Abstract
PURPOSE: To develop a deep learning radiomics of multiparametric magnetic resonance imaging (DLRMM)-based model that incorporates preoperative and postoperative signatures for prediction of local tumor progression (LTP) after thermal ablation (TA) in hepatocellular carcinoma (HCC). METHODS: From May 2017 to October 2021, 417 eligible patients with HCC were retrospectively enrolled from three hospitals (one primary cohort [PC, n = 189] and two external test cohorts [ETCs][n = 135, 93]). DLRMM features were extracted from T1WI + C, T2WI, and DWI using ResNet18 model. An integrative model incorporating the DLRMM signature with clinicopathologic variables were further built to LTP risk stratification. The performance of these models were compared by areas under receiver operating characteristic curve (AUC) using DeLong test. RESULTS: A total of 1668 subsequences and 31,536 multiparametric MRI slice including T1WI, T2WI, and DWI were collected simultaneously. The DLRMM signatures were extracted from tumor and ablation zone, respectively. Ablative margin, multiple tumors, and tumor abutting major vessels were regarded as risk factors for LTP in clinical model. The AUC of DLRMM model were 0.864 in PC, 0.843 in ETC1, and 0.858 in ETC2, which was higher significantly than those in clinical model (p < 0.001). After integrating clinical variable, DLRMM model obtained significant improvement with AUC of 0.870-0.869 in three cohorts (all, p < 0.001), which can provide the risk stratification for overall survival of HCC patients. CONCLUSIONS: The DLRMM model is essential to identify LTP risk of HCC patients who underwent TA and may potentially benefit personalized decision-making.