Radiomics-based kidney lesion classification: Mitigating batch effect with nested combat harmonization

The increased use of CT imaging has elevated the incidental detection of renal masses, necessitating accurate differentiation between benign and malignant nodules. Radiomics offers potential for improved diagnostics; however, it is limited by variability in imaging parameters such as slice thickness, highlighting the need for effective harmonization techniques.

The purpose of this study is to conduct a comprehensive radiomics analysis, evaluating the impact of slice thickness in distinguishing between kidney cysts and tumors using machine learning techniques, thus contributing to more precise and effective patient management strategies.

We utilized a publicly available dataset, KITS23, and extracted radiomic features from contrast-enhanced computed tomography (CT) scans using the PyRadiomics library. The dataset consists of 599 cases, which were divided into training (60%) and testing (40%) cohorts to develop and validate predictive models. Six feature selection methods and ten machine learning classifiers were employed. Additionally, the Nested Combat harmonization technique was applied to address variations in imaging protocols across institutions.

We observed improvements in AUC values across various feature selection methods and classifiers after harmonization, with the highest AUC reaching 0.95. This represents significant enhancements in model performance, with mean AUC improvements ranging from 0.7% to 7.7% across different feature selection methods, bringing our results in line with, and in some cases surpassing, the AUCs reported in the literature.

These findings underscore the potential of radiomics-based machine learning models to enchance diagnostic accuracy and patient management in clinical practice. The use of harmonization techniques, such as, Nested Combat is crucial in achieving reliable and generalizable predictive models for renal oncology.