Fig. 3:

A sample of fixed (T2) and moving image (T1) in our training dataset used for deriving a deep metric for registration. The moving image is misregistered by a random affine transformation. Two classes of patches are shown on the right (we crop 3D patches; the middle cross-section of each patch is shown in 2D). The registered class (z = 1) contains patches that are cropped from the same location in the space of images, and the unregistered class patches (z = 0) are randomly picked. Fixed and moving image patches (for both classes) are concatenated in the channel dimension and used for training a deep binary classifier by minimizing the cross-entropy loss.