| Objective:With the popularity of low-dose computed tomography（CT）and lung cancer screening,the diagnosis of ground-glass nodules（GGO）in clinical practice has been increasing,and has become an important clinical problem in oncology.It has been reported that the cancer incidence of GGO is as high as 63%,so the differentiation of benign and malignant GGO is of great significance for prolonging the survival time of tumor patients[1].Deep learning is a rapidly developing new discipline,which can provide certain help for clinical diagnosis and treatment of patients through quantitative analysis and feature extraction of patients’CT images.The aim of this study was to develop a computer-aided diagnosis（CAD）protocol to classify ground glass nodules as benign and malignant,and to integrate the clinical characteristics of patients,serum early antibodies,and thymidine kinase 1characteristics to improve the accuracy of the model.Methods:Retrospective analysis was performed on 478 patients with pulmonary nodules who underwent surgical resection in the Department of Thoracic Surgery of the First Affiliated Hospital of China Medical University from July 2014 to April2018.The preoperative CT images of the patients were automatically segmented and features extracted using Inception V4 as the main framework to establish a model for predicting GGO subtypes.In addition,173 GGO patients admitted to the Department of Oncology of the First Affiliated Hospital of China Medical University from June2018 to November 2020 were included in this study.Serological tests of tumor markers CEA,serum antibody 7 and thymidine kinase 1 were performed in all patients.Finally,the predictive probabilities of serum seven antibodies and thymidine kinase 1,which have diagnostic value for early lung cancer,were fused with the predictive probabilities of the model by information fusion method to obtain the final model.In addition,this study conducted a bibliometric analysis of the Web of Science Core Collection（WOSCC）articles on pulmonary nodules over the past 10 years,using Vos Viewer 1.6.15 to identify the characteristics of authors,institutions,and journals.The Vos Viewer is used to generate network visualization,and the co-occurrence analysis is conducted with all keywords as a unit.The data is imported into the literature online econometric analysis platform（http://bibliometric.com/）to visualize the international cooperation between countries.Results:The 478 images of pulmonary nodules were randomly divided into the training set and the validation set at an 8:2 rate.The sensitivity and specificity of the model constructed through the Inception V4 network for the prediction of GGO subtypes in the training set were close to 100%,the area under the ROC curve was0.955,and the sensitivity and specificity in the validation set were 0.588 and 0.741,respectively.The area under the ROC curve was 0.705,and the accuracy of the prediction of the subtype category with the highest prediction probability was 70%.Among the subtypes of the top3 prediction probabilities,the accuracy consistent with the actual results is about 95%.In addition,the sensitivity,specificity and area under the ROC curve of CEA,serum seven antibody and TK1 were calculated respectively in this study.The sensitivity and specificity of TK1 were the highest among the three,which were 37%and 96%,respectively.The positive predictive value and negative predictive value were 0.939 and 0.471,respectively.When the Inception V4 model was fused with serum seven antibodies and TK1,the sensitivity and specificity of the final model were improved to 64.7%,82%and 0.791（95%CI:[0.715,0.867],P=0.002）.Conclusion:The prediction models of different subtypes of GGO were constructed based on the pulmonary CT image features of patients before surgery,which can provide certain help for clinical diagnosis and treatment of patients.Moreover,the model after integrating serological features can further improve the efficacy of benign and malignant classification of GGO. |
PDF Full Text Request