Figure 2.

Deep Symmetry-Sensitive Convolutional Neural Network (DeepSymNet) architecture overview. Longitudinal images go through a Siamese network composed of 3-D Inception modules that share weights. These outputs are passed through a L-1 merge layer that computes differences between these two sessions. Then, this is passed through a final 3-D Inception layer to learn from these differences. Lastly, these outputs are flattened and put through a dense layer for the final AD-progression prediction.