Figure 1.
Flowchart demonstrating the multi-atlas learner fusion (MLF) framework. A large collection of training images are processed offline using a typical multi-atlas segmentation pipeline. The dimensionality of the training images is then reduced, and learners are constructed to map a weak initial estimate to the multi-atlas segmentation. Finally, for a new testing image, the image needs to be projected into the low-dimensional space and the locally appropriate learners can be fused to efficiently and accurately estimate the final segmentation.