Exploring the predictive value of MRI-based clinical-radiomics models for biochemical recurrence after radical prostatectomy in prostate cancer / 中华放射学杂志

ABSTRACT

Objective:To construct a clinical-radiomics model based on MRI, and to explore its predictive value for biochemical recurrence (BCR) after radical prostatectomy in prostate cancer patients.Methods:A total of 212 patients with prostate cancer who underwent radical prostatectomy in the First Affiliated Hospital of Soochow University from January 2015 to December 2018 and had complete follow-up data were retrospectively analyzed. The random toolkit of Python language was used to randomly sample the patients at a ratio of 7∶3 without replacement, and they were divided into a training set (149 cases) and a test set (63 cases). The endpoint of follow-up was BCR or at least 3 years. BCR occurred in 50 patients in the training group and 21 patients in the test group. The imaging features of the main lesion area in the preoperative T 2WI, diffusion-weighted imaging and apparent diffusion coefficient map of patients in the training set were extracted, and the unsupervised K means clustering algorithm was used to screen the features. The selected features were fitted by a multivariate Cox regression model, and the radiomics model was constructed. Univariate Cox regression analyses were used to screen the main clinical risk factors associated with BCR, and the clinical-radiomics model was constructed combined with RadScore. In the test set, the time-dependent receiver operating characteristic (ROC) curve was constructed, and the area under the curve (AUC) was calculated to evaluate the predictive efficacy of the radiomics model, clinical-radiomics model and prostate cancer risk assessment after radical resection (CAPRA-S) score for the occurrence of BCR. Harrell consistency index (C-index) was used to evaluate the model to predict BCR consistency. The calibration curve was used to evaluate the degree of variation of the model. The decision curve was used to evaluate the clinical application value of the prediction model. Results:A total of 26 radiomics features were screened to establish the radiomics model. The univariate Cox showed that the preoperative clinical features included preoperative prostate-specific antigen level (HR=1.006, 95%CI 1.002-1.009, P=0.001), Gleason score of biopsy (HR=1.422, 95%CI 1.153-1.753, P=0.001), clinical T stage (HR=1.501, 95%CI 1.238-1.822, P<0.001). The multivariate Cox showed that the RadScore was an independent predictor of BCR after radical prostatectomy (HR=51.214, 95%CI 18.226-143.908, P<0.001). The selected preoperative clinical features were combined with RadScore to construct a clinical-radiomics model. In the test set, the AUCs of the time (3 years)-dependent ROC curves of the radiomics model, the clinical-radiomics model, and the CAPRA-S score were 0.824 (95%CI 0.701-0.948), 0.841 (95%CI 0.714-0.968), and 0.662 (95%CI 0.518-0.806), respectively. The C-index of the radiomics model, clinical-radiomics model and CAPRA-S score were 0.784 (95%CI 0.660-0.891), 0.802 (95%CI 0.637-0.912) and 0.650 (95%CI 0.601-0.821), respectively. The calibration curve showed that the predicted probability and actual probability of BCR by radiomics model, clinical-radiomics model and CAPRA-S score were in good agreement (χ 2=7.64, 10.61, 6.37, P=0.465, 0.225, 0.498). The decision curve showed that the clinical net benefit of the clinical-radiomics model and the radiomics model was significantly higher than the CAPRA-S score. When the threshold probability was 0.20-0.30, 0.40-0.50, and >0.55, the clinical net benefit of the clinical radiomics model was higher than that of the radiomics model. Conclusions:The clinical-radiomics model can effectively predict the occurrence of BCR in patients with prostate cancer after radical prostate ctomy, and the prediction efficacy is better than the radiomics model and CAPRA-S score.