Fig. 1.

Overview of the latent space derived from the proposed method. All trajectories form a vector field (blue box) modeling brain aging (red curve). Each trajectory is assigned to a cluster centroid (orange star). During training, mini-batches are sampled from neighboring clusters (orange blob) until it reaches the desired size for the mini-batch. Mapping into a neighborhood is regularized according to brain age and progression consistency. For example, $z^1$ in the cyan trajectory in the upper-right box is first encouraged to be similar (inward red arrow) to its corresponding cluster centroid and dissimilar (outward gray arrow) to other centroids. Furthermore, as shown in the lower right green box, we encourage the direction of this vector to be consistent with $\Delta h$ (purple), a vector aggregated from the neighborhood of $z^1$. (blue circle).