Abstract:
Renal cell carcinoma (RCC) constitutes %85 to %90 of all kidney malignancies. In 2020, 430,000 new cases were diagnosed and 179,000 of them lost their lives. Clear cell renal cell carcinoma (ccRCC) is the most common sub-type of RCC with approxi mately %80 occurrence rate. Accurate, non-invasive and preoperative determination of the International Society of Urological Pathology (ISUP) based tumor grade is impor tant for the e ective management of patients with ccRCC. Recent studies showed that CT radiomics can o er the means to predict this grade but there are some problems about data such as scarcity, unbalancing and standardization. In this study, we aimed to improve discrimination power between grades via using 3D and 2D radiomics fea tures and ensemble machine learning methods. Radiomics features were extracted from 143 CT images obtained from the publicly available data set from The Cancer Imaging Archive. Over sampling methods and series of feature selection methods were applied to reduce the number of features. Besides the actual tumor volume, 5 additional VOIs were created to consider peritumor regions and test the robustness of the model against variations in segmentation for three ensemble machine learning algorithms. The best result was found when SMOTE was used in combination with Light Gradient Boosting Method (LightGBM) AUC of 0.89 ± 0.02. As a result, ccRCC tumor grade can be pre dicted from 3D CT images with a high reliability despite the inadequacy of a dataset. The algorithm is moderately robust against deviations in segmentation by observers.|Keywords : Radiomics, WHO/ISUP Grade, Peritumor, ccRCC, Machine Learning.
