Accurate preoperative classification of renal tumor subtypes, including clear cell RCC (ccRCC), papillary RCC (pRCC), chromophobe RCC (chRCC), oncocytoma, and angiomyolipoma (AML), has major implications for surgical planning and decision-making about active surveillance versus intervention.
Many renal tumors share overlapping imaging features on standard CT, making definitive subtype diagnosis challenging without surgical pathology. Machine learning applied to quantitative radiomic features extracted from multiphase CT offers an opportunity to improve preoperative classification accuracy beyond what radiologists can achieve by visual inspection alone.
This multicenter study trained and externally validated a radiomics-based XGBoost classifier across five renal tumor subtypes using arterial and venous phase CT data from two German tertiary care centers, providing a realistic assessment of cross-institutional generalizability.
A training cohort of 297 patients was assembled from one German tertiary center, with an independent testing cohort of 121 patients from a second institution. Tumor subtypes included ccRCC (64.3%), pRCC (13.5%), chRCC (7.4%), oncocytoma (9.4%), and AML (5.4%), reflecting realistic clinical incidence proportions.
Radiomic features were extracted from both arterial and venous phase CT images. To address class imbalance in the multi-class setting, SMOTE (Synthetic Minority Over-sampling Technique) was applied to the minority subtype classes during training, preventing model bias toward the dominant ccRCC class.
The classification algorithm selected was XGBoost, an ensemble gradient boosting method known for strong performance on structured tabular data. Models were trained separately for arterial phase, venous phase, and combined multiphase inputs to determine which CT phase provided optimal discriminative information.
Performance was evaluated by per-class AUC and multi-class overall AUC on both the training and independent testing cohorts, with particular attention to whether performance degraded substantially in the external test set.
The venous phase model achieved an overall AUC of 0.81 in training and 0.75 in independent testing. The arterial phase model reached AUC of 0.64 and 0.67 respectively, while the combined multiphase model achieved AUC of 0.80 and 0.75, nearly identical to venous phase alone.
Subtype-specific performance varied considerably. AML was the easiest subtype to classify, achieving AUC of 0.90 to 0.94 across phases, likely due to its distinctive fat-containing imaging phenotype. Oncocytoma was the hardest, with AUC of only 0.57 to 0.69, reflecting its known imaging overlap with chRCC.
The external testing AUC of 0.75 was slightly lower than the training AUC of 0.81, but the gap was relatively modest for a genuine cross-institutional validation, suggesting reasonable generalizability of the venous phase radiomics model across different CT scanner and protocol environments.
Reliable preoperative identification of AML versus malignant subtypes is clinically urgent, as AML can be managed with surveillance or embolization rather than surgery, potentially sparing patients the morbidity of unnecessary nephrectomy. The high AUC for AML supports this specific clinical application.
Distinguishing oncocytoma from chRCC remains an unsolved preoperative challenge. Both have similar imaging appearances, and oncocytoma is benign while chRCC requires surgery. The low AUC for oncocytoma classification in this study highlights the ongoing need for additional biomarkers or multimodal approaches to resolve this diagnostic dilemma.
The multicenter validation design is particularly relevant for clinical translation. A model that maintains acceptable performance across institutions with different CT scanners and protocols is substantially more ready for real-world deployment than a single-center model.
This study provides one of the most rigorous validations of renal tumor subtype classification using radiomics, employing a genuine independent external test set from a different institution rather than a held-out split of the same center's data.
The venous phase alone provides performance equivalent to combined multiphase input, which has practical implications: venous-only protocols expose patients to less contrast and radiation while delivering equivalent diagnostic information for this application.
Future research should focus on increasing dataset size for minority subtypes, integrating clinical and laboratory variables alongside radiomic features, and developing ensemble approaches that combine radiomics with pathomics or molecular data to improve oncocytoma and chRCC discrimination.