When kidney cancer spreads beyond the kidney to other parts of the body - known as advanced or metastatic renal cell carcinoma (RCC) - it becomes much harder to treat. Survival rates drop significantly, and determining the best treatment strategy for each individual patient is a complex challenge that current methods do not always handle well.
Artificial intelligence (AI), particularly techniques that analyze medical images and genomic data, offers new tools for this challenge. This review article systematically examines how AI is being applied across the management of advanced and metastatic RCC, from detecting the cancer at initial staging to predicting how patients will respond to treatment.
The review covers a range of AI approaches including radiomics (extracting quantitative features from CT or MRI scans), deep learning (training neural networks to recognize patterns in images or data), and radiogenomics (linking imaging features to underlying genetic characteristics of tumors). Together, these technologies represent a new frontier in kidney cancer management.
At the staging phase, AI models have been developed to automatically detect tumor features that predict aggressive behavior - including renal vein invasion, lymph node involvement, and tumor thrombus (tumor growing into blood vessels). These high-risk features change surgical planning and affect prognosis, but they can be subtle and easily missed or inconsistently assessed by human readers on CT scans.
Several AI models have also been trained to predict which patients are likely to develop metastatic disease after kidney removal surgery. If such a model could reliably identify high-risk patients before metastasis occurs, it might be possible to intervene earlier with adjuvant therapies. Current clinical tools for this prediction (such as the UCLA Integrated Staging System) remain imperfect, and AI models using radiomic features have shown promise for improving accuracy.
Radiogenomics - pairing CT imaging features with genomic data - represents one of the most exciting applications. Research has shown correlations between specific CT texture features and mutations in genes like VHL, PBRM1, and BAP1 that drive ccRCC. If imaging features can non-invasively estimate a tumor's genetic profile, this could guide treatment selection without the need for additional biopsies.
One of the most clinically impactful potential applications of AI in advanced RCC is predicting which patients will respond to immune checkpoint inhibitors - immunotherapy drugs that have transformed the treatment of metastatic kidney cancer. Not all patients respond to these drugs, and treating non-responders exposes them to side effects without benefit while delaying other potentially effective therapies.
Radiomics-based models have been developed to predict response to drugs like nivolumab (anti-PD-1) and combination immunotherapy regimens. These models analyze CT scan features before treatment begins to estimate likelihood of response. Early studies have shown encouraging accuracy, though most remain limited by small patient numbers and single-institution designs.
AI tools have also been applied to predict response to targeted therapies - drugs that block specific molecular pathways that RCC cells depend on for growth, such as VEGF inhibitors (sunitinib, pazopanib) and mTOR inhibitors (everolimus). Predicting which of these multiple options is best for a given patient is a key unmet clinical need that AI may help address.
Despite the promise of AI in advanced RCC management, the review identifies important limitations that currently prevent widespread clinical adoption. Most published studies are retrospective - they analyze data from patients who were already treated, rather than prospectively testing AI tools in new patients. Retrospective studies are prone to bias and may not perform as well when applied to new patient populations.
A second major limitation is small cohort sizes. Many AI models for RCC have been developed and tested on fewer than 100 patients. Machine learning models trained on small datasets may overfit - performing well on the data they were trained on but failing when applied to different patients at different hospitals. Validation across multiple large, diverse institutions is essential before any of these tools can be used clinically.
A third challenge is the lack of standardization in how radiomic features are extracted from CT scans. Different scanners, scanning protocols, and feature extraction software can produce different feature values for the same tumor, making it difficult to compare or combine results across studies. Ongoing efforts to standardize radiomics methodology are critical to moving the field forward.
AI represents a genuine opportunity to improve outcomes for patients with advanced and metastatic kidney cancer - a group where current treatments are still insufficient for many. By extracting more information from existing imaging studies, AI could help detect disease earlier, identify high-risk patients who need more aggressive treatment, and match patients with the treatments most likely to work for their specific tumor biology.
For patients and families, the near-term impact of this research is likely to come from AI tools that assist radiologists in detecting subtle features on CT scans that affect staging and treatment decisions. Longer-term, the integration of imaging data with genomic and clinical data through AI may enable truly personalized treatment planning for kidney cancer.
The authors conclude that rigorous prospective validation studies - comparing AI-guided decisions to standard care in randomized clinical trials - are the necessary next step to move these tools from research publications into real-world clinical benefit for patients.