Artificial intelligence has emerged as a transformative technology across nearly every domain of medical imaging, and renal imaging is no exception. This state-of-the-art review, authored by researchers at the National Institutes of Health, provides a comprehensive synthesis of the current landscape of AI applications in kidney cancer detection, characterization, and management as of 2025.
The review covers a broad scope including segmentation of renal anatomy and tumors, differentiation of malignant from benign masses, histologic grade prediction, radiogenomics linking imaging to molecular alterations, and emerging technologies including foundation models, large language models (LLMs), large vision-language models (LVLMs), federated learning, and multi-center validation challenges.
By organizing the literature into these thematic domains, the review provides clinicians and researchers with a structured map of where AI adds the most value in renal imaging, where significant limitations remain, and what directions are most promising for near-term clinical translation.
Automated segmentation of renal anatomy and tumors is foundational to nearly all subsequent AI analysis tasks. The review traces the evolution of segmentation techniques from early simple methods to state-of-the-art deep learning approaches. Early methods included thresholding (separating tissue by Hounsfield unit values), region growing (expanding a seed point based on intensity similarity), and edge detection (identifying boundaries between tissues).
More sophisticated traditional approaches included watershed algorithms (which treat image intensity as a topographic surface and identify drainage basins), active contour models (deformable curves or surfaces guided by image gradients), level set methods (representing contours as implicit functions that evolve over time), and atlas-based segmentation (registering a pre-labeled template to new images). While these methods advanced beyond simple thresholding, they required careful parameter tuning and often failed in the presence of heterogeneous tumors or imaging artifacts.
Deep learning architectures, particularly the U-Net and its variants, have achieved significantly superior performance for renal segmentation. U-Net uses an encoder-decoder architecture with skip connections that allow fine spatial detail preserved in early layers to be combined with high-level semantic features learned in deeper layers. Extensions such as YOLO-based detection and Mask R-CNN for instance segmentation have further expanded the toolkit for automated renal mass analysis.
Radiomics refers to the high-throughput extraction of quantitative features from medical images, capturing information about tumor shape, texture, intensity distribution, and internal heterogeneity that is invisible to visual inspection. When combined with machine learning classifiers, radiomic features can be used to differentiate benign from malignant renal masses, distinguish RCC subtypes, and predict histologic grade.
The review summarizes evidence that radiomic models can reliably distinguish clear cell RCC from oncocytoma, a benign tumor that is frequently indistinguishable from RCC on conventional imaging. Prediction of Fuhrman nuclear grade, a histopathologic grading system that strongly predicts RCC prognosis, has also been demonstrated using radiomic features extracted from CT, with models achieving AUC values consistently above 0.80 in well-designed studies.
The shift from hand-crafted radiomic features to deep learning-derived features represents the current frontier in this domain. Convolutional neural networks can learn task-specific representations directly from image data, often outperforming engineered radiomic features, particularly when sufficient training data is available. Hybrid approaches combining conventional radiomics with deep features are also being explored.
Radiogenomics is an emerging field that links imaging phenotypes to underlying molecular alterations, aiming to non-invasively characterize the genetic landscape of a tumor using imaging alone. In RCC, key molecular targets include VHL (von Hippel-Lindau gene, mutated in most clear cell RCC), PBRM1 (polybromo-1, associated with favorable prognosis), and BAP1 (BRCA1-associated protein-1, associated with aggressive disease and poor survival).
The review summarizes studies demonstrating that CT and MRI radiomic features correlate with VHL mutation status, PBRM1 loss, and BAP1 expression in ccRCC. While these correlations are statistically significant, effect sizes are often modest, and the biological mechanisms linking imaging texture to specific gene mutations are not fully understood. Nevertheless, the potential to identify BAP1-mutated tumors non-invasively, which would have direct prognostic and potentially therapeutic implications, is a compelling motivation for continued research.
Radiogenomics studies face methodological challenges including small sample sizes (requiring tumor biopsies for molecular annotation), potential sampling bias from heterogeneous tumors, and the difficulty of identifying which imaging features are truly mechanistically linked to molecular alterations versus spuriously correlated. Multi-institutional datasets with standardized molecular profiling protocols will be essential to advance this field.
Foundation models are large-scale AI models trained on massive and diverse datasets that can be adapted to specific downstream tasks through fine-tuning on relatively small task-specific datasets. In medical imaging, general-purpose foundation models such as SAM (Segment Anything Model) have been applied to renal image segmentation with promising results, demonstrating that general visual understanding can transfer effectively to specialized medical domains.
Large language models (LLMs) such as GPT-4 are being evaluated for automated radiology report generation, clinical decision support, and summarizing complex imaging findings in accessible language. Large vision-language models (LVLMs) combine visual and language understanding, enabling tasks such as answering clinical questions about image content or generating structured reports from imaging without explicit feature engineering.
The review notes that while these technologies are exciting, their deployment in renal imaging remains early-stage. Key challenges include ensuring factual accuracy of AI-generated reports, integrating these systems into radiology workflows, addressing regulatory requirements for AI-generated clinical content, and evaluating performance in diverse real-world patient populations.
Federated learning is a distributed training approach that allows AI models to be trained across multiple institutions without sharing patient data. Each institution trains the model locally on its own data, and only model weights (not patient information) are shared and aggregated. This approach addresses privacy concerns and regulatory barriers to multi-center AI training while enabling access to the scale and diversity of data needed for robust models.
Multi-center validation is the critical bottleneck for clinical translation of renal AI tools. Many promising single-institution models fail when applied to data from other centers due to differences in scanner type, imaging protocol, patient demographics, and local radiologist practices. The review highlights several multi-center efforts including the KiTS (Kidney Tumor Segmentation) challenge, which has driven substantial progress in automated renal tumor segmentation benchmarking.
Regulatory approval of AI tools for renal imaging is an active area of development. The FDA has cleared several AI imaging tools as Software as a Medical Device (SaMD), but the specific requirements for RCC detection tools including clinical trial evidence, post-market surveillance obligations, and performance standards across subpopulations are still evolving. Understanding the regulatory landscape is essential for developers aiming to bring renal AI tools into clinical practice.