A radiomics-based model for the prediction of WHO/ISUP in clear cell renal cell carcinoma using contrast-enhanced CT indicating response to TKI therapy.

Abstract

e16522 Background: WHO/ISUP grade is a significant risk factor for the prognosis of patients with clear cell renal cell carcinoma (ccRCC) and effects the response to tyrosine kinase inhibitors (TKIs) for advanced-stage patients. The purpose of this study was to develop a fully-automated model that can predict WHO/ISUP grade based on three-phase CT images, and may implicate the TKIs response. Methods: A total of 373 patients with ccRCC from three medical centers were retrospectively included in the study, with 261 in the training set and 112 in the testing set. CT images of 166 TCGA-KIRC cohort were used to explore the different expressed genes and enriched biological pathways related to the radiomics model. All CT phases were aligned to the venous phase and used to evaluate a presenting deep learning model (Kidney and kidney tumor segmentation 2023, KiTS23) for kidney tumor segmentation. Radiomics features were extracted from the tumor of original CT phases. Linear discriminant analysis was used to develop three models based on transcriptomic features, radiomics features, and both features combined. Models were evaluated by area under curve, sensitivity, and specificity. Results: The average dice coefficients of kidney tumor segmentation were 0.87 in the training set and 0.83 in the testing set. For WHO/ISUP grade prediction, in the testing set, the model based on radiomics (AUC = 0.801) outperformed the model based on transcriptomic features (AUC = 0.783). The hybrid model based on transcriptome and radiomics features achieved the best performance in both the training set (AUC = 0.911) and testing set (AUC = 0.859). Moreover, the hybrid model also provided the highest accuracy (0.930), sensitivity (0.714), specificity (0.972), positive predictive value (0.833), and negative predictive value (0.946). The TCGA-KIRC cohort were divided into high- and low-risk group based on the radiomics model prediction, and the differential expressed gene in high-risk group were significantly enriched on the pathway of EGFR tyrosine kinase inhibitor resistance. Conclusions: The fully-automated model based on transcriptome and radiomics features can accurately predict the WHO/ISUP grade of patients with ccRCC, and implicate the TKIs response for advanced-stage patients.