Accurate classification of renal masses as benign versus malignant, and stratification of malignant subtypes by aggressiveness, is essential for guiding treatment decisions in patients with incidentally discovered renal lesions. CT imaging is the primary modality used for this purpose, but radiologist interpretation is subjective and can be inconsistent, particularly for small renal lesions.
Radiomics, the extraction of quantitative imaging features from regions of interest on CT or MRI, offers a data-driven alternative that can capture subtle textural and morphologic information beyond what is apparent visually. Most radiomic studies have focused exclusively on features extracted from within the tumor itself, the intratumoral region. However, the tumor microenvironment and the tissue immediately surrounding a tumor also contain meaningful biological information.
This study systematically evaluated eight different regions of interest (ROIs) on CT, including intratumoral regions and multiple peritumoral shells of varying thickness, to determine which combination of radiomic features provides the best classification accuracy for renal lesions. The study also introduced intratumoral heterogeneity ecological diversity features as a novel dimension of tumor characterization.
The study enrolled 1,795 patients with renal lesions from three hospital sites, creating a diverse multi-institutional cohort suitable for assessing model generalizability. The dataset included a range of lesion types including clear cell RCC, papillary RCC, chromophobe RCC, oncocytoma, and cysts, providing a realistic distribution of lesion types seen in clinical practice.
Eight radiomic ROIs were defined: the intratumoral region (ITR) and seven peritumoral regions (PTR) at different expansion distances (1 mm, 3 mm, 5 mm, and combinations such as 3 mm expanded outward plus 3 mm further expansion). Each ROI was independently characterized using standard radiomic feature extraction including first-order statistics, shape descriptors, and texture features.
Intratumoral heterogeneity ecological diversity features were derived by analogy to ecological diversity metrics used in biology. These features quantify the diversity and evenness of radiomic subregion phenotypes within a tumor, similar to how ecologists measure species diversity within an ecosystem. This novel approach captures spatial heterogeneity that single-value radiomic features may miss.
Classification was performed using AutoGluon-Tabular, an automated machine learning (AutoML) framework that automatically trains and ensembles multiple model types including gradient boosting, neural networks, and linear models. AutoML frameworks reduce the need for manual hyperparameter tuning and model selection, making the pipeline more reproducible and accessible.
AutoGluon-Tabular was selected because it consistently achieves strong performance on tabular datasets with minimal user configuration. The framework uses bagging and stacking ensembles, where predictions from multiple base models are combined by a meta-learner, typically outperforming any individual model. This is particularly beneficial when dealing with high-dimensional radiomic feature sets.
The study assessed all possible combinations of the eight ROIs plus the ecological diversity feature to identify which combination maximized AUC. This systematic feature combination analysis required training and evaluating many model variants, demonstrating the practical value of AutoML in enabling large-scale feature selection experiments.
The best-performing model, which combined intratumoral radiomic features (ITR-3mm), peritumoral radiomic features (PTR-3 to +3mm), and the ecological diversity feature, achieved an AUC of 0.946 on the whole cohort. This substantially outperformed radiologist performance, which achieved an AUC of 0.823 on the same cases.
The performance advantage was even more pronounced for small renal lesions, which represent the most diagnostically challenging subset. For lesions below a defined size threshold, the AI model achieved AUC 0.935 compared to radiologist AUC of 0.745. Small lesions are disproportionately important because they are the most commonly encountered incidental finding and the population in which management decisions are most uncertain.
The combined model including peritumoral features consistently outperformed models using intratumoral features alone, confirming the hypothesis that the tissue surrounding a renal tumor carries diagnostically relevant information beyond what is encoded in the tumor itself. The peritumoral microenvironment likely reflects the biological interaction between tumor and host tissue, including angiogenesis, immune infiltration, and fibrotic remodeling.
When the ecological diversity feature was added to the best radiomic feature combination, AUC improved from 0.929 to the reported optimum, confirming that intratumoral heterogeneity captured through ecological metrics provides complementary information beyond conventional radiomic features.
Ecological diversity in this context measures the evenness and variety of radiomic subregion phenotypes within the tumor, analogous to measuring whether a forest contains many species of similar abundance versus one dominant species. Tumors with high radiomic diversity tend to be more biologically aggressive, and this diversity index may serve as a surrogate for intratumoral genetic heterogeneity.
The consistent benefit of ecological diversity features across multiple outcome comparisons suggests that this novel feature class captures a biologically meaningful dimension of tumor heterogeneity. Future studies should explore whether ecological diversity features correlate with molecular markers of genomic heterogeneity in RCC.
The management of small incidentally discovered renal lesions represents one of the most common clinical dilemmas in urology. Current guidelines recommend biopsy, active surveillance, or treatment based on lesion size, enhancement characteristics, and patient factors, but decision-making remains uncertain for a large proportion of cases.
An AI model with AUC 0.946 for distinguishing benign from malignant lesions could reduce the number of unnecessary biopsies and surgeries for benign lesions such as oncocytoma or angiomyolipoma without visible fat. It could also identify patients with small malignant lesions who would benefit most from early intervention rather than surveillance.
The multi-site validation at three hospitals supports the generalizability of the model across different imaging protocols and patient populations. Prospective validation and integration into clinical decision support workflows would be the next steps toward clinical deployment. This study represents one of the largest and most methodologically rigorous radiomic studies for renal lesion classification to date.