Discrimination of clear cell and non-clear cell renal cell carcinomas using computerized tomography radiomics features and machine learning methods

ABSTRACT Mert Can ÖZMAN, Discrimination of Clear Cell and Non-Clear Cell Renal Cell Carcinomas Using Computerized Tomography Radiomics Features and Machine Learning Methods, Zonguldak Bülent Ecevit University Faculty of Medicine, Department of Radiology, Department of Radiology, Zonguldak 2025. Objective: This study aimed to perform radiomics analysis on lesions detected in pre-operative CT images of patients diagnosed with Renal Cell Carcinoma (RCC), to extract features, and to investigate characteristics that differentiate Clear Cell Renal Cell Carcinoma (ccRCC) from non-Clear Cell Renal Cell Carcinoma (non-ccRCC) using machine learning algorithms. Materials and Methods: Patients diagnosed with RCC between 2012 and 2023 at Zonguldak Bülent Ecevit University Faculty of Medicine Hospital, who had CT scans, were retrospectively reviewed through the Hospital Information System A total of 70 cases with CT images of suitable quality and parameters for radiomics analysis were included. Histopathological findings were recorded from pathology reports, and CT images were retrieved from the hospital's PACS (Picture Archiving and Communication System). Radiological findings were assessed and collected for statistical analysis. Manual 3D volumetric segmentation was performed to extract features from the lesions. Radiological findings were compared based on histological subtypes. Extracted features were categorized into shape, histogram, and texture characteristics. Feature selection and dimensionality reduction were conducted using methods such as LASSO (Least Absolute Shrinkage and Selection Operator), PSO (Particle Swarm Optimization), and WT (Wavelet Transformation). Subsequently, 34 different machine learning algorithms were used for modeling. High-performing algorithms were identified, tested using 10-fold cross-validation, and evaluated with performance metrics, confusion matrices, ROC curves and AUC values. The results of the high-performing algorithms were compared and documented. Results: Following segmentation and image preprocessing, 863 features were extracted. After feature selection and dimensionality reduction processes, it was observed that Support Vector Machines (SVM), Linear Discriminant Analysis (LDA), Bagged Trees, Logistic Regression (LR), and k-Nearest Neighbors (k-NN) algorithms performed more successfully in classifying Clear Cell and Non-Clear Cell Renal Cell Carcinomas based on different performance parameters. The k-NN algorithm achieved an accuracy of 68.57% with an AUC of 0.6498, linear SVM achieved 74.29% accuracy with an AUC of 0.7982, LDA achieved 82.86% accuracy with an AUC of 0.8311, LR achieved 71.43% accuracy with an AUC of 0.5796, and Bagged Trees achieved 77.1% accuracy with an AUC of 0.7378. Conclusion: The study results indicate that radiomics-based machine learning methods demonstrated varying degrees of high performance in distinguishing between clear cell and non-clear cell subtypes of renal cell carcinoma. Keywords: Renal Cell Carcinoma, Computed Tomography, Radiomics Analysis, Machine Learning