Comparison Between Radiological Semantic Features and Lung‐RADS in Predicting Malignancy of Screen‐Detected Lung Nodules in the National Lung Screening Trial

Micro‐Abstract In this study, we investigated the predictive value of radiological semantic features and lung‐RADS in pulmonary nodules malignancy risk at 3 screening rounds. We obtained 199 patients (139 nodule‐positive controls and 60 incident lung cancers) from the National Lung Screening Trial. It was found that semantic features outperformed lung‐RADS at baseline and were comparable to lung‐RADS at subsequent scans. We find addition of semantic features to lung‐RADS improves malignancy risk prediction. Rationale: Lung computed tomography (CT) Screening Reporting and Data System (lung‐RADS) has standardized follow‐up and management decisions in lung cancer screening. To date, little is known how lung‐RADS classification compares with radiological semantic features in risk prediction and diagnostic discrimination. Objectives: To compare the performance of radiological semantic features and lung‐RADS in predicting nodule malignancy in lung cancer screening. Methods: We used data and low‐dose CT (LDCT) images from the National Lung Screening Trial (NLST). The training cohort contained 60 patients with screen‐detected incident lung cancers who had a positive baseline screen (T0) that was not diagnosed and then was diagnosed at second follow‐up (T2), and 139 nodule‐positive controls who had 3 consecutive positive screens (T0 to T2) that were not diagnosed as lung cancer. The testing cohort included 40 patients with incident lung cancers that were diagnosed at first follow‐up (T1) and 40 nodule‐positive controls. Twenty‐four semantic features were scored on a point scale from the LDCT images. Multivariable linear predictor model was built on the semantic features and the performances were compared with lung‐RADS in 3 screening rounds. We also combined non–size‐based semantic features with lung‐RADS to improve malignancy detection. Results: At T0, the average area under the receiver operating characteristic curve (AUROC) for border definition in risk prediction was 0.72. The average AUROC for contour at T1 in risk prediction and T2 in diagnostic discrimination was 0.82 and 0.88, respectively. By comparison, the average AUROC of lung‐RADS at T0, T1 and T2 were 0.60, 0.76 and 0.87, respectively. The combined model of the semantic features and lung‐RADS shows improvement with AUROCs of 0.74, 0.88 and 0.96 at T0, T1, and T2, respectively, achieved by adding border definition (at T0) or contour (at T1 and T2). Conclusion: We find semantic features defined by border definition and contour performed similar to lung‐RADS at follow‐up time point and outperformed lung‐RADS at baseline. These semantics alongside of lung‐RADS shows improved performance to detect malignancy.