Abstract
BACKGROUND: Due to the prognosis and treatment differences between primary renal sarcomas and renal cell carcinoma (RCC), preoperative differentiation between them is important but challenging. This study aims to explore and develop a diagnostic method based on computed tomography (CT) and clinical data for preoperatively differentiating primary renal sarcomas from RCCs. METHODS: Patients pathologically diagnosed with primary renal sarcoma from two centers were retrospectively included, and RCCs were probably 2:1 compared to renal sarcomas as the control group. Clinical data, standard contrast-enhanced CT images and histological findings were obtained. A clinical model was established with independent indicators based on logistic regression analysis. The region of interest was outlined in each of the three modal CT images [unenhanced phase (UP), corticomedullary phase (CMP) and nephrographic phase (NP)] and formed 7 modal imaging datasets for deep learning (DL) models' development. Reported performance metrics included accuracy (ACC) and area under the curve (AUC). RESULTS: Totally, 7,482 images were obtained from 85 patients. Multivariate logistic regression showed that the independent indicators of renal sarcoma were intratumoral arteries and Gerota's fascia invasion (P<0.05). The AUC of clinical model was 0.77 [95% confidence interval (CI): 0.67-0.87], sensitivity 0.74, specificity 0.67, positive predictive value (PPV) 0.51, and negative predictive value (NPV) 0.85. The DL models yielded effective discrimination. The UP model yielded AUC of 0.95±0.09 and ACC of 0.94±0.07, and UP + NP model yielded nearly AUC =0.95±0.06, and ACC =0.94±0.07. CONCLUSIONS: The multimodal artificial intelligence (AI) models show good performance for differentiating renal sarcomas from RCCs, which assist in individualized management.