT scans.
Published Tuesday in the “Proceedings of the National Academy of Sciences of the United States of America,” the study details an algorithm that could make the process of reading CT scans more efficient and accurate, which is especially critical as radiologists are having to read an increasing number of scans. According to Weicheng Kuo, a co-author of the study and a full-time researcher at Google Brain, when radiologists read scans more rapidly, it increases the likelihood of making life-threatening errors.
“The computed found some abnormalities that were missed by the radiologists,” Esther Yuh, co-corresponding author of the study and UCSF associate professor of radiology, said in an email. “This clinical application requires fast interpretation for potentially life-threatening abnormalities, and requires a very high level of accuracy, even for tiny abnormalities.”
According to Yuh, radiologists often have to read thousands of scans per day. She added that this is problematic, as there are severe medical consequences to misreading a scan.
Additionally, many cases of acute hemorrhages are found in emergency room patients and need to be treated within an hour, which is difficult because radiologists typically need between five and 30 minutes to read a CT scan, Kuo said.
“The consequences of missing acute hemorrhages is very serious since hemorrhages can cause loss of brain function,” Kuo said. “There is a saying that ‘time is brain,’ meaning every second counts when having a hemorrhage.”
According to Kuo, the algorithm can read a three-dimensional scan of the head in one second and does so more consistently than radiologists, as they can get tired after reading many scans. He added that this is especially significant because it is difficult to get a machine to do things as well as a human.
The algorithm beat two out of four expert radiologists, making it more accurate than any previous algorithms despite using significantly fewer scans to train it, according to Kuo. This algorithm was trained with 4,396 CT scans, Kuo added, stating that it typically takes hundreds of thousands of scans to train algorithms.
Unlike previous algorithms, the researchers labeled the exact pixels on the screen that had acute hemorrhages, rather than having a machine guess where the acute hemorrhages were from radiology reports.
Although the new algorithm has been proven to be efficient and accurate, according to Kuo, there will still be a need for radiologists in the medical field, as it can only detect acute hemorrhages. Yuh added that, if used, the algorithm would most likely be incorporated into whatever system radiologists receive images from.
“The project was one of the most fun things I’ve ever done,” Yuh said. “I hope to collaborate with UC Berkeley in the future.”