Abstract Background The performance of contrast‐enhanced computed tomography (CECT) in staging clear cell renal cell carcinomas (ccRCCs) and assessing tumor aggressiveness remains limited by heterogeneous and poor sensitivity. Purpose This study aims to design and validate a human‐AI interactive network, Kidney Tumor Staging Network (KtSNet), which leverages weakly supervised learning for real‐time, efficient detection of high‐grade aggressive ccRCC (HGRCC) using CECT. Materials and methods A total of 1,092 patients with ccRCC were enrolled across five cohort datasets (training/internal testing/external testing, n = 611/153/328). To achieve precise pre‐surgical detection of HGRCC on CT imaging, we pretrained a self‐supervised foundation model (SSFM) using a large cross‐modal dataset ( n = 40 000) for image restoration‐based transfer learning. To develop human‐AI interactive capabilities, we trained KtSNet by integrating SSFM with weakly supervised learning, enabling real‐time determination of HGRCC on CT imaging through human‐AI interaction. Results In the internal test cohort comprising 153 patients, KtSNet demonstrated significantly higher Area Under the Curve (AUC) values for both ROC and PR curves (ROC‐AUC = 0.76; PR‐AUC = 0.29, F 1 max : 0.441) compared to B2Net (ROC‐AUC = 0.68, p = 0.040; PR‐AUC = 0.22, F 1 max : 0.366), RML‐XGB (ROC‐AUC = 0.53, p < 0.001; PR‐AUC = 0.14, F 1 max : 0.264), and Likert scoring (ROC‐AUC values of 0.57, 0.58, and 0.70 for the three readers) in staging HGRCC. In the external validation cohort, KtSNet maintained superior AUCs on both ROC and PR curves (ROC‐AUC = 0.85; PR‐AUC = 0.42, F 1 max : 0.529) compared to B2Net (ROC‐AUC = 0.74, p = 0.002; PR‐AUC = 0.21, F 1 max : 0.328) and exhibited significantly higher AUCs than RML‐XGB (ROC‐AUC = 0.63, p = 0.002; PR‐AUC = 0.23, F 1 max : 0.359). Conclusions The weakly human‐supervised KtSNet may serve as a promising opportunity for real‐time determination of HGRCC using CT imaging.
Zhi et al. (Sun,) studied this question.