Renal cell carcinoma (RCC) accounts for over 90% of all kidney cancers and is the most lethal urogenital malignancy, carrying a mortality rate of 30 to 40%. That is roughly double the mortality rate of bladder and prostate cancer, which hover around 20%. The disease encompasses several histological subtypes. Clear cell RCC (ccRCC), papillary RCC (types I and II), and chromophobe RCC (chRCC) are the most common solid forms, together representing 70 to 90% of all cases.
A major diagnostic challenge is that most RCC cases are discovered incidentally during imaging ordered for unrelated conditions. Only about 30% of patients receive a diagnosis based on symptoms such as pain, hematuria, or palpable masses. Additionally, 10 to 17% of surgically removed kidney tumors turn out to be benign on final histopathological assessment, meaning a significant number of patients undergo unnecessary surgery. Renal mass biopsy (RMB) offers high sensitivity (99.1%) and specificity (99.7%) for detecting malignancy, but it carries risks including bleeding, tumor seeding, and hematomas at a rate of about 4.3%. Furthermore, only 64.6% of oncocytomas diagnosed by RMB were confirmed benign after surgical resection, highlighting accuracy gaps in current biopsy methods.
The authors argue that artificial intelligence can fill these gaps by extracting patterns from digitized health records, virtual pathology slides, and imaging data that are not perceptible to the human eye. Machine learning (ML) and its more advanced subset, deep learning (DL), can analyze extensive datasets and classify tissue structures with speed and consistency that manual review cannot match. This review examines AI advancements across four diagnostic domains: histopathology, multi-omics, imaging, and perioperative planning.
The authors performed a non-systematic literature review by searching the PubMed and Google Scholar databases. Search terms included "artificial intelligence," "diagnostics," "renal cell carcinoma," "histology," "multi-omics," "imaging," and "perioperative diagnostics." All relevant studies published up to January 15, 2025, were considered, and only English-language articles were included.
Inclusion criteria required that publications (1) discuss artificial intelligence in relation to renal cell carcinoma, and (2) be full-text original articles or review articles containing all standard subsections: introduction, materials and methods, results, and discussion with conclusions. Preliminary screening was conducted by examining titles and abstracts, followed by a comprehensive review of full-text articles. The authors also identified additional pertinent articles from the reference listings of those already selected.
The reviewed studies were organized into four thematic areas of AI-enhanced diagnostics for RCC: histological analysis, multi-omics, imaging, and perioperative planning. Each area is summarized with tables that catalog the specific AI application, the model or algorithm used, and the main results or conclusions from each study. This structure allows clinicians and researchers to quickly identify which AI approaches have shown the most promise for each diagnostic domain.
Deep learning for whole slide images (WSIs): Nyman et al. built a deep learning framework using ResNet50, a convolutional neural network (CNN), to generate quantitative high-resolution whole slide images of ccRCC. The system creates spatial maps and regional adjacency graphs from a single WSI, identifying intratumor heterogeneity features that go beyond what manual pathology review can detect. Graph-based microheterogeneity analysis improved prognostic predictions, including forecasting how patients respond to immune checkpoint inhibitors.
Nuclear grading and subtype classification: Holdbrook et al. developed an automated pipeline for measuring nuclear pleomorphic patterns in ccRCC across a cohort of 59 patients. A five-fold cross-validation of the patch classifier showed a correlation of r = 0.59 between predicted feature values and regional multigene scores. In a separate study, Abu Haeyeh et al. used decision-fusion deep learning to distinguish ccRCC from chRCC and non-RCC tissues, achieving 93.0% accuracy, 91.3% sensitivity, and 95.6% specificity. Their method outperformed the standard ResNet-50 model for RCC subtype classification.
Pyramidal CNN pipelines: Abdeltawab et al. designed a pyramidal deep learning pipeline incorporating three CNNs for pixel-wise and patch-wise classification of kidney WSIs. This approach surpassed other neural networks, including ResNet, and complemented pathologists' expertise in automated histopathological diagnosis. Meanwhile, Cai et al. integrated texture feature descriptors with deep learning platforms. Their best model used an SVM classifier combined with AlexNet and Gabor filter features, reaching 98.54% accuracy. Other texture combinations also performed well: AlexNet with HOG (93.76%), GLCM (94.52%), LBP (93.45%), and MRF (97.39%).
Mutation prediction from histology: Zheng et al. developed a self-supervised attention-based multiple instance learning (SSL-ABMIL) model to predict tumor mutation burden (TMB) and VHL mutation status from hematoxylin and eosin-stained slides. The Wang-ABMIL variant produced AUROC scores of 0.83 for TMB prediction and 0.80 for VHL mutation prediction. Attention heat maps confirmed the model focused on tumor regions for high-TMB patients and detected stromal lymphocytic infiltration patterns for VHL prediction.
While most AI histopathology studies focus on morphological pattern recognition through neural networks, He et al. took a different approach by integrating numerical data from specific marker proteins. They examined autophagy-related proteins (ATGs) obtained from immunohistochemical (IHC) images of RCC tissue microarrays. Using IHC staining and automated quantification, the team observed significant reductions in ATG1, ATG5, and LC3B levels in RCC tissue, indicating decreased basal autophagy in renal cell carcinoma.
