Preoperative prediction of renal fibrous capsule invasion in clear cell renal cell carcinoma using CT-based radiomics model

Abstract Objectives To develop radiomics-based classifiers for preoperative prediction of fibrous capsule invasion in renal cell carcinoma (RCC) patients by CT images. Methods In this study, clear cell RCC (ccRCC) patients who underwent both preoperative abdominal contrast-enhanced CT and nephrectomy surgery at our hospital were analyzed. By transfer learning, we used base model obtained from Kidney Tumor Segmentation challenge dataset to semi-automatically segment kidney and tumors from corticomedullary phase (CMP) CT images. Dice similarity coefficient (DSC) was measured to evaluate the performance of segmentation models. Ten machine learning classifiers were compared in our study. Performance of the models was assessed by their accuracy, precision, recall and area under the receiver operating characteristic curve (AUC). The reporting and methodological quality of our study was assessed by the CLEAR checklist and METRICS Score. Results This retrospective study enrolled 163 ccRCC patients. The semiautomatic segmentation model using CMP CT images obtained DSCs of 0.98 on training cohort and 0.96 on test cohort for kidney segmentation, and DSCs of 0.94 and 0.86 for tumor segmentation on training and test set, respectively. For preoperative prediction of renal capsule invasion, the AdaBoost had best performance in batch1, with accuracy, precision, recall and F1-score equal to 0.8571, 0.8333, 0.9091 and 0.8696, respectively; and the same classifier was also the most suitable for this classification in batch 2. The AUCs of AdaBoost for batch 1 and batch 2 were 0.83 (95% CI: 0.68-0.98) and 0.74 (95% CI: 0.51-0.97), respectively. Nine common significant features for classification were found from two independent batch datasets, including morphological and texture features. Conclusions The CT-based radiomics classifiers performed well for preoperative prediction of fibrous capsule invasion in ccRCC. Advances in knowledge Noninvasive prediction of renal fibrous capsule invasion in RCC is rather difficult by abdominal CT images before surgery. A machine learning classifier integrated with radiomics features shows a promising potential to assist surgical treatment options for RCC patients.