Fig. 3

The ResNet-18 model produced the best area under the curve (AUC) compared to other

CNN models on the training set. Nine different CNN models (e.g. VGG, ResNet, ResNext and DenseNet) and various numbers of layers in these models were tested to search for the best CNN model that utilized CT images of arterial phase to predict MVI in HCC. The numbers of layers are indicated by the postfix numbers in the model names. For each CNN model, three repetitions of 5-fold cross validation were performed on the training set. The dot represents the AUC of each 5-fold cross validation and the bar represents the mean of the three repeats.