基于T2WI的纹理分析和机器学习在鉴别肾乏脂血管平滑肌脂肪瘤和肾癌中的价值

Objective: To distinguish between renal angiomyolipoma without visible fat (AMLwvf) and renal cell carcinoma (RCC) using T2WI texture analysis and machine learning. Materials and Methods: 80 cases of renal tumors were analyzed retrospectively, including AMLwvf (n=20), clear cell renal cell carcinoma (n=20), papillary renal cell carcinoma (n=20) and chromophobe renal cell carcinoma (n=20). Lesions were delineated on software by two radiologists to extract the corresponding volumes of interest (VOI) and then 93 features were generated. The Kruskal Wallis test showed that there was no significant difference between renal carcinoma subtypes, so renal carcinoma subtypes were combined into one group (renal carcinoma, n=60). Univariable analysis was carried out through Mann-Whitney U test and Holm-Bonferroni method to find the best features and analyze the diagnostic performance. Modeling with multiple features: after the primary selection of features by Pearson correlation coefficient, the C5.0 node of IBM SPSS modeler software calculated the relative importance ranking of features. Top 2, 3, 4 and 5 most important features were used to form 4 feature subsets. Decision tree C5.0 model was built with or without boosting. The differentiation and generalization ability of each model was evaluated to find the best one as the final model. Results: Univariable analysis: after Holm-Bonferroni correction, four different features were screened: minimum, 10 percentile, difference variance and contrast. The area under the curve was 0.888, 0.837, 0.789 and 0.777, respectively. The range of positive predictive value was 50.00%—69.57%. Modeling with multiple features: 8 decision tree C5.0 models were constructed. The area under the curve of final model was 0.950. The sensitivity, specificity, positive predictive value, negative predictive value and accuracy of final model were 90.00%, 100%, 100%, 96.77% and 97.5%, respectively. The accuracy based on cross validation is 95.0%. Conclusions: Univariable analysis based on T2WI has limited clinical application value because of its low positive predictive value. Decision tree C5.0 model has high accuracy and good generalization ability to distinguish AMLwvf and RCC, which is helpful to make reasonable treatment plan in clinic.