Precision Diagnosis of Kidney Tumors through Radiogenomics

: Integrating Radiomics, Machine Learning, and Genomic Chromosomal Copy Number Variation

Abstract

Renal tumors, including chromophobe renal cell carcinoma (ChRCC), renal oncocytoma (RO), and clear cell renal cell carcinoma (ccRCC), require accurate diagnosis and classification for optimal patient management. While ChRCC and RO present similar imaging characteristics, making noninvasive differentiation challenging, ccRCC can typically be distinguished from other renal tumors. However, accurately grading ccRCC is crucial for determining tumor aggressiveness and guiding treatment strategies. Traditional diagnostic methods, such as biopsy, are invasive and may lack precision, particularly in grading ccRCC or differentiating between ChRCC and RO. In this thesis, radiomics and machine learning (ML) techniques, alongside genomic data, were employed to address these challenges by improving noninvasive methods for distinguishing tumor subtypes and predicting ccRCC grade.

The first experiment explored the differentiation between ChRCC and RO using radiomics features extracted from preoperative computed tomography (CT) scans. By applying ML-based texture analysis, predictive models were developed and evaluated to provide a ``virtual biopsy" solution that could potentially prevent unnecessary surgical resections for benign RO tumors. The models achieved strong performance with Area Under the Curve (AUC) values indicating high accuracy, demonstrating the feasibility of noninvasive classification of ChRCC and RO.

The second experiment integrated radiomics and genomics to create a radiogenomics map aimed at distinguishing RO and ChRCC. By correlating radiomic features from CT scans with single nucleotide polymorphism (SNP) array-based copy number variation (CNV) data, key radiomic-genomic associations were identified. In this context, some of the radiomic features were used as proxies for gene markers. The study demonstrated that correlating imaging features with genomic data enhances the differentiation of RO and ChRCC, suggesting the potential of radiogenomics to provide an accurate and noninvasive diagnostic tool.

The third study addressed the grading of ccRCC, which accounts for the majority of renal cancer cases. Preoperative grading of ccRCC is essential for determining tumor aggressiveness and management. Using radiomics and ML models, the study predicted the WHO/ISUP grade of ccRCC and analyzed tumor subregions for grading accuracy. The findings revealed that tumor heterogeneity is a critical factor in grading, with certain subregions offering better predictive performance. Radiomics models outperformed biopsy-based grading, demonstrating high accuracy and underscoring their potential as a preoperative tool for ccRCC grading compared to traditional biopsy histology.

Overall, this thesis advances the field of renal tumor diagnosis by integrating radiomics, ML, and genomic data, offering novel approaches for the noninvasive classification and grading of renal tumors. These findings may contribute to accurate, personalized treatment strategies and reduce the need for invasive procedures.

Date of Award	2025
Original language	English
Awarding Institution	University of Dundee
Supervisor	Douglas Steele (Supervisor) & Benjie Tang (Supervisor)