Clear cell renal cell carcinoma (ccRCC) is the most common kidney cancer. One of the most important pieces of information about this cancer is its WHO/ISUP grade, a system developed by the World Health Organization and the International Society of Urological Pathology that classifies tumors from Grade I (least aggressive) to Grade IV (most aggressive) based on how the cancer cells look under a microscope.
Currently, tumor grade is only confirmed after surgery by examining removed tissue. But knowing the grade before surgery would be enormously valuable: it would help surgeons decide how radical the operation needs to be, help oncologists plan additional treatments, and give patients a clearer understanding of their prognosis before they go under the knife.
Traditional biopsies can determine grade before surgery but are invasive and carry risks including bleeding and tumor seeding. This study developed a completely non-invasive AI model based on CT scans that can predict tumor grade before any tissue is removed, using only images the patient has already had as part of routine cancer workup.
The study analyzed 1,499 histologically confirmed ccRCC patients from two medical centers, divided into a training cohort (929 patients), an internal validation cohort (398 patients), and an external validation cohort (172 patients from the second center). This rigorous three-way split is considered best practice for validating AI medical tools.
Three different phases of CT imaging were used: NCP (non-contrast phase), CMP (corticomedullary phase), and NP (nephrographic phase). Each phase captures different aspects of the tumor's appearance as contrast dye moves through the blood vessels, providing complementary information about the tumor's structure and blood supply.
The model integrates three types of information into a single prediction tool called a DLRN (Deep Learning Radiomics Nomogram). First, radiomics: a computer extracts 5,502 mathematical features from the shape, texture, and intensity patterns of the tumor in the CT images. Second, deep learning features: a DenseNet201 neural network automatically learns additional visual patterns from the images. Third, clinical data: standard patient information such as age, sex, and tumor size is incorporated.
The DLRN (integrated model) achieved an AUC of 0.935 in training, 0.901 in internal validation, and 0.911 in external validation. The AUC measures how well the model distinguishes low-grade from high-grade tumors, with 1.0 being perfect. These results are consistently high across all three cohorts, demonstrating the model is robust and generalizes well to new patients.
Critically, the integrated DLRN significantly outperformed each of its three components when used alone. The clinical-only model achieved an AUC of only 0.730, the radiomics-only model reached 0.845, and the deep learning-only model achieved 0.868. The DLRN surpassed all of these because the three information sources each capture different aspects of tumor biology that complement each other.
In the external validation cohort, the model demonstrated an accuracy of 83.1%, with a sensitivity of 75.0% (correctly identifying 75% of high-grade tumors) and a negative predictive value (NPV) of 92.8%. The high NPV means that when the model says a tumor is low grade, it is correct almost 93% of the time, a particularly useful property for reassuring patients and guiding conservative surgical approaches.
SHAP (SHapley Additive exPlanations) analysis revealed that the deep learning features from DenseNet201 were the dominant contributors to the model's predictions. This means the neural network's ability to recognize complex visual patterns in the CT images contributes more than the hand-crafted mathematical features or clinical variables alone.
To understand what the deep learning model was actually looking at in the images, the researchers used Grad-CAM (Gradient-weighted Class Activation Mapping), a technique that highlights which parts of an image most influenced the model's prediction. Grad-CAM maps showed the model focused on regions of tumor heterogeneity, areas where the cancer tissue appears irregular or mixed, which pathologists know correlate with higher grade tumors.
This alignment between what the AI focuses on (tumor heterogeneity) and what experienced radiologists and pathologists look for when assessing grade provides strong evidence that the model is learning clinically meaningful features rather than spurious correlations in the data.
For patients, this technology could change the pre-surgical conversation significantly. Instead of knowing only that they have kidney cancer, patients and their surgeons could receive a reliable estimate of tumor grade before the operation, enabling a more informed discussion about surgical options ranging from partial nephrectomy (kidney-sparing) to radical nephrectomy.
High-grade tumors identified preoperatively might prompt more aggressive surgical planning or immediate discussion of systemic therapy, while patients with predicted low-grade tumors might be candidates for active surveillance in appropriate cases or more conservative surgical approaches. This individualization of care is a core goal of modern precision oncology.
Because the model uses CT scans that are already performed as part of standard kidney cancer workup, implementation would require no additional imaging or costs for patients. The AI analysis would simply be added to the existing imaging workflow, generating a grade prediction alongside the radiologist's standard report.
This two-center validation study demonstrates that an integrated deep learning radiomics nomogram can predict kidney cancer tumor grade non-invasively with high accuracy. The consistent performance across training, internal validation, and external validation cohorts provides strong evidence that the model is robust and generalizable.
The approach of combining radiomics, deep learning, and clinical information represents a template for building future AI diagnostic tools: each data type contributes unique information, and integration outperforms any single approach alone. This principle is likely applicable to tumor grading and staging problems across many cancer types.
Future directions include expanding validation to more institutions and diverse patient populations, integrating genomic data to further improve accuracy, and designing prospective clinical trials to test whether using this tool to guide treatment decisions actually improves patient outcomes compared to current standard-of-care approaches.