fig1

Figure 1. Architecture of the proposed multi-task neural network. The model utilizes a DenseNet backbone with shared parameters to extract features from triphasic CT images (arterial, portal venous, and delayed). The network branches into three task-specific outputs: a classification head for MVI prediction and two survival heads for OS and RFS estimation. CT: Computed tomography; 3D Conv: three-dimensional convolution; BatchNorm: batch normalization; ELU: exponential linear unit; MaxPool: max pooling; MVI: microvascular invasion; OS: overall survival; RFS: recurrence-free survival.