Predicting common solid renal tumors using machine learning models of classification of radiologist-assessed magnetic resonance characteristics

Purpose: Solid renal masses (SRM) are difficult to differentiate based on standard MR features. The purpose of this study was to assess MR imaging features of SRM to evaluate performance of ensemble methods of classifying SRM subtypes. Materials and methods: MR images of SRM (n = 330) were retrospectively evaluated for standard and multiparametric (mp) features. Models of MR features for predicting malignant and benign lesions as well as subtyping SRM were developed using a training dataset and performance was evaluated in a test data-set using recursive partitioning (RP), gradient booting machine (GBM), and random forest (RF) methods. Results: In the test dataset, GBM and RF models demonstrated an accuracy of 86% (95% CI 75% to 93%) for predicting benign versus malignant SRM compared to 83% (95% CI 71% to 91%) for the RP model. RF had the greatest accuracy in predicting SRM subtypes, 81.2% (95% CI 69.5% to 89.9%) compared with GBM 73.4% (95% CI 60.9% to 83.7%) or RP 70.3% (95% CI 57.6% to 81.1%). Marginal homogeneity was reduced by the RF model compared with the RP model (P < 0.001), but not the GBM model (P = 0.135). All models had high sensitivity and specificity for clear cell and papillary renal cell carcinomas (RCC), but performed less well in differentiating chromophobe RCC, oncocytomas, and fat-poor angiomyolipomas. Conclusion: Ensemble methods for prediction of SRM from radiologist-assessed image characteristics have high accuracy for distinguishing benign and malignant lesions. SRM subtype classification is limited by the ability to categorize chromophobe RCCs, oncocytomas, and fat-poor angiomyolipomas.