Kidney cancer is one of the most common solid tumors in both men and women. Its management is challenging from the very first step - distinguishing whether a kidney mass is benign or cancerous - through to choosing the right treatment and monitoring for recurrence.
A key problem is small renal masses: many small kidney growths turn out to be completely benign, yet current imaging tools cannot reliably tell benign from cancerous without a biopsy. Biopsies, however, fail to provide a diagnosis in about 20% of cases, limiting how widely they are used.
Beyond initial diagnosis, doctors struggle to predict which patients will do well after treatment and which face higher risks of recurrence or spread. Standard staging and grading tools provide only rough estimates, because kidney cancers are genetically diverse - tumors at the same stage can behave very differently depending on their underlying biology.
This review examines a rapidly developing approach called radiogenomics - combining information extracted from imaging scans with genetic data from the tumor - to improve diagnosis, treatment selection, and prognosis prediction for kidney cancer patients.
Radiomics is the process of extracting hundreds or thousands of quantitative measurements from medical imaging scans (such as CT or MRI) using computer software. Instead of a radiologist visually reading a scan and forming a subjective impression, radiomics converts scan information into precise, reproducible numbers.
These measurements capture features that human eyes miss - subtle variations in tissue texture, shape, density patterns, and internal structure across the entire scan - and then use statistical and AI methods to find patterns associated with specific cancer behaviors.
The radiomics pipeline includes data selection, acquiring the imaging scan, extracting features from the image, statistical analysis, and building a predictive model. Each step must be carefully controlled to ensure that results are reproducible and reliable across different hospitals and scanners.
A quality scoring system (the radiomics quality score) has been developed to evaluate how rigorously radiomics studies are conducted - covering everything from how the tumor outline is defined on the scan to whether the results were validated in independent patient groups.
Current research on radiomics in kidney cancer covers several major clinical questions. The first is distinguishing benign from cancerous masses - particularly important because up to 30% of surgically removed kidney masses turn out to be completely benign, meaning those surgeries could have been avoided.
AI algorithms applied to CT and MRI images have shown promise in separating angiomyolipomas (a common benign tumor) and oncocytomas (another benign type) from kidney cancer - with some AI models outperforming radiologists in studies. This could help many patients avoid unnecessary surgery.
Radiomics has also been applied to subtyping kidney cancer - distinguishing clear cell from papillary from chromophobe subtypes - which matters because different subtypes respond differently to targeted therapies. AI neural networks have shown promising accuracy in this area.
Additionally, researchers are using radiomics to predict tumor grade before surgery (estimating how aggressive the cancer cells are from the scan) and to assess treatment response in patients receiving targeted therapies - trying to detect whether a cancer is responding to treatment before it is visibly shrinking on conventional scans.
Clear cell renal cell carcinoma (ccRCC), the most common subtype, typically starts with changes to chromosome 3p. This region contains several important tumor suppressor genes, especially VHL, which when mutated leads to uncontrolled activation of genes that promote blood vessel growth and cancer proliferation. VHL mutations are present in about 92% of ccRCC cases.
Von Hippel-Lindau disease is a hereditary syndrome caused by inherited VHL mutations. Affected individuals develop kidney cysts, kidney tumors (often bilateral), and tumors in the nervous system, retina, and adrenal glands. Kidney tumors in VHL disease tend to appear about 20 years earlier than sporadic kidney cancers.
Papillary kidney cancer (type 1) is driven by MET gene mutations and gains of chromosomes 7 and 17. Type 2 papillary RCC is more genetically diverse and more aggressive. Chromophobe RCC is characterized by the loss of multiple chromosomes and typically has a better prognosis. Each subtype has distinct genetic drivers that influence how patients respond to different treatments.
Understanding these genetic differences matters because targeted therapies work through specific molecular pathways. A treatment targeting the VHL pathway may not help a papillary cancer with MET mutations. Knowing a tumor's genetic profile is essential for selecting the most appropriate therapy - which is precisely what radiogenomics aims to determine from imaging alone.
Standard CT and MRI scans show the size and shape of tumors but say little about their biological activity. Molecular imaging uses special tracers or contrast agents that target specific molecules present in cancer cells, making the biological behavior of tumors visible on a scan.
One promising approach uses a radiotracer called 99mTc-sestamibi with a SPECT/CT scan. This tracer accumulates preferentially in cells with high numbers of mitochondria (the energy factories of cells). Oncocytomas have very high mitochondrial content, while ccRCC has low mitochondrial content - so the scan can potentially distinguish these two types non-invasively, avoiding biopsies in some patients.
