Diagnostic Study Of Nodular Pulmonary Cryptococcosis In CT Image Based Radiomic Features

Objective To evaluate the practical value of CT image based radiomic features in the diagnosis and differential diagnosis of nodular pulmonary cryptococcosis.Materials and methods Retrospective analysis was performed on44 cases of pulmonary cryptococcosis(PC),58 cases of Tuberculosis(TB)and60 cases of Lung Cancer(LC)diagnosed with pulmonary nodulesin CT images confirmed by surgical pathology.A total of 107 image histological features of 7categories were extracted with radcloud3.0 software.The most relevent histological features of prediction model were screened by Lassomethod.The prediction model of nodular pulmonary cryptococcosis was established by SVM.Then the model was tested with cross validation method.Finally,the area under the curve(AUC),sensitivity and specificity of the subject operating characteristic(ROC)were selected to evaluate the effectiveness of the prediction model.At the same time,CT features of nodules and clinical features in the three groups were observed and analyzed.P<0.05 was considered statistically significant,and the efficacy of statistically significant para-meters in the 3groups was evaluated by ROC.Finally,the individualized prediction models were further established combining imaging histological characteristics,clinical characteristics and CT morphological characteristics(diagnostic efficacy of medium or above),and were divided into individualized prediction model 1(PC group and TB group)and individualized prediction model 2(PC group and LC group).Results There were statistically significant differences between 3group in CT features of nodules and clinical features,such as gender,age,distribution,internal structure,margin,the relationship between nodules and bronchus or blood vessel,pleural depression,satellite lesions,aCER and vCER.In the PC group and TB group,the distribution,pleural depression and vCER had the better discrimination performance(AUC 0.754,0.712,0.786,respectively),diagnosis efficiency is better.In the PC group and LC group,the distribution,pleural depression,aCER and vCER had medium diagnostic efficacy(AUC 0.754,0.712,0.706,0.780,respectively),and the relationship between nodules and blood vessel had the highest diagnostic efficacy(AUC 0.877).After dimensionality reduction processing of all radiomic features,5 most relevant radiomic features were selected from the prediction model 1(PC group and non-pc group).In the validation set of model 1,the prediction accuracy was 0.49,AUC was 0.472,sensitivity was 0.11,and specificity was 0.88.In the prediction model 2(PC group and TB group),the 6most relevant radiomic features were selected.In the validation set of model2,the prediction accuracy was 0.80,AUC was 0.815,sensitivity was 0.78,and specificity was 0.83.In the prediction model 3(PC group and LC group),the 6most relevant radiomic features were selected.The accuracy,AUC,sensitivity and specificity of the validation set were 0.76,0.806,0.67 and 0.83 respectively.In PCgroup and TB group,an individualized prediction model 1 was established,combined with CT morphological features such as pulmonary lobe distribution and pleural depression sign and image histology.The accuracy,AUC,sensitivity and specificity of the validation set were 0.80,0.870,0.67 and 0.92,respectively.In the PC group and LC group,the individualized prediction model2 was established,combined with CT morphological characteristics and image histology,such as pulmonary lobe distribution,pleural depression sign,and the relationship between nodules and adjacent vessels.The accuracy of validation set was 0.90,AUC0.926,sensitivity was 0.78,and specificity was 1.00.Conclusion(1)the PC group lesions located in the lower lobe and more close to the pleura or pulmonary peripheral,the relationship between nodules and blood vessel was more performance for Type ? & ?,pleural indentation sign was relatively rare,mild to moderate degree of enhancement.(2)The prediction models based on radiomic features are feasible in the differential diagnosis of PC,TB and LC,and the diagnostic efficiency is higher than single clinical features and traditional CT features.And the individualized prediction models combined with image histological features and some traditional CT morphological features have the best diagnostic performance.