spark139@jhmi.edu

The low dose CT (LDCT) screening criteria used in National Lung Screening Trial (NLST) has a 26.6% false positive rate at baseline. Even when updated more stringent Lung RADS criteria was retrospectively applied to the NLST data, the false positive rate remained at 12.8%. Deep learning, a form of artificial intelligence, has the potential to improve risk stratification of pulmonary nodules. The purpose of this study is compare the performance of deep learning vs. radiologists in the risk stratification of pulmonary nodules.

264 patients with one solid nodule reported in NLST database up to 20mm (mean±standard deviation: 7.5±3.4mm) in size (223 benign, 41 malignant) were retrospectively selected from the NLST baseline LDCT (T0). All malignant nodules were confirmed pathologically and benign nodules were diagnosed based on pathology or clinical follow-up by NLST investigators. The nodules were semi-automatically segmented using our in-house software. 3D deep convolutional networks (CNN) was used for the deep learning classification of malignancy based on 64x64x64 input patch bounding intramodular and perinodular areas. 4-fold cross-validation was performed. Data augmentation by scaling and rotating was used to increase the number of training dataset. Two radiologists who were blinded to the diagnosis reviewed the cases independently and scored the nodules based on Lung RADS criteria. Scores 1 and 2 were considered negative and scores >= 3 were considered positive.

The selected cohort was 62.0±5.1 year-old-patients at T0 (150 male and 114 female). The average accuracy, sensitivity, and specificity of the review of radiologists were 0.67, 0.73, and 0.67, respectively. 4-fold cross validation result of deep learning was 0.88, 0.90, and 0.88 in the same terms of accuracy, sensitivity, and specificity. Especially, the false positive rate showed significant improvement from 0.33 to 0.12, which represents to reduce false positive cases from 73 to 27, using CNN.

Deep learning achieved improved sensitivity, specificity, and accuracy in risk stratification of pulmonary nodules compared with radiologists.

Deep learning can improve the accuracy in risk stratification of pulmonary nodules compared with radiologists. This has the potential of achieving earlier cancer detection and reducing unnecessary work up in the lung screening population.

Presenters or authors on this event have been recognized as RSNA Honored Educators for participating in multiple qualifying educational activities. Honored Educators are invested in furthering the profession of radiology by delivering high-quality educational content in their field of study. Learn how you can become an honored educator by visiting the website at: https://www.rsna.org/Honored-Educator-Award/ Elliot K. Fishman, MD - 2012 Honored EducatorElliot K. Fishman, MD - 2014 Honored EducatorElliot K. Fishman, MD - 2016 Honored EducatorElliot K. Fishman, MD - 2018 Honored Educator