Fig. 5

The developed ResNet-18 model utilizing arterial phase (AP) images and clinical factors (CF) produced the highest area under the receiver operation characteristic curve (AUC) on the validation and external data sets. The receiver operation characteristic curves of (A) the 2 deep learning and (B) the 2 machine learning models on validation and external data sets. (C) The AUC scores of the four models on training, validation and external data sets. During ResNet-18 model development, 5 repetitions of model training and validation were performed. The best of the 5 repetitions that produced the highest AUC on the validation set was then used to make predictions on the external data set. No repetition was performed in machine learning model development because their predictions were stable and consistent with repetitions. The receiver operation characteristic curves in (A) were plotted from the models that had the highest AUC among the 5 repetitions. SVM = support vector machine