The Research Of Feature Comparison Analysis And Biological Activity In Hepatic Alveolar Echinococcosis By Radiomics

Objective: To analyze the imaging characteristics and classification of hepatic alveolar echinococcosis(HAE)by spectral CT and MRI,and to compare with histopathological indexes and 18FDG-PET/CT.To explore the effectiveness of spectral CT and MRI in evaluating the biological activity of HAE lesions,find the best imaging examination method,and establish the biological activity prediction model of hepatic alveolar echinococcus by using radiomics and artificial intelligence technology.Methods: 156 patients with HAE admitted to the First Affiliated Hospital of Xinjiang Medical University from January 2012 to June 2020 were collected.The diagnostic criteria were derived from the WHO Information Working Group on echinococcosis(WHO-IWGE),and 136 patients were enrolled.MRI and PET/CT were performed in all patients to obtain the maximum standardized uptake value(SUVmax).If the SUVmax of the lesion is higher than the uptake value of the normal liver parenchyma,it is defined as having biological activity;otherwise,if the uptake value is equal to or lower than the uptake value of the normal liver parenchyma,it is defined as having no biological activity.18 patients underwent spectrum CT scanning.CT value,iodine quantification,best CNR(contrast to noise ratio)value and curve slope were analyzed and measured.18 patients’ tissue sections were performed to HE,Masson and CD34 staining.The correlation between the iodine value of spectral CT and the SUVmax in the edge area of the lesion,the iodine value of spectral CT and the SUVmax in the edge area of the lesion was analyzed.To explore the correlation between Kodama classification and PET/CT results.Fleiss’ s kappa and Cohen’s kappa were used to test the inter observer consistency of MRI and PET/CT image classification.According to the results of PET/CT,136 cases were divided into training set(active group)and test set(inactive group).Through image segmentation,feature extraction and dimension reduction,the prediction model is established,and the effectiveness of the model is verified by the test set T2 WI image data.Python 3.6 is mainly used to implement standardization,feature selection and model construction.Spss20.0 was used for statistical analysis of clinical information.The test in the study was a two tailed test,P < 0.05 was considered statistically significant.Result: The mean iodine concentration in arterial phase,portal phase and venous phase of the lesion margin was 7.25± 3.80,28.40 ± 7.59 and 26.26 ± 6.74 respectively,which was different from that of solid components and normal liver parenchyma in the same period(P=0.000).The mean value of MVD in the edge of HAE lesions was 27.81 ± 7.17,which was statistically significant compared with the MVD values of the solid part(0.72 ± 1.17)and the normal liver parenchyma(4.24 ± 2.05).The correlation coefficients between iodine quantification and MVD were: arterial phase(r=0.029,P=0.909),portal phase(r=0.775,P=0.000)and venous phase(r=0.659,P=0.003).The correlation coefficients between SUVmax and iodine quantification in the edge of HAE lesions were as follows: arterial phase(r=0.644,P=0.000),portal phase(r=0.812,P=0.000)and venous phase(r=0.697,P=0.000).The results of PET/CT showed that 17(94.4%)cases had biological activity,and 16(88.9%)had enhancement pattern on the spectrum CT images.The two methods had good consistency in the evaluation of biological activity,kappa value was 0.364,P=0.546.Kodama classification of 136 cases: type 1,4/136(2.7%);type 2,33/136(24.3%);type 3,83/136(61.3%);type 4,12/136(8.9%);type 5,4/136(2.8%).According to the results of PET/CT,90 lesions were active,and the SUV values ranged from 4.4 to 22.9(average 11.3),which were mainly distributed in 4 of type 1(100%),27 of type 2(81.8%),53 of type 3(63.9%),5 of type 4(41.7%)and 1 of type 5(25.0%).On the basis of T2 WI image segmentation and extraction,the optimal features selected by radiomics method include first-order statistics feature(n=2)and texture feature(n=1),and first-order statistics feature(n=16)and texture feature(n=29)after filter transformation.The performance of the three classifier models in the training set is as follows: AUC of LR was 0.855 ± 0.025,accuracy was 0.806,sensitivity was 0.836,specificity was 0.775;AUC of MLP was 0.925 ± 0.057,accuracy was 0.886,sensitivity was 0.883,specificity was 0.889;AUC of SVM was 0.907 ± 0.037,accuracy was0.806,sensitivity was 0.836,specificity was 0.775.In the test set: AUC of LR was 0.809 ±0.046,accuracy was 0.794,sensitivity was 0.778,specificity was 0.811;AUC of MLP was0.830 ± 0.053,accuracy was 0.817,sensitivity was 0.822,specificity was 0.811;AUC of SVM was 0.804 ± 0.035,accuracy was 0.794,sensitivity was 0.778,specificity was 0.811.Conclusion: 1)Spectral CT imaging can provide more information about the tissue structure and blood supply of HAE lesions than conventional CT.The quantitative value of iodine in the edge of the lesion has a high correlation with the MVD value and SUVmax of the corresponding region,to a certain extent,which reflects the biological activity of HAE lesions.2)MRI images have more advantages in showing the fine structure of HAE lesions.Kodama classification can better reflect the pathological state and process of HAE lesions.Compared with PET/CT results,some of its classification can reflect the biological activity of HAE lesions.3)The prediction model of HAE biological activity established by MRI radiomics and artificial intelligence technology has a very good predictive value,which is similar to the results of PET/CT.This model may be a supplementary tool for monitoring and follow-up evaluation of the activity of echinococcosis in clinical work.