The researchers applied K-Nearest Neighbor (KNN) machine learning to this protein expression data and identified LC3B as a strong indicator for clear cell RCC (ccRCC). High ROC curve values also demonstrated that p62 could serve as a robust marker for distinguishing chromophobe RCC subtypes from other forms. Combinations of multiple autophagy-related proteins further improved the ability to differentiate RCC subtypes from normal tissue.
This protein-based approach is significant because it offers a complementary pathway to the purely image-based deep learning methods. Rather than relying solely on visual pattern recognition in tissue slides, it quantifies specific biological markers that reflect underlying cellular processes. The authors suggest this enhances the potential for precision oncology applications, where treatment decisions can be guided by both morphological and molecular characteristics of the tumor.
Transcriptomic classification: Jagga et al. used four supervised learning algorithms (J48, Random Forest, SMO, and Naive Bayes) to distinguish early-stage from late-stage ccRCC based on transcriptomic signatures from The Cancer Genome Atlas. The Random Forest model outperformed the others, achieving 89% sensitivity, 77% accuracy, and an AUC of 0.80. It identified 62 differentially expressed genes between the two staging groups, reinforcing the link between molecular mechanisms and disease progression. Separately, Liu et al. applied bioinformatics and neural networks to identify ten hub genes (including TPX2, AURKB, NCAPG, and CCNA2) as significant diagnostic biomarkers for ccRCC, all associated with poor overall survival.
Urine metabolomics: Bifarin et al. demonstrated a non-invasive approach using liquid chromatography-mass spectrometry and nuclear magnetic resonance data fed into an optimized machine learning platform. Their urine-based assay for RCC detection achieved an AUC of 0.98, accuracy of 88%, sensitivity of 94%, and specificity of 85%. This represents one of the strongest non-invasive diagnostic performances reported in the review.
Lipidomics and glycan-based biomarkers: Manzi et al. combined AI with mass spectrometry-based lipidomics, building support vector machine models that yielded a 16-lipid panel discriminating ccRCC from controls with 95.7% accuracy in training and 77.1% in testing. A second model distinguished early- from late-stage ccRCC with 82.1% accuracy. Iwamura et al. analyzed N-glycan signatures from 100 serum subjects with RCC using supervised machine learning, producing a scoring system with an AUC of 0.99, 90% sensitivity, and 99% specificity for RCC detection across all pathological stages.
DNA methylation for difficult cases: Differentiating RCC from renal oncocytoma is challenging due to their morphological similarity. Brennan et al. profiled DNA methylation in fresh-frozen oncocytoma and chRCC tumors using machine learning. A model based on 30 differentially methylated CpG sites distinguished oncocytoma from chRCC with an AUC of 0.96 in 10-fold cross-validation. It also separated chRCC from other RCC subtypes with an AUC of 0.87, establishing DNA methylation as a reliable standalone biomarker.
Deep learning on CT scans: Xu et al. pre-trained deep learning models using stochastic gradient descent (SGD) with cross-entropy loss for 60 epochs, developing four separate DL networks with individual weights on 9,978 images. The single deep learning model achieved an AUC of 0.864 in the validation cohort, while the ensembled model demonstrated superior performance with an AUC of 0.882, showing that AI can offer accuracy comparable to traditional biopsy while remaining non-invasive.
CT plus clinical metadata for surgical selection: Mahmud et al. integrated AI with CT scans and clinical metadata to grade cancer patients and determine the most suitable surgical procedure. Their system achieved 85.66% accuracy, 84.18% precision, and an F1-score of 84.92% for cancer grading. For selecting surgical procedures specifically for malignant RCC tumors, performance was even stronger at 90.63% accuracy, 90.83% precision, and an F1-score of 90.50%. Feature ranking identified tumor volume and cancer stage as the key factors driving diagnosis and treatment decisions.
Distinguishing RCC from fat-poor angiomyolipoma: Fat-poor angiomyolipoma (AML) variants are frequently misdiagnosed as RCC. A Cleveland Clinic report found that imaging errors led to preoperative misdiagnosis in 55% of AML patients. Yao et al. developed a multichannel deep learning classifier trained on unenhanced whole-tumor CT images that achieved an average AUC of 0.951 in five-fold cross-validation, with validation AUCs of 0.966 (internal) and 0.898 (external), performing especially well for large tumors of 40 mm or greater. These models trained on unenhanced CT outperformed those using contrast-enhanced CT images.
Contrast-enhanced ultrasound (CEUS): Luo et al. developed a radiomic ML model based on renal cancer CEUS images to predict preoperative differentiation grades. The model achieved an AUC of 0.811, specificity of 0.786, and accuracy of 0.784, reinforcing its utility for WHO/ISUP nuclear grading and the non-invasive diagnosis of ccRCC. Yang et al. assessed CNN-based automatic segmentation of CEUS images in renal tumors, testing seven models. The UNet++ model delivered the best results, with a mean Intersection over Union (mIOU) of 93.04%, Dice Similarity Coefficient (DSC) of 92.70%, precision of 97.43%, and recall of 95.17%.
