Subregion-based radiomics analysis for predicting the histological grade of clear cell renal cell carcinoma.

PURPOSE

We explored the feasibility of constructing machine learning (ML) models based on subregion radiomics features (RFs) to predict the histological grade of clear cell renal cell carcinoma (ccRCC) and explore the molecular biological mechanisms associated with RFs.

METHODS

Data from 186 ccRCC patients from The Cancer Imaging Archive (TCIA) and 65 ccRCC patients from a local hospital were collected. RFs were extracted from entire tumor regions and subregions, which were segmented via a Gaussian mixture model (GMM). ML models and radiomics scores (radscores) were developed on the basis of candidate RFs. A RFs-related gene module was identified. Key signaling pathways were enriched, and hub genes were identified.

RESULTS

Two subregions were segmented. The logistic regression (LR) and support vector machine (SVM) models constructed using 3 candidate RFs selected from subregion 1 demonstrated the best predictive performance, with AUCs of 0.78 and 0.77 for the internal test set and 0.74 and 0.77 for the external validation set, respectively. Radscores stratified ccRCC patients into high- and low-risk groups, with high-risk individuals exhibiting poorer overall survival (OS) for the internal test set. Radiogenomic analysis revealed that RFs were associated with signaling pathways related to cell migration, cell adhesion, and signal transduction. The hub genes CTNNB1 and KDR were identified as being associated with RFs.

CONCLUSION

We revealed an association between RFs and tumor biological processes. The proposed subregional radiomics models demonstrated potential for predicting the histological grade of ccRCC, which may provide a novel noninvasive predictive tool for clinical use.