Ultrasound-based Radiomics Study For Differentiating Subcutaneous Tissue Hemangioma And Kaposiform Hemangioendothelioma

Background:Subcutaneous tissue Hemangioma reach about 5%in infants,which is a benign tumor and has good prognosis;Kaposiform hemangioendothelioma（KHE）is an invasive vascular tumor similar to sarcoma,infants and young children are more common with high mortality rate and poor prognosis^[1-2].Early accurate differential diagnosis of both diseases is critical for clinician decision making and improving patient outcomes.Radiomics is a research method for quantitative description and quantitative analysis of medical imaging,at present,ultrasound-based radiomics has been widely used at home and abroad,and shows great potential,but it is rare in superficial soft tissue tumors,this paper will identify subcutaneous HE from KHE by ultrasound-based radiomics.Objective:To evaluate the application value of ultrasound-based radiomics in distinguishing subcutaneous HE and KHE by constructing ultrasound-based radiomics model.Methods:252 Ultrasound images of 213 subcutaneous tissue HE or KHE patients from Henan Provincial People’s Hospital department of hemangioma from October 2014 to February 2022,which were confirmed either clinically or pathologically,were selected.Image features were extracted by using Pyradiomics,including morphology,first order,texture,wavelet and so on（951 in total）,The minimum absolute shrinkage and selection algorithm（LASSO）was used to select 22 stable features with non-zero coefficients after reducing the feature dimension.Support Vector Machine（SVM）were used to construct the radiomics model,and the combined model was established by combining clinical data and radiomics characteristics.The area under the ROC curve（AUC）,along with 95%confidence interval（95%CI）,accuracy,sensitivity,specificity,positive predictive value and negative predictive value were calculated to evaluate model’s efficiency,then the diagnostic efficacy of different models was compared.Results:A total of 951 features were extracted from the gray scale ultrasound image ROI,185 features were removed by t-test and Mann-Whitney U test,and the remaining 766 features were screened by dimension reduction,then 22 stable features with non-zero coefficients were selected.The SVM model were established by 22 selected features.The AUC（along with 95%CI）、accuracy、sensitivity、specificity、positive predictive value and negative predictive value,in the training group and the validation group,were0.91（0.89～0.93）、91.41%、83.20%、93.92%、95.79%、89.00%and 0.85（0.83～0.87）、90.78%、79.32%、97.90%、96.71%、88.68%.Statistical analysis of clinical data showed that the location of lesion distribution and platelet count were statistically different between HE and KHE（P<0.05）.The combined model was established by 2clinical features and 22 radiomics features.The AUC（along with 95%CI）、accuracy、sensitivity、specificity、positive predictive value and negative predictive value,in the training group and the validation group,were 0.94（0.92～0.96）、94.33%、90.77%、96.38%、94.23%、94.90%和0.90（0.88～0.92）、92.14%、85.69%、95.76%、93.33%、92.30%.Conclusions:1.Ultrasound-based radiomics can distinguish HE and KHE.2.The diagnostic efficiency of the combined model established by combining clinical data and ultrasound imaging data is better than that of the simple ultrasound-based radiomics model.