Shear wave elastography with machine learning: Sagreiya et al. integrated four machine learning algorithms (logistic regression, naive Bayes, quadratic discriminant analysis, and SVM) to evaluate data from the tumor, cortex, and combined tumor-cortex-medulla inputs. SVMs achieved the highest performance at 94% accuracy and an AUC of 0.98, significantly outperforming the median shear wave velocity (SWV) measurement. Most models using combined inputs demonstrated strong performance, highlighting the algorithms' ability to differentiate between RCC and AML with high classification accuracy.
MRI-based deep learning: Xi et al. implemented a ResNet deep learning network to analyze MRI data (T1C and T2WI sequences) for differentiating RCC from benign renal masses. The model outperformed both traditional radiomics approaches and expert radiologist assessments. In a related study, Zheng et al. used ResNet on T2-weighted fat saturation MRI sequences to discriminate between high-grade and low-grade RCCs (AJCC grade I and II). The model achieved an overall accuracy of 60.4% and a macro-average AUC of 0.82, with subtype-specific AUCs of 0.94 for ccRCC, 0.78 for chRCC, 0.80 for AML, and 0.76 for pRCC.
Surgical workflow recognition: Nakawala et al. developed the Deep-Onto network to optimize the surgical workflow for robotic-assisted partial nephrectomy (RAPN). Using over 700,000 frames extracted from nine comprehensive RAPN videos, the system delineated ten distinct surgical phases and categorized the data into test, training, and validation sets. Although still in early implementation, the model achieved 74.0% precision and 74.3% accuracy in predicting the steps of RAPN, demonstrating that AI can provide real-time intraoperative feedback and adapt to unanticipated surgical scenarios.
Hidden vessel detection: Amir-Khalili et al. built an automated guidance system to reduce intraoperative complications by identifying hidden accessory vessels, particularly those concealed within fat. The system analyzed temporal motion characteristics in endoscopic video, using pulsatile motion alongside static visual attributes such as texture, intensity, and color as features for vessel segmentation. The AUC for this technique was 0.72, indicating promising initial results for reducing surgical complications. Nosrati et al. extended this approach by training random decision forest ML programs on data from 15 RAPN videos, integrating preoperative imaging to estimate three-dimensional anatomical abnormalities. Their technique achieved a 45% improvement in structure identification within noisy endoscopic environments.
3D surgical planning: Nguyen et al. used Fujifilm's Synapse AI Platform to handle a rare case of RCC with a concurrent duplex kidney. The Synapse 3D platform generated precise three-dimensional simulations to refine tumor identification and enhance visualization of the feeding vessels' spatial relationships. The AI automated the creation of vascular maps, reduced preparation time, and mitigated inconsistencies associated with traditional manual tracing of CT scans, such as operator-dependent errors and variations in scan timing and contrast conditions.
Dataset and validation issues: The authors identify several recurring limitations across the reviewed studies. Selection bias remains a significant obstacle, and randomized controlled trials for RCC diagnostics are unfeasible due to the nature of the disease. Many studies relied on small or homogeneous datasets. For example, Bifarin et al. adjusted for five confounding variables (gender, BMI, age, race, and smoking history), but larger and more geographically diverse cohorts are still needed. Mahmud et al. raised concerns about the generalizability of the KiTS dataset, calling for datasets specifically curated for AI that include a balanced representation of tumor stages and subtypes.
Histopathology-specific challenges: Nyman et al. noted that tissue samples were limited to pre-treatment tumor states, creating a discrepancy with ongoing tumor evolution. The scarcity of pixel-level annotation of all images remains a barrier, and nuclear grading lacks a definitive framework for producing consistent annotations due to diverse tumor microenvironment phenotypes. In computational pathology, the separation of different slide regions caused independent categorizations despite originating from one specimen, and models may lack contextual awareness due to an inability to learn nonlinear interactions among instances.
Imaging-specific challenges: Radiomic variability from multiple CT scanners, restriction to single-phase CT images, and 5 mm slice thickness limitations all affected imaging study performance. Label noise (mismatches between medical image labels and actual images) undermines deep learning methodology quality. For CEUS, the absence of video formatting for automated segmentation leads to loss of temporal information, and single-operator studies compromise reliability. MRI faces difficulty differentiating small renal lesions from surrounding tissue and lacks sufficient published literature on advanced neural network applications.
Path forward: Despite these challenges, the authors conclude that AI has the potential to transform RCC diagnostics across all four domains. In histopathology, AI reveals microscopic patterns invisible to the human eye. Multi-omics AI enables earlier detection through comprehensive molecular profiling. Imaging AI helps clinicians differentiate RCC from other renal lesions and plan personalized treatments. Perioperative AI improves decision-making, patient safety, and recovery time. The authors emphasize that future work must focus on controlling selection bias, assembling larger and more diverse datasets, ensuring reliable external validation, and advancing AI capabilities for real-world clinical deployment.