Fig. 1

Flowchart of DL model architecture. The DL model has three inputs, which are regions of interest (ROIs) cropped from three raw MRI sequences (T1, T1D, and T1V). In order to make all the ROIs have the same size, we resized them into 320 × 320 pixels, then the processed ROI were input into the conventional neural network (CNN). The separate CNNs were utilized for feature extraction from each of the three ROIs. The features extracted from these three branches were fused and subsequently fed into fully connected (FC) layers combined with a SoftMax classifier to obtain the predicted results. The DL model has three outputs, including predicted results of MVI absent, MVI-grade 1, and MVI-grade 2. MVI-grade 1 and 2 categories were together deemed as MVI present. MVI, microvascular invasion; DL, deep learning; T1, T1-weighted imaging; DWI, diffusion-weighted imaging; T2, T2-weighted imaging; T1A, T1-weighted imaging at arterial phase; T1V, T1-weighted imaging at portal venous phase; T1D, T1-weighted imaging at delayed phase