Interpretable CT Radiomics-based Machine Learning Model for Preoperative Prediction of Ki-67 Expression in Clear Cell Renal Cell Carcinoma

Abstract

Rationale and objectives: To develop and externally validate interpretable CT radiomics-based machine learning (ML) models for preoperative Ki-67 expression prediction in clear cell renal cell carcinoma (ccRCC).

Methods: 506 patients were retrospectively enrolled from three independent institutes and divided into the training (n=357) and external test (n=149) sets. Ki67 expression was determined by immunohistochemistry (IHC) and categorized into low (<15%) and high (≥15%) expression groups. Radiomics features were extracted from segmented tumor regions in the corticomedullary phase (CMP) CT images using the "PyRadiomics" package. The least absolute shrinkage and selection operator (LASSO) regression was applied to select the most relevant radiomics features for Ki-67 expression, which were then used to train five ML models. Models' performances were evaluated via the receiving operator curve analysis and compared using Delong test. Calibration and decision curve analyses assessed the models' clinical utility. Kaplan-Meier analysis and Log-rank tests were conducted to determine the prognostic value of radiomics-predicted Ki-67 expression status. The optimal model was interpreted using SHapley Additive exPlanations (SHAP).

Results: Eight radiomics feature were selected to build models using Random forest (RF), eXtreme Gradient Boosting (XGBoost), Logistic regression (LR), Support vector machine (SVM), and K-nearest neighbor (KNN). The RF model exhibited the best performance, achieving the highest area under the curve (AUC) in both the training (0.910, 95% confidence interval [CI]: 0.881-0.936) and external test (0.885, 95% CI: 0.826-0.934) sets, as confirmed by Delong test (all P values＜0.05). Calibration and decision curves further demonstrated the superior clinical utility of the RF model. Both IHC-based and RF-predicted high Ki-67 expression groups were significantly associated with a higher risk of tumor recurrence in the training and external test sets (all P values＜0.05). The prediction process of the RF model was uncovered in the globe and individualized terms using the SHAP.

Conclusion: The interpretable CT radiomics-based RF classifier exhibited robust predictive performance in assessing Ki-67 expression levels preoperatively, offering valuable prognostic insights and aiding clinical decision-making in ccRCC patients.

Keywords: Clear cell renal cell carcinoma; Computed tomography; Ki-67; Machine learning; Radiomics.