A predictive model for WHO/ISUP pathologic grading of renal clear cell carcinoma based on CT radiomics: a multicenter study

Abstract

Objective: This study aims to evaluate the predictive value of CT radiomics combined with clinical-imaging features for the WHO/ISUP pathological grade of clear cell renal cell carcinoma(ccRCC).

Methods: In this multicenter retrospective study enrolled 169 patients (110 males, 59 females) with pathological confirmed ccRCC between November 2017 and February 2022. Based on the WHO/ISUP pathological grading criteria, patients were stratified into two groups: low-grade (grades I-II, n = 93) and high-grade (grades III-IV, n = 76). Three-dimensional tumor segmentation was performed on CT cortical-phase images using ITK-SNAP software. The segmented data were subsequently processed through the United Imaging Intelligent Scientific Research Platform for radiomic feature extraction and selection. Logistic regression analyses were conducted to identify independent predictive factors. Based on these factors, an optimized predictive model was developed through random forest classification and evaluated using calibration curves, ROC analysis, Delong test and decision curve analysis.

Results: Significant differences were observed in tumor size, morphology, hemorrhage, necrosis, tumor thrombus, capsular invasion and tumor extending beyond the renal margin between the two groups. Logistic regression analysis identified tumor size, hemorrhage and tumor thrombus as independent predictors. Six radiomic features were selected to establish prediction model. In the training cohort, the combined model demonstrated superior discriminative performance, achieving an AUC of 0.895, compared to the radiomics model (AUC = 0.873) and the clinical-imaging model (AUC = 0.712). The combined model also exhibited strong predictive ability in both the validation cohort (AUC = 0.885) and the external cohort (AUC = 0.860). DeLong tests revealed statistically significant differences in AUC between the combined model and the clinical-imaging model (Z = 3.023, P = 0.002), as well as between the radiomics model and the clinical-imaging model (Z = 2.560, P = 0.010). However, no significant difference in AUC was found between the combined model and the radiomics model (Z = 1.627, P = 0.103). Decision curve analysis revealed the combined model yielded enhanced net benefit across threshold ranges (0.1-0.84).

Conclusion: The combined model based on CT radiomics combined with clinical- imaging features can effectively predict the WHO/ISUP pathological grade of ccRCC, providing more basis for the prognosis assessment of patients.

Keywords: Computed tomography; Predictive model; Radiomics; Renal clear cell carcinoma.

Fig. 1

The flowchart of patient recruitment process

Fig. 2

Image omics characteristics of the training set obtained by LASSO screening. (a) λ parameter selection plot; (b) Feature coefficient change plot; (c) Six selected optimal feature weight plots; (d) Violin diagram of maximum weight features

Fig. 3

ROC curve, calibration curve, and decision curve of WHO/ISUP grading model (e-g). e. ROC curve (ROC curves were used to assess discrimination between high and low grade ccRCC by predictive models). f. Combined model calibration curve. (Calibration data show the relationship between predicted risk and actual risk). g. Decision curve (the decision curve was used to observe whether the model has clinical effectiveness)