Not every kidney mass discovered on a scan is cancer. Renal hamartoma without visible fat (RH-WVF), also called angiomyolipoma without visible fat, is the most common benign kidney tumor in clinical practice. When these tumors do not contain enough fat to be identified by standard CT imaging, they can appear almost identical to early-stage kidney cancer on scans, making the correct diagnosis extremely difficult for radiologists.
The tumor they are most commonly confused with is homogeneous clear cell renal cell carcinoma (hm-ccRCC), a malignant kidney tumor without obvious internal breakdown, fluid-filled areas, or bleeding. Because both tumor types enhance similarly on CT imaging and appear as solid masses, conventional imaging alone is often insufficient to tell them apart before surgery.
This diagnostic confusion has serious clinical consequences. Patients with benign RH-WVF who are mistakenly suspected of having cancer may undergo unnecessary surgery with its associated risks, recovery time, and potential loss of kidney function. Conversely, a patient with early ccRCC that is incorrectly classified as a benign hamartoma may face delayed treatment, allowing the cancer time to progress or spread.
To resolve this challenge, this study developed and validated a non-invasive preoperative diagnostic model that combines two types of molecular data: radiomic features extracted from CT scans and protein biomarkers measured in urine samples. By integrating these two data types with machine learning, the researchers aimed to create a reliable tool that can distinguish benign from malignant kidney masses before a patient ever has surgery.
The study retrospectively analyzed 371 patients from two hospitals in China (January 2015 to February 2024), all with confirmed diagnoses of either RH-WVF or hm-ccRCC based on surgical specimens. Patients were randomly divided 70/30 into training (259 patients) and validation (112 patients) cohorts, with stratification by tumor stage to ensure balanced clinical characteristics between groups.
CT-based radiomic features were extracted from two imaging phases: the corticomedullary phase and the parenchymal (nephrographic) phase of contrast-enhanced CT. Using the Pyradiomics software, a total of 1,050 radiomic features were extracted from each tumor region, including 14 shape features, 18 first-order statistical features, and hundreds of texture features from gray-level co-occurrence matrices, wavelet transforms, and Gaussian-Laplacian filters.
In parallel, urine samples from matched patients with RH-WVF and hm-ccRCC were analyzed using advanced mass spectrometry (timsTOF Pro in PASEF mode), which identified proteins present in the urine that differed significantly between the two tumor types. 113 differentially expressed proteins were found, and the top candidates were validated using ELISA (enzyme-linked immunosorbent assay) blood and urine tests. The three most promising urinary protein biomarkers selected for the final model were S100A14, FABP4 (fatty acid binding protein 4), and C3 (complement component 3).
LASSO regression with 10-fold cross-validation was used to narrow down the radiomic and urinary proteomic features to the most predictive combination. The final model incorporated a radiomics score plus three urinary protein measurements (S100A14, FABP4, and C3) into a nomogram (a visual scoring tool) and separately into a decision tree model. Both were compared using ROC analysis, calibration curves, and decision curve analysis.
The nomogram model combining the CT radiomics score with the three urinary protein biomarkers achieved an AUC of 0.889 (95% CI: 0.832 to 0.946) in the training cohort and an AUC of 0.895 (95% CI: 0.838 to 0.952) in the independent validation cohort. The consistency of performance between training and validation sets confirms that the model generalizes well to new patients, a critical requirement for clinical usefulness.
The nomogram significantly outperformed the decision tree model, which achieved AUCs of 0.821 and 0.808 in the training and validation cohorts respectively. Calibration curve analysis showed that the predicted probabilities from the nomogram aligned closely with actual observed diagnoses (Hosmer-Lemeshow test P = 0.573), meaning the model's numerical predictions accurately reflect real-world probability of each diagnosis.
Of the four individual predictor variables in the final nomogram, all four were independently significant in multivariate analysis. The S100A14 urinary protein had the highest odds ratio (OR = 3.27), meaning it was the single most powerful individual predictor of malignant hm-ccRCC versus benign RH-WVF. The radiomics score (OR = 2.29) and C3 (OR = 2.87) were also strongly predictive, while FABP4 (OR = 1.26) contributed a more modest but still significant effect.
Applying the nomogram with a threshold of 120 total points (corresponding to a 65% predicted probability of malignancy) for risk stratification, approximately 32% of validation cohort patients (36 patients) in the low-risk group could have safely avoided surgery. The false negative rate in this low-risk group was only 4.7% (5 patients), meaning 95.3% of the patients classified as low risk were truly benign, a clinically acceptable miss rate for a non-invasive preoperative test.
