摘要
Background Computer-aided diagnosis （CAD） of lung cancer is the subject of many current researches. Statistical methods and artificial neural networks have been applied to more quantitatively characterize solitary pulmonary nodules （SPNs）. In this study, we developed a CAD scheme based on an artificial neural network to distinguish malignant from benign SPNs on thin-section computed tomography （CT） images, and investigated how the CAD scheme can help radiologists with different levels of experience make diagnostic decisions. Methods Two hundred thin-section CT images of SPNs with proven diagnoses （135 small peripheral lung cancers and 65 benign nodules） were analyzed. Three clinical features and nine CT signs of each case were studied by radiologists, and the indices of qualitative diagnosis were quantified. One hundred and forty nodules were selected randomly to form training samples, on which the neural network model was built. The remaining 60 nodules, forming test samples, were presented to 9 radiologists with 3-20 years of clinical experience, accompanied by standard reference images. The radiologists were asked to determine whether a nodule was malignant or benign first without and then with CAD output. Diagnostic performance was evaluated by receiver operating characteristic （ROC） analysis. Results CAD outputs on test samples had higher agreement with pathological diagnoses （Kappa=0.841, P〈0.001）. Compared with diagnostic results without CAD output, the average area under the ROC curve with CAD output was 0.96 （P〈0.001） for junior radiologists, 0.94 （P=0.014） for secondary radiologists and 0.96 （P=0.221） for senior radiologists, respectively. The differences in diagnostic performance with CAD output among the three levels of radiologists were not statistically significant （P=0.584, 0.920 and 0.707, respectively）. Conclusions This CAD scheme based on an artificial neural network could improve diagnostic performance and assist radiologists in distinguishing malignant from benign SPNs on thin-section CT images.
Background Computer-aided diagnosis （CAD） of lung cancer is the subject of many current researches. Statistical methods and artificial neural networks have been applied to more quantitatively characterize solitary pulmonary nodules （SPNs）. In this study, we developed a CAD scheme based on an artificial neural network to distinguish malignant from benign SPNs on thin-section computed tomography （CT） images, and investigated how the CAD scheme can help radiologists with different levels of experience make diagnostic decisions. Methods Two hundred thin-section CT images of SPNs with proven diagnoses （135 small peripheral lung cancers and 65 benign nodules） were analyzed. Three clinical features and nine CT signs of each case were studied by radiologists, and the indices of qualitative diagnosis were quantified. One hundred and forty nodules were selected randomly to form training samples, on which the neural network model was built. The remaining 60 nodules, forming test samples, were presented to 9 radiologists with 3-20 years of clinical experience, accompanied by standard reference images. The radiologists were asked to determine whether a nodule was malignant or benign first without and then with CAD output. Diagnostic performance was evaluated by receiver operating characteristic （ROC） analysis. Results CAD outputs on test samples had higher agreement with pathological diagnoses （Kappa=0.841, P〈0.001）. Compared with diagnostic results without CAD output, the average area under the ROC curve with CAD output was 0.96 （P〈0.001） for junior radiologists, 0.94 （P=0.014） for secondary radiologists and 0.96 （P=0.221） for senior radiologists, respectively. The differences in diagnostic performance with CAD output among the three levels of radiologists were not statistically significant （P=0.584, 0.920 and 0.707, respectively）. Conclusions This CAD scheme based on an artificial neural network could improve diagnostic performance and assist radiologists in distinguishing malignant from benign SPNs on thin-section CT images.
基金
This work was supported by a grant from Beijing Natural Science Foundation(No.7062020).
