Deep learning model may predict lung cancer risk from a single CT scan

Trained and developed on over 35,000 low-dose computed tomography (LDCT) scans and validated in three independent data sets, a deep learning algorithm demonstrated an average area under the curve (AUC) of 90.6 percent for predicting lung cancer within one year.

Deep learning assessment of a single low-dose computed tomography (LDCT) scan may allow a highly accurate prediction of future lung cancer, according to newly published research.

In a retrospective study, published in the Journal of Clinical Oncology, researchers trained and developed a deep learning algorithm with a total of 35,001 LDCT scans from participants in the National Lung Screening Trial (NLST). For predicting future lung cancer based solely on LDCT assessment, the deep learning algorithm had a 1-year area under the curve (AUC) of 92 percent, a 2-year AUC of 86 percent and a 6-year concordance index (C-index) of 75 percent, according to the study.

In subsequent external validation testing, the researchers tested the deep learning algorithm with a 13,309 LDCT (6,392 patients) data set from the Massachusetts General Hospital (MGH) and a 12,480 LDCT (10,696 patients) data set from the Chang Gung Memorial Hospital (CGMH) in Taiwan. The study authors found the deep learning algorithm had an 86 percent 1-year AUC and an 81 percent 6-year C-index in the MGH data set as well as a 94 percent 1-year AUC and an 80 percent C-index in the CGMH data set.

While acknowledging that prospective clinical trials are necessary to confirm the clinical utility of these study finding, the researchers suggested the deep learning algorithm may help reduce unnecessary follow-up imaging in patients deemed to have low-risk lung nodules.

“On the basis of our clinical results, one potential clinical application is to use (the deep learning algorithm) to decrease follow-up scans or biopsies among patients with nodules that are low risk,” wrote Regina Barzilay, Ph.D., who is affiliated with the Department of Electrical Engineering and Computer Science and the Jameel Clinic at the Massachusetts Institute of Technology in Cambridge, Mass, and colleagues. “ … In our assessment of the NLST test set, (the deep learning algorithm) further reduced the FPR (false-positive rate) to 8% for baseline scans, compared with 14% for Lung-RADS 1.0, while maintaining equivalent sensitivity.”

The researchers also noted a correlation with the deep learning algorithm between the prediction of high cancer risk and targeted identification of where the malignant nodules would occur.

“We noted an association between (the ability of the deep learning algorithm) to correctly lateralize the location of future cancers and the likelihood that an LDCT receives a high-risk score, indicating that when (the deep learning algorithm) predicts high future lung cancer risk, the signal it uses localizes to specific at-risk regions rather than being equally spread over the entire thorax,” pointed out Barzilay and colleagues.

In regard to study limitations, the authors acknowledged the retrospective nature of the study. Noting that the study cohorts were comprised of patients in lung cancer screening programs, they said they could not assess the deep learning model’s ability to predict lung cancer among people who are not in lung cancer screening programs. The study authors also conceded suboptimal diversity in the study cohort and the lack of a true comparator model. Diagnostic Imaging