To develop a CT-based radiomics model and a combined model for preoperatively discriminating infiltrative renal cell carcinoma (RCC) and pyelocaliceal upper urinary tract urothelial carcinoma (UTUC), which invades the renal parenchyma.
Eighty patients (37 pathologically proven infiltrative RCCs and 43 pathologically proven pyelocaliceal UTUCs) were retrospectively enrolled and randomly divided into a training set (n = 56) and a testing set (n = 24) at a ratio of 7:3. Traditional CT imaging characteristics in the portal venous phase were collected by two radiologists (SPH and ZXL, who have 4 and 30 years of experience in abdominal radiology, respectively). Patient demographics and traditional CT imaging characteristics were used to construct the clinical model. The radiomics score was calculated based on the radiomics features extracted from the portal venous CT images and the random forest (RF) algorithm to construct the radiomics model. The combined model was constructed using the radiomics score and significant clinical factors according to the multivariate logistic regression. The diagnostic efficacy of the models was evaluated using receiver operating characteristic (ROC) curve analysis and the area under the curve (AUC).
The RF score based on the eight validated features extracted from the portal venous CT images was used to build the radiomics model. Painless hematuria as an independent risk factor was used to build the clinical model. The combined model was constructed using the RF score and the selected clinical factor. Both the radiomics model and combined model showed higher efficacy in differentiating infiltrative RCC and pyelocaliceal UTUC in the training and testing cohorts with AUC values of 0.95 and 0.90, respectively, for the radiomics model and 0.99 and 0.90, respectively, for the combined model. The decision curves of the combined model as well as the radiomics model indicated an overall net benefit over the clinical model. Both the radiomics model and the combined model achieved a notable reduction in false-positive and false-negativerates, resulting in significantly higher accuracy compared to the visual assessments in both the training and testing cohorts.