Abstract: Objective To explore convolutional neural network (CNN) based on multi-parameter magnetic resonance (MR) sequences, combined with traditional radiomics signature and clinical indicators, in predicting microvascular invasion (MVI) of hepatocellular carcinoma (HCC) preoperatively.  
Methods Two hundred and seventy-five patients with pathologically confirmed HCC were enrolled in this study. The data set was randomly divided into a training set (n=192) and a test set (n=83). A MVI predictive classifier for HCC was developed by using CNN technique, which fused 2D multi-parameter MR tumor images, 3D traditional radiomics signature and clinical indicators. Using the receiver operating characteristic (ROC) curve, the performance of combined model (ModelCom), CNN model (ModelD), radiomics model (ModelR) and clinical model (ModelC) were compared. 
Results The area under the ROC curve (AUC) of ModelD was 0.914 in the training set and 0.842 in the test set, which was better than that of ModelC (training set: P＜0.001; test set: P=0.032) and ModelR (training set: P＜0.001; test set: P=0.044). The AUC of ModelCom in the training set and test set were 0.951 and 0.881, respectively, which was better than that of ModelD in the training set (P=0.012), but there was no significant difference in the test set (P=0.157). The calibration curve showed that ModelCom had a good goodness of fit (Hosmer-Lemeshow test, P=0.402 for training set, P=0.689 for test set). Decision curve analysis showed that the net benefit of ModelCom in identifying positive MVI and negative  MVI was higher than that of other models. 
Conclusion The ModelCom based on CNN can accurately predict the MVI status of HCC.

Key words: liver neoplasms, convolutional neural network, radiomics