Clear cell kidney cancer (clear cell renal cell carcinoma, or ccRCC) is the most common and most dangerous type of kidney cancer, making up about 70-75% of all cases. It tends to be more aggressive than other kidney cancer types, more likely to spread, and harder to treat at advanced stages.
We now know that kidney cancer is not just one disease - it is characterized by four important genetic mutations (changes in genes called VHL, PBRM1, BAP1, and SETD2) and four distinct molecular subtypes based on patterns of gene activity. These differences matter enormously for treatment: different mutations respond differently to different drugs, and different molecular subtypes have different survival outcomes.
The problem is that finding out which mutations a patient's tumor has - and which molecular subtype it belongs to - currently requires either a tissue biopsy or analysis of surgically removed tissue. These procedures are invasive, carry risks, and are expensive. Once a tumor is removed in surgery, the molecular information is available, but by then it is too late to use it for planning the operation or choosing initial treatment.
This study, from researchers at West China Hospital and Sichuan University, explored a bold question: could a regular CT scan, combined with sophisticated computer analysis, reveal the genetic mutations and molecular subtype of a patient's kidney tumor - without any biopsy at all?
Researchers used CT scan images and detailed biological data from 207 kidney cancer patients in The Cancer Genome Atlas (TCGA) - a large U.S. national database that pairs cancer imaging with genetic, protein, and clinical data. A second, independent group of 175 patients from West China Hospital was used to test whether the findings held up in a completely different population.
For each patient, an expert oncologist carefully traced the outline of the tumor on the CT scan, creating a precise 3D map of the tumor's boundaries. A computer program then extracted 107 different measurements from within that boundary - capturing the tumor's shape (how round or irregular it is), the distribution of brightness within it, and the fine texture patterns at different scales. These types of measurements are called radiomics.
The researchers then built two sets of computer models. The first set asked whether the radiomics measurements could predict each patient's genetic mutations and molecular subtype. The second set asked whether combining radiomics data with actual genetic data, gene activity data, and protein data could predict how long each patient would survive - and whether adding more types of biological data made the prediction better.
All models were trained on one group of patients and tested on a separate group the model had never seen, to ensure the results were not just a coincidence of the specific training data.
The first major finding was remarkable: an AI analysis of CT scan images alone could predict which genetic mutations were present in the tumor with very high accuracy. For three of the most important kidney cancer mutations (VHL, PBRM1, and BAP1), accuracy scores exceeded 0.95 out of 1.0, meaning the computer was nearly as accurate as a laboratory genetic test.
This matters because certain mutations guide treatment decisions. For example, patients with PBRM1 mutations may respond better to certain immunotherapy drugs, while patients with BAP1 mutations tend to have worse outcomes and may not respond well to some targeted therapies. Knowing this before surgery - from a routine CT scan - could help doctors plan the best treatment from the start.
Even more impressively, the computer could also predict which of the four molecular subtypes a patient's tumor belonged to - a classification that previously required analyzing thousands of genes from tumor tissue. All four subtypes were predicted with accuracy scores above 0.95. This appears to be the first time CT scans have been used to predict these gene activity-based subtypes in kidney cancer.
The best-performing AI method was called a Random Forest - a technique that builds hundreds of slightly different decision trees and combines their predictions. This method consistently outperformed seven other approaches that were also tested.
The second major investigation asked whether combining the CT scan measurements with actual genetic laboratory data could better predict how long patients would live. The answer was clearly yes. A CT scan-based model alone predicted 5-year survival with a score of 0.775, which is already quite good. When genetic mutation data was added, the score rose to 0.784. Adding gene activity data (transcriptomics) raised it further to 0.794. Adding protein data (proteomics) pushed it to 0.816.
The model that combined all four types of data - CT scans plus mutations plus gene activity plus protein levels - achieved the best overall prediction, with a 5-year accuracy score of 0.846. This is substantially better than any single data type alone.
Crucially, when this combined model classified patients into high-risk and low-risk groups, the difference in survival was dramatic: high-risk patients were more than six times more likely to die during the study period than low-risk patients. This kind of powerful risk separation is exactly what doctors need to make meaningful treatment decisions.
One key insight from this work is that CT scan data and genetic data are providing different types of information - they are not duplicating each other. CT scans capture the physical appearance of the whole tumor (shape, texture, internal complexity), while genetic data captures the molecular drivers at the cellular level. Because they measure different aspects of the cancer, combining them gives a more complete picture.
A critical test of any research finding is whether it holds up when applied to patients who were not part of the original study. The researchers tested the CT scan-based model on 175 patients from West China Hospital in China - a completely separate group with somewhat different characteristics (on average younger, and with a higher proportion of less aggressive tumors).
Even in this different population with different tumor profiles, the CT scan analysis still predicted survival meaningfully, with a 5-year accuracy score of 0.755. Patients flagged as high-risk by the model were twice as likely to die during the study period as those flagged as low-risk. This cross-institutional validation is an important sign that the findings are real and generalizable, not just specific to the original dataset.
Unfortunately, the detailed genetic and protein data available in the U.S. database was not available for this Chinese cohort, so the combined model could only be tested using CT measurements alone. Future studies that collect all types of data at multiple hospitals will be needed to fully validate the multi-data approach.
The fact that CT measurements alone were predictive across both populations gives confidence that the imaging signals being captured - the shape and texture patterns of kidney tumors - reflect genuine biological differences that are consistent regardless of which hospital's scanner was used or which country the patient was in.
If these findings are confirmed in larger future studies, the implications for patients could be significant. Before any surgery, a CT scan (which patients already have routinely to assess their kidney mass) could potentially reveal not only the size and location of the tumor, but also its genetic character - which mutation it carries, which molecular subtype it is, and how aggressive it is likely to be.
This information could help doctors make more informed decisions before the operation: how urgently surgery needs to be done, whether a less extensive operation might be appropriate, whether the patient should be enrolled in a clinical trial, and which drugs should be considered immediately after surgery. Today, much of this molecular information only becomes available after the tumor has been removed and analyzed in a laboratory - sometimes weeks later.
A particularly important application is for patients who are not good candidates for surgery, where a biopsy is risky or technically difficult, or where the tumor is being monitored rather than immediately removed. In these cases, a scan-based molecular assessment could guide decisions without requiring any additional invasive procedure.
The researchers acknowledge that the study needs to be validated prospectively (following patients forward in time, rather than looking backward at historical records) and in multi-center settings with larger patient numbers before this approach could be used in clinical practice. This is common at this stage of research - the findings provide a strong scientific basis to justify those next steps.
This study demonstrated two important things. First, that a CT scan - analyzed with AI - can predict with remarkable accuracy which genetic mutations are present in a kidney cancer tumor, and which molecular subtype that tumor belongs to. This was achieved without any tissue sampling, purely from patterns in medical images.
Second, that combining CT imaging data with genetic, gene activity, and protein data improves predictions of patient survival substantially. A model that uses all four types of information can separate patients into groups whose survival outcomes differ by more than sixfold - which is a powerful tool for medical decision-making.
The concept of reading a tumor's molecular identity from medical images has been called radiogenomics, and this study is one of the most comprehensive demonstrations of it in kidney cancer to date. If this approach matures, CT scans could evolve from being purely anatomical tools (showing where the tumor is) into molecular diagnostic tools (showing what kind of tumor it is, at the genetic level).
Much work remains before this becomes clinical reality - including larger prospective studies at multiple hospitals. But this research provides strong proof-of-concept evidence that the idea is scientifically sound and worth pursuing, with the goal of giving every kidney cancer patient better, earlier, and less invasive access to the molecular information that could guide their care.