Combining clinical characteristics with CT radiomics to predict Ki67 expression level of small renal mass based on artificial intelligence algorithms.

Background: Most small renal masses (SRMs) grow slowly and have good prognosis, but a portion of SRMs can also demonstrate aggressive characteristics, which can be explored by the proliferation-related marker Ki67.

Methods: A total of 241 patients collected from the two centers were included in the study, of which 145 patients from the First Affiliated Hospital of Guangzhou Medical University were divided into training and validation cohort, while 96 patients from Sun Yat-sen Memorial Hospital were served as test cohort. To ensure the class balance of the outcome measures, the training cohort was oversampled, resulting in an increase of 77 cases in the minority class. After variables processing and feature selecting, optimal artificial intelligence-based model was constructed to predict the Ki67 expression level, and the model performance, interpretation and application development was performed.

Results: The baseline characteristics of enrolled patients were described, and no statistically significant differences were found between two centers and cohorts, both before and after oversampling. The optimal model, regularized random forest, was constructed showing AUROC values of 0.802, 0.878, and 0.668, and balanced accuracy of 0.744, 0.808, and 0.679 in the oversampling training, validation, and test cohort, respectively. Model interpretation was performed, and a web application was built.

Conclusions: An artificial intelligence-based predictive model for non-invasively assessing the Ki67 expression level of SRMs was developed, thus providing valuable reference for clinical decision-making in these patients.