Original Article
Quantitative radiomic model for predicting malignancy of small solid pulmonary nodules detected by low-dose CT screening
Abstract
Background: It is a permanent challenge to differentiate small solid lung nodules. Massive data, extracted from medical image through radiomics analysis, may help early diagnosis of lung cancer. The aim of this study was to assess the usefulness of a quantitative radiomic model developed from baseline low-dose computed tomography (LDCT) screening for the purpose of predicting malignancy in small solid pulmonary nodules (SSPNs).
Methods: This retrospective study included malignant and benign SSPNs (6 to 15 mm) detected in baseline low-dose CT screening. The malignancy was confirmed pathologically, and benignity was confirmed by long term follow-up or pathological diagnosis. The non-contrast CT images were reconstructed with a lung kernel of a slice thickness of 1 mm and were processed to extract 385 quantitative radiomic features using Analysis-Kinetic software. A predictive model was established with the training set of 156 benign and 40 malignant nodules, and was tested with the validation set of 77 benign and 21 malignant nodules through the analysis of R square. The performance of the radiomic model in predicting malignancy was compared with that of the ACR Lung Imaging Reporting and Data System (ACR lung-RADS).
Results: In 294 cases of SSPNs, 61 lung cancers and 24 benign nodules were confirmed pathologically and the remaining 209 nodules were stable with long-term follow-up (4.1±0.9 years). Eleven non-redundant features, including 8 high-order texture features, were extracted from the training data set. The sensitivity and specificity of the prediction model in malignancy differentiation were 81.0% and 92.2% respectively. The accuracy was superior to ACR-lung RADS (89.8% vs. 76.5%).
Conclusions: A radiomic model based on baseline low-dose CT screening for lung cancer can improve the accuracy in predicting malignancy of SSPNs.
Methods: This retrospective study included malignant and benign SSPNs (6 to 15 mm) detected in baseline low-dose CT screening. The malignancy was confirmed pathologically, and benignity was confirmed by long term follow-up or pathological diagnosis. The non-contrast CT images were reconstructed with a lung kernel of a slice thickness of 1 mm and were processed to extract 385 quantitative radiomic features using Analysis-Kinetic software. A predictive model was established with the training set of 156 benign and 40 malignant nodules, and was tested with the validation set of 77 benign and 21 malignant nodules through the analysis of R square. The performance of the radiomic model in predicting malignancy was compared with that of the ACR Lung Imaging Reporting and Data System (ACR lung-RADS).
Results: In 294 cases of SSPNs, 61 lung cancers and 24 benign nodules were confirmed pathologically and the remaining 209 nodules were stable with long-term follow-up (4.1±0.9 years). Eleven non-redundant features, including 8 high-order texture features, were extracted from the training data set. The sensitivity and specificity of the prediction model in malignancy differentiation were 81.0% and 92.2% respectively. The accuracy was superior to ACR-lung RADS (89.8% vs. 76.5%).
Conclusions: A radiomic model based on baseline low-dose CT screening for lung cancer can improve the accuracy in predicting malignancy of SSPNs.
