Predicting kidney tumor subtype from ct images using radiomics and clinical features

Purpose: This study aims to evaluate the performance of machine learning methods in predicting the subtype (clear-cell vs. non-clear-cell) of kidney tumors using clinical patient and radiomics data from CT images. Method: CT images of 192 malignant kidney tumor cases (142 clear-cell, 50 other) from TCIA’s KiTS-19 Challenge were used in the study. There were several different tumor subtypes in the other group, most of them being chromophobe or papillary RCC. Patient clinical data were combined with the radiomic features extracted from CT images. Features were extracted from 3D images and all of the slices were included in the feature extraction process. Initial dataset consisted of 1157 features of which 1130 were radiomics and 27 were clinical. Features were selected using Kruskal Wallis – ANOVA test followed by Lasso Regression. After feature selection, 8 radiomic features remained. None of the clinical features were considered important for our model as a result. Training set classes were balanced using SMOTE. Training data with the selected features were used to train the Coarse Gaussian SVM and Subspace Discriminant classifiers. Results: Coarse Gaussian SVM was faster compared to Subspace Discriminant with a training time of 0.47 sec and ~11000 obs/sec prediction speed. Training duration of Subspace Discriminant was 4.1 sec with ~960 obs/sec prediction speed. For Coarse Gaussian SVM; validation accuracy was 67,6% while the accuracy of test was 80%, with and AUC of 0.86. Similarly, Subspace Discriminant had 68,8% validation accuracy and 80% test accuracy; AUC was 0.85. Conclusion: Both models produced promising results on classifying malignant tumors as ccRCC or non-ccRCC. However, Coarse Gaussian SVM might be more preferable because of its training and prediction speed.