The discovery of three urine protein biomarkers (S100A14, FABP4, and C3) that reliably distinguish benign from malignant kidney masses is a scientifically meaningful finding. Urine is an easily obtainable, non-invasive sample that contains proteins released by kidney tissue, making it an ideal source of cancer biomarkers. This study is among the first to demonstrate that urinary proteomics can be used to differentiate between specific types of kidney masses before surgery.
S100A14, the strongest predictor in the model, belongs to a family of calcium-binding proteins that regulate cell movement and signal communication between cells. Previous studies have linked the related proteins S100A8 and S100A9 to kidney cancer detection, and this study confirms that S100A14 in urine can similarly distinguish between benign and malignant kidney masses with high predictive value.
FABP4 (fatty acid binding protein 4) is a protein involved in lipid metabolism and fat transport, known to be highly expressed in fat cells and macrophages. In clear cell kidney cancer, lipid metabolism plays a central role in tumor biology, and FABP4 levels in urine reflect the distinct lipid metabolic activity of cancer cells compared to the smooth muscle and blood vessel tissue that makes up benign hamartomas.
C3 (complement component 3) is a key protein in the immune system's complement cascade, a molecular pathway that helps identify and destroy abnormal cells. Prior research has found that complement system proteins C3, C3AR1, and C5 are involved in multiple cancer types, and elevated C3 in the urine of hm-ccRCC patients reflects the immune system's response to the malignant tumor, which differs from the benign inflammatory environment around hamartoma tissue.
The most direct clinical benefit of this model is its potential to identify a subgroup of patients with radiologically ambiguous kidney masses who can be safely managed without surgery. In this study, 32% of validation cohort patients were classified as low risk, and 95.3% of them were truly benign, suggesting that over one-third of currently operated patients with these ambiguous masses could potentially avoid surgery if this tool were used clinically.
For patients with benign RH-WVF, surgery carries real risks: complications from general anesthesia, bleeding, infection, and potential long-term reduction in kidney function. If a reliable non-invasive test could confidently classify a mass as benign, those patients could instead be managed with active surveillance, monitoring the mass over time with repeat imaging to ensure it remains stable.
The model was developed into an online clinical decision support tool using the DynNom software package, making it accessible to clinicians without requiring specialized data analysis skills. A doctor can input a patient's radiomics score and the results of three standard urine protein measurements (S100A14, FABP4, C3) to receive an immediately calculated predicted probability of malignancy, visualized through the nomogram framework.
Patients classified as high risk by the nomogram (above the 120-point threshold) would still be recommended for surgery or biopsy, ensuring that the model does not compromise care for patients who truly need aggressive evaluation. The combination of high sensitivity for cancer detection and the ability to spare low-risk patients from surgery represents exactly the kind of precision medicine decision support that modern kidney cancer management needs.
This study is among the first to combine CT radiomics with urine proteomics in a machine learning framework for preoperative kidney mass diagnosis. The multi-omics approach achieved better predictive performance (AUC 0.895) than either data type could achieve alone, demonstrating the value of integrating complementary biological information from imaging and liquid biopsy in a single clinical decision tool.
The discovery and validation of S100A14, FABP4, and C3 as urinary protein biomarkers for distinguishing benign hamartomas from malignant kidney cancer represents a meaningful contribution to kidney cancer biology. These proteins have established biological roles in cancer pathways, giving mechanistic credibility to their diagnostic utility and suggesting potential avenues for future therapeutic investigation.
The study has important limitations. It is retrospective, covering a 10-year period at only two hospitals, which may introduce selection bias. The manual tumor segmentation required for radiomic feature extraction is time-consuming and may not be feasible in all clinical settings. The urine sample collection protocol involved careful biobank storage for older samples, which could introduce variability. Larger, multicenter prospective validation studies are planned to confirm the findings.
If validated prospectively, this model could become a standard preoperative tool that significantly improves the management of radiologically ambiguous kidney masses. Patients whose masses are predicted to be benign could avoid unnecessary surgery, while those with high predicted malignancy could be triaged for early surgical intervention. The development of automated segmentation tools and validated clinical ELISA kits for S100A14, FABP4, and C3 measurement will be essential next steps toward routine clinical adoption.