| Part 1Differentiating minimally invasive and invasive adenocarcinomas in pure ground-glass nodules larger than 10 mm Objective:To explore the possibility of computed tomography(CT)quantitative analysis combined with radiomics to accurately identify the two main types of lung adenocarcinoma: minimally invasive adenocarcinoma(MIA)and invasive adenocarcinoma(IA).Materials and methods:Data were retrospectively collected from pGGN patients with postoperative pathologically confirmed MIA or IA and lesions ≥10 mm in diameter on preoperative CT images.In the training cohort,patients’ clinical data,CT semantic and quantitative features were analyzed using univariate and multivariate regression,and a standard CT model was established.The lesions were manually segmented,and the radiomics features were extracted and screened.Then the radiomics models were built by six machine learning methods to discriminate between MIA and IA.The performance of the standard CT model,the radiomics model and the combined model combining the standard CT model with the best radiomics model was evaluated using the area under receiver operating characteristic curve(AUC).Four experts were recruited to evaluate the pGGNs in the validation cohort separately,and the diagnostic performance of the experts and models were compared.Results:Two hundred pGGNs from 198 patients were included in this study.Shape and quantitative features included volume and mean CT value of the largest cross-section were used to build the standard CT model.In the training cohort,10 radiomics features were identified as the optimal feature subset,and the performance of radiomics models built by six machine learning methods was higher than that of the standard CT model.In the validation cohort,the logistic regression model performed consistently and had higher performance(AUC=0.877)than the standard CT model(AUC=0.828),which produced Radiomics＿score as an independent predictor of IA after multivariate analysis.The combined model did not show a significant improvement in performance over the radiomics model.The radiomics model performed better in distinguishing between IA and MIA compared to the diagnosis of four experts.Conclusions:For pGGN larger than 10 mm,the radiomics model performed excellent in distinguishing IA from MIA,which helps clinicians in treatment plan selection and preoperative evaluation.Part 2Identifying lung adenocarcinomas with predominant lepidic growth presenting as pure ground-glass nodules larger than 10 mm Objective:To explore the value of quantitative CT features and radiomics the diagnosis of the lung adenocarcinomas with predominant lepidic growth in pure ground-glass nodules(pGGNs)larger than 10 mm.Materials and methods:Retrospective analysis of CT images of 204 patients with large-size(≥10 mm)pGGN with postoperative pathological diagnosis of lung adenocarcinomas with predominant lepidic growth,including minimally invasive adenocarcinoma(MIA)and lepidic predominant invasive adenocarcinoma(LPA),or non-lepidic predominant invasive adenocarcinoma(NLPA).The pGGN in both groups(MIA/LPA and NLPA)were randomly divided into the training and validation cohorts.Patients’ clinical data and CT semantic and quantitative parameters in the training cohort were analyzed to build a baseline model;and a radiomics model with optimal radiomics features was built;finally,a combined model was built using best radiomics signature and baseline independent predictors.The performance of the three models and the differential diagnostic ability of the three radiologists were evaluated and compared,and external validation was performed using a test cohort of 47 patients from other hospitals.Results:The radiomics model built by the logistic regression method(AUC=0.833 in the training cohort;AUC=0.804 in the validation cohort;AUC=0.792 in the external test cohort)and the combined model(AUC values of 0.849,0.820,and 0.775,respectively)had better differential diagnostic ability than the baseline model including two features of tumor location and mean CT value(AUC values of 0.756,0.762,and 0.725,respectively).The DeLong test showed that the AUCs of the combined model and the radiomics model increased significantly.The highest diagnostic performance for radiologists was 0.600.Conclusions:The application of CT radiomics improved the identification performance of lung adenocarcinomas with predominant lepidic growth appearing as GGNs larger than 10 mm.Part 3Comparative study of CT images and pathological findings in lung adenocarcinomas manifesting as mixed ground-glass nodules larger than 10 mm Objective:To predict the histological aggressiveness of mixed ground-glass nodules(mGGN)-like lung adenocarcinoma larger than 10 mm by preoperative CT quantitative analysis combined with radiomics.Materials and Methods:This retrospective study included 112 patients who underwent surgical resection between November 2011 and December 2018 and were diagnosed with minimally invasive adenocarcinomas(MIAs)or invasive adenocarcinomas(IAs)with lesions presenting as mGGNs ≥10 mm on CT images.Patients’ CT images were analyzed for semantic features,quantitative features and radiomics,and then baseline model,radiomics model and combined model were established using logistic regression method.Results:In the training cohort,the baseline model(lobulation sign,mass),the radiomics model and the combined model obtained AUC values of 0.792,0.822 and 0.843,respectively,while the three models in the validation cohort obtained AUC values of 0.779,0.786 and 0.806,respectively.The nomogram showed that rad＿score was the only independent predictor compared to the two baseline features.Conclusions:For mGGN larger than 10 mm,radiomics can still provide valuable information about the invasiveness of lung adenocarcinomas.Although quantitative features are also efficient for pathological type prediction to some extent,combining radiomics often results in higher classification performance and clinical benefit,enabling accurate diagnosis of lung large-size ground-glass nodules and helping physicians in medical decision making. |
PDF Full Text Request