DIPY Classibundler - Classifying fibers (better than Recobundles)

Announcements
1

Our team developed a novel algorithm for classifying fibers in tractography. It’s much similar to current Recobundles in Dipy, except that it classifies each fiber to which bundle it belongs. It was battle tested on ~200 patients so far and shows some good qualities:

  1. It’s much faster - it takes <1 min to classify all fibers in a patient. You can then extract each bundle as a set of fibers with a given label.

  2. Classification accuracy is ~95%, which is comparable to a human accuracy.

  3. There’s fewer tuning parameter comparing to Recobundles, that you basically don’t need to touch at all - it’s already tuned for the classifier based on HCP842 atlas.

We’d like to contribute it to Dipy and would love some guidance and feedback. Thank you!

Classibundler

2

Sure, you need to make a PR at DIPY’s GitHub. Note that we usually merge new methods after they have been peer reviewed in a journal or conference. Your contribution here seems using Numba which we do not currently support. The use of kdtree was also used in previous work. Therefore, make sure you give a good explanation of what is the contribution and what is new that improves results.

DIPY’s GitHub is available here

Blockquote