AI-driven radiogenomic analysis of clear cell renal cell carcinoma: perinephric adipose tissue stranding as a key feature of the NIPAL4-associated imaging pattern

Background: Radiogenomics offers a non-invasive approach to correlate imaging features with tumor molecular profiles. This study aims to identify computed tomography (CT) imaging characteristics associated with positive NIPA-like domain containing 4 (NIPAL4) expression in clear cell renal cell carcinoma (ccRCC) and to develop a radiogenomic predictive model to support personalized risk stratification. Methods: A retrospective analysis was conducted on 241 ccRCC patients from The Cancer Genome Atlas (TCGA) and The Cancer Imaging Archive (TCIA) databases. Clinical, pathological, and CT features were compared between NIPAL4- positive and NIPAL4-negative groups. A penalized logistic regression model was built to predict NIPAL4 expression, and its performance was assessed using Receiver Operating Characteristic (ROC) and Decision Curve Analysis (DCA). Additionally, unsupervised K-means clustering was used to identify radiologic phenotypes, and a nomogram was developed to enable individualized risk estimation. Results: Among 241 ccRCC patients, 29 (12.03%) showed positive NIPAL4 expression. Compared to NIPAL4-negative cases, positive expression was significantly associated with larger tumor size (median 70.5 mm vs. 52 mm, p = 0.0371), illdefined margins (61.5% vs. 32.4%, p = 0.0077), perinephric adipose tissue stranding (76.9% vs. 50.0%, p = 0.0114), renal vein thrombosis (24.0% vs. 4.7%, p = 0.021), Gerota’s fascia thickening (61.5% vs. 35.2%, p = 0.0163), and collecting system invasion (52.0% vs. 26.5%, p = 0.0171). A multivariate penalized logistic regression model incorporating these features achieved an AUC of 0.973% and 92.1% accuracy in predicting NIPAL4 positivity. Conclusions: Positive NIPAL4 expression in ccRCC is significantly associated with aggressive CT features—particularly perinephric adipose tissue stranding, ill-defined margins, and renal vein thrombosis. A radiogenomic model based on these features achieved excellent predictive performance (AUC = 0.973), supporting its potential role in noninvasive risk stratification and personalized clinical decisionmaking.