Differentiation of renal angiomyolipoma without visible fat from renal cell carcinoma by machine learning based on whole-tumor computed tomography texture features.

Abstract

BackgroundMorphological findings showed poor accuracy in differentiating angiomyolipoma without visible fat (AMLwvf) from renal cell carcinoma (RCC). PurposeTo determine the performance of a machine learning classifier in differentiating AMLwvf from different subtypes of RCC based on whole-tumor slices of CT images. Material and MethodsIn this retrospective study, 171 pathologically proven renal masses were collected from a single institution. Texture features were extracted from whole-tumor images in three phases including the pre-contrast (PCP), corticomedullary (CMP), and nephrographic (NP) phases. A support vector machine with the recursive feature elimination method based on fivefold cross-validation (SVM-RFECV) with the synthetic minority oversampling technique (SMOTE) was utilized to establish classifiers for differentiating AMLwvf from all subtypes of RCC (all-RCC), clear cell RCC (ccRCC), and non-ccRCC. The performances of the classifiers based on three-phase and single-phase images were compared with each other and morphological interpretations. ResultsA machine learning classifier achieved the best performance in differentiating AMLwvf from all-RCC, ccRCC, and non-ccRCC. The performance of the best machine learning classifier for differentiating AMLwvf from all-RCC (area under the curve [AUC] = 0.96) and ccRCC (AUC = 0.97) was higher than that for differentiating AMLwvf from non-ccRCC (AUC = 0.89); morphological interpretations achieved lower performance for differentiating AMLwvf from all-RCC (AUC = 0.67), ccRCC (AUC = 0.68), and non-ccRCC (AUC = 0.64). ConclusionMachine learning can be a useful non-invasive technique for differentiating AMLwvf from all-RCC, ccRCC, and non-ccRCC, and it can be more accurate than morphological interpretation by radiologists.