Aims
The purpose of this study was to establish a preoperative clinical-radiomics prediction model of post-hepatectomy liver failure (PHLF) in patients with hepatocellular carcinoma(HCC), and to predict clinical outcome of HCC patients who suffered PHLF.
Methods
The study included 528 HCC patients who underwent hepatectomy first time from January 2015 to December 2019 in The First Affiliated Hospital of Sun Yat-sen University and Sun Yat-sen University Cancer Center.ICGR15 detection and Gd-EOB-DTPA-Enhanced MRI was performed within 30 days before surgery.Patients in this study didn`t have other tumors or serious organic diseases and were followed up after liver resection for 90 days.
We obtained 57766 Gd-EOB-DTPA-Enhanced MR images from 528 patients.Through multiple-model fusion algorithm,such as H-DenseUnet,etc, we extracted 1044 features per patient from his Gd-EOB-DTPA-Enhanced MR images,screening out feature sets of high contribution by RFE-SVM algorithm and transforming them to FF scores.
Clinical indicators, radiologic features and FF scores were included in the significance test of difference or univariate analysis as well as multivariate analysis. Then we managed to establish a preoperative clinical-radiomics prediction model of PHLF in HCC patients,namely LF scoring model through LDA(Linear discriminant Analysis) algorithms.
Results
The AUC of LF scoring model reached 0.953 (95%CI 0.950-0.984) and 0.945 (95%CI 0.941-0.980) in the training set and the testing set, respectively. The accuracy of the model were 0.909 in the testing set, as well as the sensitivity and the specificity were 0.882 and 0.938 in it. Moreover,the positive predictive value and the negative predictive value achieved 0.938 and 0.882 respectively in the testing set.
The model showed Maximum diameter of Nodes, PLT, Excision site, iMELD score and FF score were independent risk factors of PHLF in HCC patients (OR=1.258, 0.986, 4.670, 1.237, 320.382; P < 0.05). Moreover,the AUC of the test set of LF scoring model was significantly higher than clinical-imaging model whose AUC of the test set was 0.842 (P<0.05) and single radiomics model whose AUC of the test set was 0.710 (P<0.05).
Meanwhile, the accuracy of the model in predicting overall survival (OS) and progression-free survival (PFS) of HCC patients with PHLF were 0.771 and 0.762, respectively.
Conclusions
The study drew a conclusion that Maximum diameter of Nodes, PLT, Excision site, iMELD score and FF score are independent risk factors of PHLF in HCC patients. In addition,it also indicated that LF scoring model has a good predictive value for the occurrence of PHLF in HCC patients and plays a vital role in predicting clinical outcome of HCC patients who suffered PHLF.