Predicting the Grade of Clear Cell Renal Cell Carcinoma from CT Images Using Random Subspace-KNN and Random Forest Classifiers

Abstract

Accurate and non-invasive determination of the International Society of Urological Pathology (ISUP) based tumor grade is important for the effective management of patients with clear cell renal cell carcinoma (cc-RCC). In this study, the radiomic analysis of 3D computed tomography (CT) images are used to determine ISUP grades of cc-RCC patients by exploring machine learning (ML) methods that can address small ISUP grade image datasets. 143 cc-RCC patient studies from The Cancer Imaging Archive (TCIA) USA were used in the study. 1133 radiomic features were extracted from the normalized 3D segmented CT images. Correlation coefficient analysis, Random Forest feature importance analysis and backward elimination methods were used consecutively to reduce the number of features. 15 out of 1133 features were selected. A k-nearest neighbors (KNN) classifier with random subspaces and a Random Forest classifier were implemented. Model performances were evaluated independently on the unused 20% of the original imbalanced data. ISUP grades were predicted by a KNN classifier under random subspaces with an accuracy of 90% and area under the curve (AUC) of 0.88 using the test data. Grades were predicted by a Random Forest classifier with an accuracy of 83% and AUC of 0.80 using the test data. In conclusion, ensemble classifiers can be used to predict the ISUP grade of cc-RCC tumors from CT images with sufficient reliability. Larger datasets and new types of features are currently being investigated.