Table 3.
Characteristics and Performance of Radiomic Models Predicting Malignancy in Pulmonary Nodules
| Publication | Training Population | Radiomics Technique | Feature Descriptions | External Validation |
Performance | Comments |
|---|---|---|---|---|---|---|
| Massion et al,78 2020 | NLST data set 14,761 CT scans (5,972 patients) with benign nodules, 932 CT scans (575 patients) with malignant nodules |
Deep learning- LCP-CNN, developed by Optellum | Unknown/not reported (Dense Convolutional Network) Clinical variables excluded from model |
Vanderbilt: incidentally detected pulmonary nodules: 52 benign, 64 malignant Oxford: incidentally detected pulmonary nodules: 400 benign, 63 malignant |
Internal validation: AUC = 0.921 vs Brock model, AUC = 0.856 External validation: Vanderbilt: LCP-CNN: AUC = 0.835 vs Mayo Clinic model AUC = 0.781 Oxford: LCP-CNN: AUC = 0.919 vs Mayo Clinic model AUC = 0.819 |
Optellum received FDA clearance for commercial use of its Virtual Nodule Clinic LCP-CNN showed superiority in net reclassification of benign and malignant nodules over Mayo Clinic model in both validation sets |
| Ardila et al,79 2019 | NLST data set 42,290 CT scans (14,851 patients) with and without nodules (70% used for training, 15% for tuning, 15% for testing); 638 malignant nodules |
Deep learning- Convolutional Neural Network, developed by Google Inc. | 1,024 learned features | Northwestern: screening CT scans: 1,139 CT scans (907 patients); 27 malignant nodules | Internal (training): AUC = 0.944 (Sens 83%, Spec 95%) External validation: AUC = 0.955 (Sens 84%, Spec 96%) |
Model showed improved performance in predicting malignancy in 1 year over retrospective Lung-RADS criteria when prior imaging was not available Similar performance to applied Lung-RADS when prior imaging was available |
| Peikert et al,80 2018 | NLST data set 318 benign nodules, 408 malignant nodules |
Conventional radiomic model (BRODERS) developed at the Mayo Clinic | Eight features selected by LASSO multivariate modeling from 57 features Notable variable categories include nodule location, nodule shape, nodule surface characteristics, and texture analysis |
Vanderbilt (Maldonado et al, 202181) incidentally detected pulmonary nodules: 79 benign nodules, 91 malignant nodules | Internal validation: BRODERS model AUC = 0.939 External validation: AUC = 0.900 (Sens 92%, Spec 62%) vs Brock model AUC = 0.870 |
Modest improvement over Brock model Brock model (clinical risk calculator developed on a screen-detected population) was applied to an external validation set with high rate of malignancy |
| Lv et al,82 2021 | NLST data set and Jinling Hospital (incidentally detected nodules) 1,078 benign nodules, 1,028 malignant nodules |
Deep learning- Filter-guided pyramid network | Unknown/not reported | Incidentally detected pulmonary nodules from three different hospitals 80 benign nodules, 261 malignant nodules |
External validation: AUC = 0.847 (Sens 85%, Spec 68%) | Performance of filter-guided pyramid network was similar to that of a panel of radiologists Established the utility of coupling deep learning strategies with human review for indeterminate pulmonary nodules |
AUC = area under the curve; BRODERS = Benign Versus Aggressive Nodule Evaluation using Radiomic Stratification; FDA = US Food and Drug Administration; LASSO = Least Absolute Shrinkage and Selection Operator; LCP CNN = Lung Cancer Predictor Convolutional Neural Network; NLST = National Lung Screening Trial.