Fig. 2.

A sketch of the architecture of AVRA, with the example of a GCA-F prediction. Briefly, each slice is processed through the residual attention network and the extracted features are passed to a 2-layer Long-Short Term Memory (LSTM) network. Once all MRI slices have propagated through these two stages, a fully connected network (FC) makes a prediction of the visual atrophy rating. The MTA and PA models followed the same structure.