Clear cell renal cell carcinoma (ccRCC) is the most common type of kidney cancer. One of the most important factors for predicting how aggressive the cancer will behave is its WHO/ISUP grade, a score from 1 to 4 assigned by a pathologist after examining the tumor under a microscope. Higher grades mean more aggressive cancer and worse outcomes.
Getting an accurate grade before surgery is valuable -- it can help surgeons and oncologists plan treatment. But current imaging techniques only look at the tumor itself. This study asked a different question: can the fat tissue surrounding the kidney (called perirenal adipose tissue, or PRAT) give us additional information about tumor grade?
There is growing scientific evidence that adipose (fat) tissue around a tumor can influence cancer behavior through inflammation, hormone signaling, and other biological pathways. Researchers hypothesized that changes in PRAT texture visible on CT scans might reflect these biological interactions and help predict grade.
The researchers enrolled 614 patients with ccRCC from multiple hospitals, creating a robust dataset spanning training (383 patients), internal validation (88 patients), and external validation (143 patients from the public KiTS19 dataset). All patients had pre-surgery CT scans and confirmed pathology results.
A key innovation was the use of TransUNet, an advanced AI model that automatically identifies and outlines the kidneys, tumors, and visceral fat on CT images. From this segmentation, a special algorithm expanded the kidney boundary outward by 6 voxels to define the perirenal adipose tissue ring for analysis.
The team then extracted radiomics features -- hundreds of numerical measurements describing texture, shape, and intensity patterns -- from three regions: the tumor itself, the surrounding kidney tissue, and the PRAT. A support vector machine (SVM) classifier was trained to predict whether tumors were low grade (WHO/ISUP 1-2) or high grade (WHO/ISUP 3-4).
The results demonstrated clear value in including PRAT. The combination model -- using features from the tumor, kidney, and perirenal fat together -- achieved an AUC of 0.814 in internal validation, significantly better than the tumor-only model (AUC 0.760). The difference was statistically significant (p = 0.034 by DeLong's test).
In the external validation cohort using the KiTS19 dataset, the combination model similarly outperformed tumor-only analysis, confirming that the benefit was real and not specific to one hospital's patient population. This external validation is critical -- it means the model generalizes to new patients from different centers.
Radiomics features from PRAT that were most informative included texture measures related to heterogeneity and surface irregularity of the fat, suggesting that biological changes in the fat surrounding a high-grade tumor are detectable on CT even without specialized scans.
TransUNet is a hybrid deep learning architecture that combines a convolutional neural network (CNN) for capturing local image features with a Transformer component for understanding long-range relationships across the image. This combination makes it especially effective for medical image segmentation tasks like outlining kidneys and tumors on CT scans.
In this study, TransUNet was trained to simultaneously segment the kidney, tumor, and body composition regions. The perirenal fat was then defined algorithmically by expanding the kidney mask by 6 voxels -- a standardized approach that ensures each patient's PRAT is defined consistently, regardless of anatomy.
This level of automation is important for clinical translation: manual segmentation by radiologists takes significant time and has operator variability. An automated pipeline that can reliably extract tissue regions from routine CT scans would allow this analysis to be run on any patient without additional effort.
Currently, WHO/ISUP grade is only confirmed after a kidney is removed and the tumor examined under a microscope. Pre-surgical grading is challenging -- biopsy carries risks, and imaging alone has limited accuracy. Better non-invasive grading tools could help patients and doctors make more informed decisions.
For example, a patient with a suspected high-grade tumor might be counseled differently about the urgency of surgery, the type of procedure best suited to their case, and the likely need for post-surgery surveillance or systemic therapy. Conversely, identifying a likely low-grade tumor could support active surveillance in appropriately selected patients.
This study suggests that routine pre-operative CT scans already contain more predictive information than we currently extract. By analyzing the tumor alongside its surrounding tissue environment, AI-assisted radiomics could provide a richer pre-surgical picture -- at no additional cost or radiation exposure to the patient.
The authors acknowledge several important limitations. The study was retrospective, meaning data was collected from existing records rather than a prospectively designed trial. All patients had surgery, which may introduce selection bias since lower-risk patients might have been managed differently. The optimal PRAT region definition (6 voxels dilation) was chosen empirically and may not be universally optimal.
The external validation used the KiTS19 public dataset, which was originally assembled for a segmentation challenge rather than a grading study. This introduces some uncertainty about whether the grading labels and imaging protocols are fully comparable to the main cohort.
Despite these limitations, this is one of the largest multicenter studies to demonstrate that perirenal fat carries independent predictive value for ccRCC grading beyond the tumor itself. Future prospective studies and integration with clinical decision support tools will be needed to translate these findings into routine care.