Fig. 3.

Architecture for automated WMH segmentation. The model first captures disease priors using a convolutional auto-encoder (top) that mimics experts' knowledge of the spatial distribution of WMH. The auto-encoder contains four sets of convolution layers, max-pooling layers, a dense layer (black arrow) to capture spatial covariance and create a fixed-length encoding, and four sets of convolution and up-sampling layers. We use ReLu activation function on all convolution layers. The inference network (bottom) uses this (fixed) prior by taking an input scan and projecting down to an encoding using a similar architecture as above with independent parameters, before using the prior decoder weights to yield a segmentation from this encoding.