Figure 6.

Five example unrolled iterative image reconstruction methods with deep networks embedded. Row 1 corresponds to taking a current image and updating it according to the data (a gradient descent step to increase agreement with the measured data), and also updating it according to the prior knowledge of how the image should appear (simply using a deep network to provide this increased agreement). Row 2 is very similar, but uses an analytically derived way for combining the two updates (rather than a simple addition, as in row 1). Row 3 corresponds to a deep learned version of a proximal gradient update. The example in row 4 is the sequential method but now using an analytical form of the update. Finally, row 5 shows how the ADMM method can have a network embedded, and it amounts to the same approach as in row 4, but with use of a residual image, u.