Multi-atlas
segmentation has often suffered from the registration error. We propose a novel
method for multi-atlas registration for multi-atlas segmentation inspired by the template generation
and deep neural network. We first add an intra-atlas registration layer which
performs image-based registration between atlas images to duplicate the
atlases. We then add a label-wise registration layer which rectifies the
registered images by label-based registration. We present preliminary results
of our multi-layered atlas registration on brain MRI segmentation.
