Clear cell renal cell carcinoma (ccRCC) is the most common type of kidney cancer. While surgery works well for early-stage disease, advanced ccRCC often requires additional treatment such as immune checkpoint inhibitor (ICI) therapy, which helps the immune system recognize and attack cancer cells.
Kidney cancer is considered highly immunogenic, meaning it tends to attract immune cells into and around the tumor. However, the level of this immune activity varies greatly between patients, and that variation affects how well immunotherapy works. Patients with higher levels of immune infiltration appear to respond better to ICI drugs such as anti-PD-1 antibodies.
Traditionally, measuring immune infiltration required an invasive biopsy or analysis of surgically removed tissue. This makes it difficult to monitor changes over time. This study aimed to develop a non-invasive way to predict immune infiltration levels using standard CT scan images.
Researchers began by analyzing genetic data from 539 ccRCC tumor samples from The Cancer Genome Atlas (TCGA), a large public cancer database. They used an algorithm called ssGSEA (Single Sample Gene Set Enrichment Analysis) to measure the activity levels of 29 different types of immune cells within each tumor.
Using these immune cell measurements, researchers created an ICI score -- a single number that summarizes the overall level of immune infiltration in a sample. They then used the Boruta algorithm to identify which genes were most important for distinguishing high from low immune infiltration.
Samples were divided into two groups: label 1 (high immune infiltration, 68 cases) and label 0 (low immune infiltration, 95 cases). These labels were later matched with CT scan images from the same patients to train machine learning models.
CT scan images for 163 patients (whose data existed in both the imaging and genetic databases) were collected. Two experienced urologists manually outlined the tumor region in each CT image using 3D Slicer software -- a technique called region of interest (ROI) segmentation.
From these outlines, the researchers extracted 1,832 radiomics features using a specialized software tool. Radiomics features are mathematical measurements that capture texture, shape, and intensity patterns within the tumor that are invisible to the naked eye. These features describe how pixels relate to each other and can reflect internal tumor biology.
After filtering for relevance and removing redundant features using Pearson correlation and LASSO regression, 7 highly informative features were selected. Three came from the GLSZM category (measuring how uniform different sized zones of similar intensity are) and others from GLCM and NGTDM categories, all describing texture complexity within the tumor.
Eight machine learning algorithms were tested on the 7 selected CT features to predict whether a tumor had high or low immune infiltration. These included Logistic Regression, Support Vector Machine, K-Nearest Neighbour, Random Forest, ExtraTrees, XGBoost, LightGBM, and Multilayer Perceptron.
The ExtraTrees model achieved the best overall performance, with an AUC (Area Under the ROC Curve) of 0.753 on the test set and 1.000 on training -- meaning it correctly distinguished high from low immune infiltration in about 75% of new cases. It also had a specificity of 0.842, meaning it was good at correctly identifying low-infiltration tumors.
Decision curve analysis confirmed that the ExtraTrees model provided meaningful clinical benefit across a wide range of decision thresholds (10-80%), suggesting it could help doctors in real treatment planning scenarios. LightGBM and Logistic Regression also showed good sensitivity, correctly identifying many high-infiltration tumors.
Statistical analysis using multifactorial Cox regression showed that the ICI score, combined with clinical factors like age, cancer stage, and tumor grade, significantly predicted how long patients lived. When ICI score was used alone, its 1-year survival AUC was 0.71; combined with clinical factors, this improved to 0.84.
A nomogram (a visual chart) was created to help clinicians estimate individual patient survival probabilities based on the ICI score and clinical factors. This tool translates complex statistical results into a format usable at the bedside.
Researchers also analyzed immunotherapy response using the TIDE algorithm and found that patients in the high immune infiltration group were significantly more likely to respond to anti-PD-1 therapy. This confirms that the immune infiltration level is not just a survival marker but also a predictor of treatment benefit.
Most radiomics studies use contrast-enhanced CT, which requires an injected dye to highlight blood vessels and tumors. However, contrast agents can cause problems for patients with impaired kidney function, which is common in kidney cancer patients.
This study deliberately used plain (non-contrast) CT scans, which are more widely available, less expensive, and safer for patients who cannot receive contrast agents. This design choice makes the model more practical and accessible in real-world clinical settings.
Using non-contrast CT also means the model could be applied to patients who already had a plain CT performed, without needing an additional imaging session. This approach addresses a significant gap, since fewer studies have used plain CT for radiogenomics in kidney cancer.
This study demonstrates that CT-based radiomics combined with machine learning can predict the level of immune cell infiltration inside a clear cell kidney tumor without any biopsy or invasive procedure. This represents an important step toward truly non-invasive tumor profiling.
For patients, this could eventually mean that a routine CT scan performed before treatment would also provide guidance on whether immunotherapy is likely to work well for them. Currently, this decision often requires waiting for biopsy results or relying on limited clinical information.
The authors caution that the study is limited by its relatively small size (163 cases with matched imaging and genetic data) and the lack of an external validation set. Larger prospective studies including patients from multiple hospitals will be needed before this approach can be used routinely in clinics.