Registration Network

Registration traditionally involves the alignment of the brain images to a statistical atlas. Atlases are built from several brain images collected from subjects of different brain images collected from subjects of different ages, genders and races. This stratified sampling approach is used to ensure that the atlas balanced in it's representation of each group. These collected brain images, typically magnetic resonances images (MRI) are combined to form a matrix of distributions, corresponding to voxel values.

More recently researchers have experimented with a variety of approaches to improve the speed of registration while maintaining accuracy of segmentation and identification of key landmarks in the brain images. The tandem approach use by Li and Fan (2018) to train the algorithm was interesting, since we wanted to build a network that would learn spatial relationships of the voxels over the whole brain. This network suited our purposes because it mimicked other networks like U-net known for learning spatial relationships well, but was not as large and did not have as many layers. Our hope was that it would allow us to train larger batches of instances on each epoch.