| Objective:Retrospectively analyzed the accuracy of intraoperative frozen section(FS)manifesting as ?1.0cm pulmonary ground-glass nodules(GGN)in CT and compared this accuracy with the accurary of based on CT Radiomics features model to predicte invasive adenocarcinoma(IA),in order to explore the feasibility of Radiomics replace FS to guide the extent of surgical resection.Methods:A total of 201 GGN cases with diameter ?1.0cm were retrospectively collected from January 2014 to June 2017 in local two tertiary hospitals for preoperative thoracic thin-section CT examination and intraoperative FS pathological diagnosis.The "gold standard" for all cases was based on the results of paraffin sections including benign lesions(BL),atypical adenomatous hyperplasia(AAH),adenocarcinoma in situ(AIS),minimally invasive adenocarcinoma(MIA)and IA.Among them,BL,AAH,AIS and MIA were classified as non-IA group,the overall consistency between FS and postoperative paraffin pathological diagnosis and the accuracy of IA diagnosis were analyzed and compared.112 cases from the first hospital were used as the training cohort to propose Radiomics model and 89 cases from the second hospital were used as the validation cohort to verify the accuracy of Radiomics model.A total of 485 features were extracted including the density,morphological,texture and wavelet features of GGN lesions based on CT images.Boruta method was used to select six of the most important features.SVM was used as a classifier to propose the Radiomics model and verify the accuracy of Radiomics model.Application of receiver operating characteristic(ROC)curve to assess the predicted or diagnostic efficacy of model and FS,P<0.05 was considered statistically significant.Results: 1.The accuracy of FS: in all 201 GNN cases,there were 60(29.9%)cases of inconsistency and the accuracy rate of FS was 70.1%.Among the cases inconsistency,14(23.3%)patients with IA were misdiagnosed as non-IA by FS,21(35.0%)patients with non-IA were overdiagnosed as IA by FS.In addition,11 cases of MIA were misdiagnosed as AIS,8 cases of AIS were misdiagnosed as AAH,3 cases of AIS were overdiagnosed as MIA,and 2 cases of AIS and 1 case of MIA were misdiagnosed as BL.The accuracy of IA diagnosis by FS was 82.6%.2.Radiomics features extraction: Among the extracted features,the number of features consistency >0.75 were 340.Six of the most important features are selected by Boruta's algorithm,including two mass features,one uniformity feature and three gray levels Co-occurrence matrix(GLCM)features.3.Prediction IA of Radiomics Model vs FS: The radiomics model showed an accuracy,sensitivity,and specificity for predicting IA were 0.857,0.862,and 0.855 in the primary cohort,respectively,while they were 0.888,0.909,and 0.875 in the validation cohort,respectively.The FS showed an accuracy,sensitivity,and specificity for determining IA were 0.810,0.724,and 0.843 in the primary cohort,respectively,while they were 0.843,0.818,0.857 and 0.810 in the validation cohort,respectively.ROC curve analysis showed that radiomics model yielded a superior area under curve(AUC)in the primary cohort(0.910(95% CI,0.846~0.973))and validation cohort(0.912(95% CI,0.875~0.981))than FS in the primary cohort 0.784(95% CI,0.692~0.875)and validation cohort 0.810(95% CI,0.728~0.893).Conclusion: 1.For GGN lesions ?1.0 cm in diameter,there is still a certain misdiagnosis rate of FS pathology in the evaluation and diagnosis of non-IA and IA,which tends to be overdiagnosed,and there seems to be room for improvement of the accuracy of FS.2.Radiomics features reflecting the mass and uniformity extracted from CT images is a feasible,stable and effective method to preoperatively predicte the ?1.0 cm GGN lesions for IA.3.The accuracy of CT-based Radiomics model to predicte the ?1.0 cm GGN lesions for invasive adenocarcinoma of is better than FS,which indicated Radiomics is expected to partially replace FS to become an important reference for preoperative surgical method of ?1.0 cm GGN. |
PDF Full Text Request