Fig. 2

(A) Schematic representation of the Probabilistic U-Net²⁷ with adversarial training⁴¹ used in this study, firstly introduced in a previous work²⁶. (B) Segmentation network of Probabilistic U-Net used in this study, which is based on the original U-Net extended into Attention U-Net only when probability maps of WMH change are used as auxiliary input. The output channel of C is either 5 or 4 depending on whether stroke lesions are jointly segmented or not, respectively. (C) Schematic of additive attention gate (AG) used in this study, firstly introduced in⁴². Input features (x_l) are from the U-Net’s skip connection while gating signals (g_l) are from the gating signal encoder (GSE). Attention coefficients (α) are learned in the training process and used to scale input features x_l to highlight important areas.