Preliminary Investigation On Radiomic Features And Its Clinical Significance In The Evaluation Of Pulmonary Ground-glass Opacity Nodules

Purpose:To establish radiomic models to distinguish between benign and early-stage lung adenocarcinoma lesions,pre-invasive and invasive lung adenocarcinoma lesions,invasive lung adenocarcinoma with different pathological types and lung adenocarcinoma with different drive gene mutation status manifesting as ground-glass opacity nodules（GGOs）.By comparing the established radiomic models with clinical-semantic models and quantitative computerized tomography（CT）models,this study aimed to explore the diagnostic efficacy and clinical application value of radiomics in the mentioned aspects.Methods:1)From January 2018 to April 2021,1269 patients（476 males and 793 females,with age ranging from 18 to 79 years old）with GGOs who were diagnosed clinically,had definite pathological results and had complete clinical and CT images with thin slice thickness in Chinese People’s Liberation Army General Hospital were enrolled in this study.The patients were then divided into a training group and a validation group according to the specific research field.Data in training group was used for model construction,while data in validation group was only used for model validation.The clinical data of the patients was searched and collected through our electronic medical record system,and CT semantic data was evaluated and recorded by respiratory physicians.2)2446 radiomic features from internal area within GGOs and 5mm ring area surrounding GGOs were acquired using 3D slicer software,and 27 CT quantitative features were obtained using FACT software from DEXIN company.In the training group,univariate analysis and least absolute shrinkage and selection operator（LASSO）were used for radiomic feature selection,and the selected features were then fitted into a radiomic model by four different machine learning methods,namely logistic regression（LR）,decision tree（DT）,support vector machine（SVM）and adaptive boosting（AB）,respectively.Clinical-semantic model and quantitative model were established using LR.A combined model and nomogram were then constructed by integrating the established radiomic,clinical-semantic and quantitative model.3)In the validation group,the established models were evaluated using area under the curve（AUC）of the receiver operating characteristic curve（ROC curve）as the primary evaluation index and accuracy（ACC）,sensitivity（SEN）,specificity（SPE）and F1 score as secondary evaluation indexes.Decision curve analysis（DCA）was accomplished to evaluate the clinical application value of the established models.R software（version 4.0.2）and SPSS software（version 22.0）were used for statistical analysis and model construction.Results:1)In this part of the research,131 benign lesions and 215 early-stage adenocarcinoma lesions were enrolled.Among the clinical and semantic parameters,smoking status,history of malignancy except lung cancer（yes/no）and spiculation sign（with/without）were identified as independent risk factors by multivariate logistic regression.The calcification volume was identified as the independent risk factor among the quantitative parameters.12 radiomic features were selected and fitted into a radiomic model based on SVM method,which exhibited better diagnostic efficacy（AUC=0.851）in the validation group than clinical-semantic model（AUC=0.506）and quantitative model（AUC=0.520）.Furthermore,the radiomic model showed the best clinical application value.2)139 pre-invasive lesions（including atypical adenomatous hyperplasia and Adenocarcinoma in situ）and 879 invasive lesions（including minimally invasive adenocarcinoma and invasive adenocarcinoma）were enrolled in this research.In the training group,abnormal lung cancer related tumor biomarkers（with/without）,nodule density（pure GGO or mixed GGO）,lobulation sign（with/without）and maximum diameter of the nodule were identified as independent risk factors among clinical and semantic parameters,while surface to volume ratio and the 90^th percentile of density were identified as independent risk factors among CT quantitative parameters.16 radiomic features were selected by LASSO regression and fitted into a LR-based radiomic model.In the validation group,the radiomic model showed an AUC of 0.828,which was significantly higher than that of the clinical-semantic model（AUC=0.746）.The radiomic model using inter-nodular features alone also exhibited good performance（AUC=0.808）.Moreover,the combined model which integrated radiomic,clinical-semantic and quantitative parameters showed an AUC of 0.875and exhibited the best clinical application value.3)According to the pathological subtypes of invasive adenocarcinoma,the GGOs were divided into lepidic predominant adenocarcinoma（LPA group）and non-lepidic predominant adenocarcinoma（non-LPA group）.After univariate analysis and LR regression,nodule density（pure GGO or mixed GGO）was incorporated in the clinical-semantic model,while surface to volume ratio and the 75^th percentile of density were included in the quantitative model.SVM method was used to establish the radiomic model,which exhibited better efficacy（AUC=0.692）in the validation group than the clinical-semantic model（AUC=0.556）.The quantitative model showed similar diagnostic efficacy（AUC=0.679）to the radiomic model.Best clinical application value was witnessed using the radiomic model rather than the clinical-semantic model or the quantitative model.4)109 lung adenocarcinoma patients with negative Epidermal Growth Factor Receptor（EGFR）mutation and 119 patients with positive EGFR mutation were included in this study.Sex（male/female）,family history of malignancy except lung cancer（yes/no）and abnormal lung cancer related tumor biomarkers（with/without）were found to be independent risk factors for EGFR status,however,all the quantitative parameters failed to show significant difference between the EGFR（+）and EGFR（-）group.The radiomic model established by LR exhibited better diagnostic efficiency（AUC=0.637）than the clinical-semantic model（AUC=0.585）in the validation group.The combined model which integrated the radiomic model and the clinical-semantic model exhibited best diagnostic efficacy（AUC=0.653 in the validation group）and clinical application value.5)654 anaplastic lymphoma kinase（ALK）mutation-negative patients and 101ALK-positive patients were enrolled in this study.In the training group,symptom（with/without）,spiculation sign（with/without）and maximum diameter of the nodule were incorporated in the clinical-semantic model,while ratio of the non-solid component and the10^th percentile of density were incorporated in the quantitative model.The radiomic model established by LR method showed an AUC of 0.736 in the validation group,comparing to0.682 of the clinical-semantic model and 0.584 of the quantitative model.Furthermore,the radiomic model also exhibited better clinical application value.The combined model,which integrated the radiomic model,the clinical-semantic model and the quantitative model,failed to show a higher diagnostic efficacy than the radiomic model alone（AUC=0.712 in the validation group）.Conclusions:1)Radiomics exhibited excellent diagnostic efficacy in distinguishing between benign and early-stage adenocarcinoma lesions,pre-invasive and invasive adenocarcinoma lesions and lesions with different ALK mutation status manifesting as GGOs.Radiomics showed good diagnostic efficacy in differentiating between LPA and non-LPA lesions,and lung adenocarcinoma with different EGFR mutation status,however,it had not reached the clinical application level.Nevertheless,radiomics showed better diagnostic efficacy and clinical application value than clinical and semantic parameters in all aspects mentioned above.2)Quantitative features had similar diagnostic abilities to radiomic features with respect to distinguishing pre-invasive lesions from invasive ones,and LPA from non-LPA lesions,however,they showed weak diagnostic efficacy in distinguishing between benign and early-stage adenocarcinoma lesions and lung adenocarcinoma with different EGFR and ALK mutation status.These results indicated a limited application range for quantitative features.3)The inter-nodular radiomic features extracted from the 5mm ring area surrounding the nodules showed good diagnostic efficacy in all aspects of this study.To summarize,radiomics exhibited good diagnostic efficacy in differentiating between benign and early-stage adenocarcinoma lesions,pre-invasive and invasive lesions,LPA and non-LPA lesions and lung adenocarcinoma with different drive mutation status manifesting as GGOs,indicating that it had the potential for future clinical utility.