Groupwise Bundle Registration

How it works

The bundle groupwise registration framework in DIPY relies on streamline linear registration (SLR) [Garyfallidis15] and an iterative process.

In each iteration, bundles are shuffled and matched in pairs. Then, each pair of bundles are simultaneously moved to a common space in between both.

After all pairs have been aligned, a group distance metric is computed as the mean pairwise distance between all bundle pairs. With each iteration, bundles get closer to each other and the group distance decreases.

When the reduction in the group distance reaches a tolerance level the process ends.

To reduce computational time, by default both registration and distance computation are performed on streamline centroids (obtained with Quickbundles) [Garyfallidis12].

Example

We start by importing and creating the necessary functions:

from dipy.align.streamlinear import groupwise_slr
from dipy.data import read_five_af_bundles
from dipy.viz.streamline import show_bundles

import logging
logging.basicConfig(level=logging.INFO)

To run groupwise registration we need to have our input bundles stored in a list.

Here we load 5 left arcuate fasciculi and store them in a list.

bundles = read_five_af_bundles()

Let’s now visualize our input bundles:

colors = [[0.91, 0.26, 0.35], [0.99, 0.50, 0.38], [0.99, 0.88, 0.57],
          [0.69, 0.85, 0.64], [0.51, 0.51, 0.63]]

show_bundles(bundles, interactive=False, colors=colors,
             save_as='before_group_registration.png')

They are in native space and, therefore, not aligned.

Now running groupwise registration is as simple as:

bundles_reg, aff, d = groupwise_slr(bundles, verbose=True)

Finally, we visualize the registered bundles to confirm that they are now in a common space:

show_bundles(bundles_reg, interactive=False, colors=colors,
             save_as='after_group_registration.png')

Extended capabilities

In addition to the registered bundles, groupwise_slr also returns a list with the individual transformation matrices as well as the pairwise distances computed in each iteration.

By changing the input arguments the user can modify the transformation (up to affine), the number of maximum number of streamlines per bundle, the level of clustering, or the tolerance of the method.