Clear cell renal cell carcinoma (ccRCC) is the most common type of kidney cancer. One of the most important factors in deciding how to treat it is the cancer's grade - a measure of how aggressive or fast-growing the tumor cells appear under a microscope.
The current standard for grading uses the ISUP (International Society of Urological Pathology) system, which replaced an older system called Fuhrman grading. ISUP grades ccRCC on a scale from 1 to 4 based on features of the tumor cell nuclei. Grade 1-2 tumors are lower-grade (less aggressive), while grade 3-4 are higher-grade (more aggressive) and require more urgent treatment.
The problem is that knowing the grade currently requires a biopsy or surgery to remove tissue for a pathologist to examine under a microscope. This is invasive and carries risks. If we could predict the grade from a CT scan - a routine non-invasive imaging test - doctors could plan treatment earlier and more safely.
This study investigated whether a special type of analysis called radiomics - which extracts hundreds of detailed measurements from CT scan images - combined with artificial intelligence, could accurately predict ISUP grade without surgery.
The study included 227 patients with confirmed clear cell kidney cancer who had CT scans before surgery. The scans were taken in three phases after injecting a contrast dye: early (corticomedullary phase), mid (nephrographic phase), and late (excretory phase). This allowed researchers to see the tumor in different lighting conditions, so to speak.
Radiologists outlined the tumor on the CT images, creating what is called a region of interest (ROI). From this region, a computer program extracted 1,029 different measurements of the tumor's appearance - covering its shape, the distribution of light and dark areas, texture patterns, and more complex mathematical features.
Since 1,029 features is far too many to use all at once, a technique called LASSO was used to filter down to only the most important ones. LASSO is a mathematical method that automatically identifies which handful of features are most useful for predicting the cancer grade.
A Support Vector Machine (SVM) - a type of AI classifier - was then trained on the selected features to predict whether each patient's tumor was low-grade (ISUP 1-2) or high-grade (ISUP 3-4). The model's performance was tested on a separate group of patients it had never seen before.
The best-performing model used features from both CT phases combined (corticomedullary and nephrographic). It achieved an accuracy score (AUC) of 0.91 when tested on new patients - meaning it correctly distinguished low-grade from high-grade tumors about 91% of the time in the test group.
In the validation (testing) group, the model had a sensitivity of 83% (it correctly identified most high-grade tumors) and a specificity of 89% (it correctly identified most low-grade tumors). This balance is important for a clinical tool - you want to catch dangerous cancers without over-alarming patients with low-grade disease.
When only one CT phase was used, the model did not perform as well. The nephrographic phase alone had a much lower accuracy of 0.56, while the corticomedullary phase alone reached 0.87 but had weaker specificity. Combining both phases gave the most reliable results.
The LASSO algorithm narrowed the 1,029 original measurements down to just 7 key features for the best model. These 7 features - including measurements of texture and gray-level patterns - contained most of the information needed to predict grade.
The AI-based model outperformed earlier studies that used older grading systems or simpler imaging features. One reason is that this study used the newer, more reliable ISUP grading system, which has clearer distinctions between grades compared to the older Fuhrman system - making it easier for the AI to learn the difference.
The texture features the model uses - measuring how uniform or varied the pixel patterns are inside the tumor - reflect something real and biologically meaningful. Higher-grade cancers tend to have more irregular, heterogeneous tissue with more areas of necrosis (dead tissue) and abnormal blood vessel growth, all of which show up as irregular patterns on CT.
The researchers also note that the CT scan approach is comparable in accuracy to an image-guided biopsy, but without the need for a needle - reducing the risk of complications such as bleeding, infection, or tumor seeding.
One limitation is the relatively small number of patients, especially in the high-grade group (only 72 patients). A larger study would be needed to confirm these results and to make the model ready for routine clinical use. The study also used 2D cross-sections of the tumor rather than analyzing the entire 3D volume.
If this type of tool were adopted in clinics, a patient with a kidney mass found on a routine scan could potentially learn whether the tumor is likely low-grade or high-grade before any surgery or biopsy. This information could guide important decisions about next steps.
For patients with lower-grade tumors, doctors might recommend active surveillance - watching and waiting - rather than immediate surgery. For higher-grade tumors, prompt surgery would be recommended. This avoids unnecessary procedures for patients who don't need them urgently.
The approach is also safe because it relies entirely on CT scans that patients often already have as part of their initial evaluation. No additional radiation, dye, or procedures would be required.
The researchers also point out that as immunotherapy becomes more important in kidney cancer treatment, understanding a tumor's characteristics before surgery could help predict how it will respond to different treatments - an exciting direction for future research.
This study demonstrates that a computer model trained on CT scan measurements can reliably distinguish between low-grade and high-grade kidney cancer, achieving an impressive AUC of 0.91 in the validation group - better than several previous approaches.
The key to the model's success was combining features from two CT imaging phases and using AI to select only the 7 most informative measurements out of over 1,000, building a lean and robust classifier.
The next steps are to test this approach in more patients from different hospitals and with different CT scanners to confirm it works broadly. Larger, multi-center studies are needed before this can become a standard clinical tool.
This research is part of a growing movement to use AI and medical imaging to give patients and doctors better information earlier in the cancer journey - without adding risk or discomfort to the diagnostic process.