Analysis of CT Image Features of CcRCC on The Basis of Machine Learning: Differentiation of High-Grade from Low-Grade Fuhrman Nuclear Grades

Previous researches have clarified clinical applications of radiomics-based prediction of tumor phenotype. The purpose of our research is to utilize radiomic features in computer-aided diagnosis (CAD) system of prediction for high and low Fuhrman nuclear grades (FNG) in clear cell renal cell carcinoma (ccRCC). We selected 110 images from 109 cases of axial contrast- enhanced computed tomography with a pathological diagnosis of ccRCC from The Cancer Imaging Achieve portal. After preprocessing, extraction and selection of features, Weka Experiment Environment was run to compare performance of different classifiers to predict high and low grade FNG. The K- Nearest Neighbors classifiers input with 23 features (Sensitivity, specificity, accuracy, precision, recall and AUROC were 91%, 89%, 90.91%, 92%, 91% and 90%, respectively) showed significant better performance than other 11 classifiers. The developed CAD system generated by the machine-learning is a non-invasive and accurate method of predicting and monitoring FNG characteristics of ccRCC.