An AI algorithm created by Google can predict lung cancer with high accuracy and improve the survival chances of those with the cancer through earlier diagnosis, according to a recent study. The findings were published in Nature Medicine on May 20.
Lung cancer is the most common cause of cancer death in the nation, with an estimated 160,000 deaths in 2018 and more than 1.7 million deaths per year around the world. However, just 2% to 4% of eligible patients in the U.S. are screened. In addition, the disease also has one of the worst survival rates among all cancers, though chances of survival are improved when the cancer is caught earlier.
Google’s project aimed to improve current methods for predicting lung cancer, which have seen improvements through lower-dose CT screening. Under current methods, radiologists look at hundreds of 2D images in a CT scan to spot cancer. Google’s model creates a 3D generation of overall lung cancer malignancy prediction that can also identify malignant tissue in the lung nodules that are subtle and taking into account previous scans that can show the growth rates of suspicious nodules.
Overall, Google’s model was able to perform on par with or better than six U.S. board-certified radiologists when using a single CT scan for a lung cancer diagnosis, according to the study. Google researchers used more than 45,000 de-identified chest CT screening cases from the National Institute of Health’s research dataset from the National Lung Screening Trial study and Northwestern University. From there, Google’s algorithm was tested on 6,716 trial cases and clinically validated on a set of 1,139 cases, achieving an area under the curve of 94.4%.
The model reduced false positive exams 11% and detected 5% more cancer cases compared to unassisted radiologists in the study. When previous scans weren’t available, the model outperformed the radiologists, detecting potential lung cancer that had previously been called normal, according to a blog post by Google.
Google is working with its Cloud Healthcare and Life Science team to make the model available through the Cloud Healthcare API and continue testing.
“While the vast majority of patients remain unscreened, we show the potential for deep learning models to increase the accuracy, consistency and adoption of lung cancer screening worldwide,” Diego Ardila, first author and senior software engineer at Google, et al. wrote in the study.