Developing a CT-Based Classifier for Indeterminate Nodules Detected by Lung Cancer Screening
Posted November 15, 2018
Fabien Maldonado, M.D., Vanderbilt University
Dr. Fabien Maldonado
In recent years, the importance of computed tomography (CT) screening of individuals at high risk for lung cancer has become readily apparent. Data from the National Lung Screening Trial (NLST) and the European NELSON study demonstrated that low dose CT (LDCT) screening leads to a 20% reduction in lung cancer mortality of at-risk individuals. While LDCT has been particularly beneficial to patients with confirmed cancers, there is a high rate of detection of false-positive pulmonary nodules. In the NLST, 96% of nodules detected by LDCT were confirmed to be benign upon further study. Currently, there is no noninvasive means for confirming whether a pulmonary nodule detected by LDCT is benign or malignant, so patients with an indeterminate nodule must wait to undergo a biopsy to rule out cancer. Clearly, improved approaches for diagnosis of lung cancer are required to minimize misdiagnosis, patient stress, and exposure to unnecessary procedures.
Dr. Fabien Maldonado, at Vanderbilt University, was awarded a Lung Cancer Research Program (LCRP) FY14 Idea Development Award - New Investigator Option to develop algorithms to use high-resolution CT to provide a means to noninvasively characterize indeterminate pulmonary nodules, as part of a multi-institutional effort including Mayo Clinic and Brown University. Using radiomic features such as general characteristics of the nodule (size and location), nodule-specific characteristics (radiodensity, texture, and surface characteristics), and features of the nodule-free surrounding lung, Maldonado and his team of Drs. Tobias Peikert, Brian Bartholmai, Srinivasan Rajagopalan and Ronald Karwoski from Mayo Clinic, and Dr. Fenghai Duan from Brown University, using data from the NLST, developed and tested a predictive multivariate model for benign and malignant pulmonary nodules. Eight features coalesced to generate a promising model that could differentiate benign from malignant lung nodules. When scored using a common measure of model sensitivity and specificity, the Area Under the Receiver Operating Characteristic (AUROC) Curve, Dr. Maldonado’s model measured 0.94. This score corresponds to the ability of a predictive test to correctly distinguish true positives (in this case, patients with malignant lesions) from false positives (patients with benign lesions) while minimizing the likelihood of false negatives (misdiagnosing a patient who has a malignant nodule). A perfect predictive model would have an AUROC of 1.0, while random chance tends to yield an AUROC of 0.50. Dr. Maldonado’s model AUROC of 0.94 improves the likelihood of a correct diagnosis significantly compared to previous measures of radiologists alone, in groups, or working with computer-aided diagnosis software, which all provide an AUROC of 0.80 or less.
At the moment, Dr. Maldonado’s model remains to be externally validated. In the coming months, he will complete his Idea Development Award effort using CT data from the LCRP-funded Detection of Early Lung Cancer Among Military Personnel (DECAMP) Consortium to conduct this external validation. If the model performance remains as excellent as in these first test sets, it has the potential to have a marked improvement on the lives of patients undergoing screening for lung cancer. Dr. Maldonado hopes this new tool will serve as a supplementary diagnostic tool for classification of indeterminate pulmonary nodules, minimizing the use of invasive tests, improper treatments, and undue stress on patients.
Multivariate analysis using LASSO on all features yielded a multivariate model with 8 selected features (selected with frequency > 50% after introducing bootstrap to reduce variability after 1000 runs) with an AUC estimate of 0.941.
Peikert T, Duan F, Rajagopalan S., et al. 2018. Novel high-resolution computed tomography-based radiomic classifier for screen-identified pulmonary nodules in the National Lung Screening Trial. PloS One 13.5:e0196910.
Last updated Monday, January 3, 2022