影像组学对磨玻璃结节型肺腺癌病理亚型的预测效能

Objective: To develop and validate the radiomics nomogram on the discrimination of lung invasive adenocarcinoma from 'non-invasive’ lesion manifesting as ground glass nodule (GGN) and compare it with morphological features and quantitative imaging. Methods: One hundred and sixty pathologically confirmed lung adenocarcinomas from November 2011 to December 2014 were included as primary cohort. Seventy-six lung adenocarcinomas from November 2014 to December 2015 were set as an independent validation cohort. Lasso regression analysis was used for feature selection and radiomics signature building. Radiomics score was calculated by the linear fusion of selected features. Multivariable logistic regression analysis was performed to develop models. The prediction performances were evaluated with ROC analysis and AUC, and the different prediction performance between different models and mean CT value were compared with Delong test. The generalization ability was evaluated with the leave-one-out cross-validation method. The performance of the nomogram was evaluated in terms of its calibration. The Hosmer-Lemeshow test was used to evaluate the significance between the predictive and observe values. Results: Four hundred and eighty-five 3D features were extracted and reduced to 2 features as the most important discriminators to build the radiomics signatures. The individualized prediction model was developed with age, radiomics signature, spiculation and pleural indentation, which had the best discrimination performance (AUC=0.934) in comparison with other models and mean CT value(P<0.05) and showed better performance compared with the clinical model (AUC=0.743, P<0.001). The radiomics-based nomogram demonstrated good calibration in the primary and validation cohort, and showed improved differential diagnosis performance with an AUC of 0.956 in the independent validation cohort. Conclusion: Individualized prediction model incorporating with age, radiomics signature, spiculation and pleural indentation, presenting with radiomics nomogram, could differentiate IAC from 'non-invasive’ lesion manifesting as GGN with the best performance in comparison with morphological features and quantitative imaging.