Figure 3.

Implementation of the weighted-sum feature fusion. Since it is not possible to depict the original 4D structure, the features are depicted in 3D. A and B are the feature sets of the first and second modalities, respectively. The dotted arrows show the separation of the i-th channel from both the red and the blue feature-set, where $A_i$ and $B_i$ are the i-th channel of A and B. A convolutional operator is then used to combine these two channels into one channel, depicted as $F_i$.