Another approach uses girentuximab - an antibody that targets a protein called CAIX (carbonic anhydrase IX) that is overexpressed in nearly all ccRCCs due to VHL mutations. Labeling this antibody with a radioactive tracer and using PET/CT scanning has shown 86% sensitivity and 86% specificity for ccRCC diagnosis in studies - better than standard CT alone.
The standard glucose-based PET tracer (18F-FDG) has limited use for diagnosing kidney masses because the kidneys naturally excrete glucose tracers, creating interference. However, FDG-PET is more useful for detecting cancer that has spread to distant sites (metastases), particularly in lymph nodes and bone, and may help predict responses to immunotherapy.
Radiogenomics combines radiomics (quantitative imaging data) with genomics (a tumor's genetic profile). The fundamental idea is that the way a tumor looks on a scan - its texture, shape, enhancement patterns, vascularity - reflects its underlying biology and genetic makeup. If true, scan features could serve as a non-invasive window into tumor genetics.
Early research confirmed that specific genetic mutations are associated with specific imaging appearances. For example, VHL mutations are linked to well-defined tumor edges and nodular patterns of blood vessel activity on CT scans. BAP1 mutations are associated with tumor invasion into veins and poorly defined tumor edges.
These associations allow AI models to predict which mutations a tumor likely carries based on its scan appearance alone - without requiring a biopsy. Several studies have built models predicting BAP1, PBRM1, and other mutation status from CT images, with accuracy measures (AUC) ranging from 0.52 to 0.987 across different studies.
Beyond mutation prediction, radiogenomics has been extended to predicting overall survival, metastasis risk, and treatment response. Models combining radiomics features with hypoxia-related gene signatures have achieved AUC values of up to 0.91 for survival prediction in ccRCC - suggesting that scan information alone can capture meaningful prognostic biology.
Despite promising research results, radiogenomics is not yet part of routine kidney cancer care. The field faces significant limitations that must be overcome before it can influence daily clinical decisions.
Most studies are retrospective - researchers analyze historical data from patients already treated, which introduces selection bias and makes it difficult to confirm that findings reflect real-world clinical scenarios. Few well-designed prospective trials (where patients are followed forward from the time of enrollment) exist.
A major technical barrier is the lack of standardization. Different hospitals use different CT or MRI scanner models, different scan settings, different ways of defining tumor boundaries, and different software for extracting radiomics features. Small changes in any of these factors can dramatically change the features measured, making it difficult to compare results across studies or apply models trained at one institution to patients at another.
Additionally, most radiogenomics studies involve small patient numbers - sometimes only a few dozen patients per study. Building reliable AI models requires much larger datasets, and validating those models in completely separate external patient groups (external validation) is rarely performed rigorously. The result is that many published models may perform worse than reported when applied to new patients in different hospitals.
If radiogenomics fulfills its promise, a kidney cancer patient undergoing a routine CT scan could receive far more information than they do today. The scan could not only show the tumor's size but also predict its genetic subtype, mutation status, and aggressiveness - all without a biopsy.
This information could directly guide treatment selection. A tumor predicted to carry VHL mutations might be prioritized for VEGF-targeted therapies like sunitinib, while one predicted to carry BAP1 mutations - associated with worse prognosis - might prompt more aggressive treatment from the outset or earlier consideration of clinical trials.
For patients with small renal masses, radiogenomics could help identify those that are truly benign - sparing them surgery. Approximately 30% of surgically removed kidney masses turn out to be benign. A reliable non-invasive tool that could identify benign masses with high confidence would prevent significant numbers of unnecessary operations.
Radiogenomics could also transform treatment monitoring. Rather than waiting months for a tumor to visibly shrink (or grow) on a follow-up scan, molecular and texture changes detectable by radiogenomics could signal whether a treatment is working weeks earlier - allowing faster switches to alternative therapies when needed.
Radiogenomics in kidney cancer is a genuinely promising field, but it is still in early development. The research done so far has established that there are real, measurable connections between how tumors appear on scans and their underlying genetics - the biological foundation for this approach is solid.
The next steps must include larger, prospective, multi-center clinical trials that follow patients from the start of their care rather than looking back at historical records. These trials need to test radiogenomics models on patients from multiple hospitals using standardized scanning protocols.
AI and deep learning will play an increasingly important role - not just in predicting single mutations, but in combining genomic, transcriptomic, proteomic, and metabolic data together with imaging to build a comprehensive, multi-dimensional picture of each tumor. The growing availability of large public databases (like The Cancer Genome Atlas) provides the raw material for training such models.
For patients and families, the key takeaway is that the field is moving toward a future where a standard scan can tell doctors far more about a kidney cancer than it does today - with the goal of delivering more personalized, better-targeted treatment while sparing patients unnecessary procedures and treatments that won't work for their specific cancer's biology.