Radiomic-Based Quantitative CT Analysis of Pure Ground-Glass Nodules to Predict the Invasiveness of Lung Adenocarcinoma.

Objectives:To investigate the performance of radiomic-based quantitative analysis on CT images in predicting invasiveness of lung adenocarcinoma manifesting as pure ground-glass nodules (pGGNs). Methods:A total of 275 lung adenocarcinoma cases, with 322 pGGNs resected surgically and confirmed pathologically, from January 2015 to October 2017 were enrolled in this retrospective study. All nodules were split into training and test cohorts randomly with a ratio of 4:1 to establish models to predict between pGGN-like adenocarcinomain situ(AIS)/minimally invasive adenocarcinoma (MIA) and invasive adenocarcinoma (IVA). Radiomic feature extraction was performed using Pyradiomics with semi-automatically segmented tumor regions on CT scans that were contoured with an in-house plugin for 3D-Slicer. Random forest (RF) and support vector machine (SVM) were used for feature selection and predictive model building in the training cohort. Three different predictive models containing conventional, radiomic, and combined models were built on the basis of the selected clinical, radiological, and radiomic features. The predictive performance of each model was evaluated through the receiver operating characteristic curve (ROC) and the area under the curve (AUC). The predictive performance of two radiologists (A and B) and our radiomic predictive model were further investigated in the test cohort to see if radiomic predictive model could improve radiologists' performance in prediction between pGGN-like AIS/MIA and IVA. Results:Among 322 nodules, 48 (14.9%) were AIS and 102 (31.7%) were MIA with 172 (53.4%) for IVA. Age, diameter, density, and nine meaningful radiomic features were selected for model building in the training cohort. Three predictive models showed good performance in prediction between pGGN-like AIS/MIA and IVA (AUC > 0.8,P< 0.05) in both training and test cohorts. The AUC values in the test cohort were 0.824 (95% CI, 0.723-0.924), 0.833 (95% CI, 0.733-0.934), and 0.848 (95% CI, 0.750-0.946) for conventional, radiomic, and combined models, respectively. The predictive accuracy was 73.44 and 59.38% for radiologist A and radiologist B in the test cohort and was improved dramatically to 79.69 and 75.00% with the aid of our radiomic predictive model. Conclusion:The predictive models built in our study showed good predictive power with good accuracy and sensitivity, which provided a non-invasive, convenient, economic, and repeatable way for the prediction between IVA and AIS/MIA representing as pGGNs. The radiomic predictive model outperformed two radiologists in predicting pGGN-like AIS/MIA and IVA, and could significantly improve the predictive performance of the two radiologists, especially radiologist B with less experience in medical imaging diagnosis. The selected radiomic features in our research did not provide more useful information to improve the combined predictive model's performance.