Textural differences between renal cell carcinoma subtypes: Machine learning-based quantitative computed tomography...

Area of Science:

• Radiology
• Oncology
• Artificial Intelligence

Background:

• Renal cell carcinoma (RCC) comprises several subtypes with varying prognoses.
• Accurate subtype differentiation is crucial for effective treatment planning.
• Quantitative computed tomography texture analysis (qCT-TA) offers potential for non-invasive tumor characterization.

Purpose of the Study:

• To develop and validate machine learning models for distinguishing RCC subtypes using qCT-TA.
• To assess the generalizability and reproducibility of qCT-TA models.
• To compare the performance of qCT-TA on unenhanced versus contrast-enhanced CT images.

Main Methods:

• Retrospective analysis of 68 RCCs for model development and internal validation, with external validation using 26 RCCs from The Cancer Genome Atlas (TCGA).
• Extraction of 275 texture features from unenhanced and corticomedullary phase CT images.
• Feature selection using radiologist reproducibility and a wrapper-based algorithm, followed by nested cross-validation.
• Classification using artificial neural networks (ANN) and support vector machines (SVM), with and without ensemble methods.

Main Results:

• Corticomedullary phase CT images yielded more reproducible texture features (232/275) than unenhanced images (93/275).
• ANN with adaptive boosting achieved the best performance for non-cc-RCC vs. cc-RCC discrimination (MCC=0.728, external accuracy=84.6%).
• Performance was poor for distinguishing three major subtypes; SVM with bagging showed the best results for papillary RCC vs. others (MCC=0.804, external accuracy=69.2%).

Conclusions:

• Machine learning-based qCT-TA can effectively distinguish non-cc-RCC from cc-RCC.
• The current qCT-TA approach demonstrates poor performance in differentiating the three major RCC subtypes.
• Corticomedullary phase CT imaging provides superior texture parameters for RCC analysis compared to unenhanced imaging.