Second Coding Week: First MI Implementation#

Hi there!

Once again, welcome to a new blog post on my GSoC 2026 journey. This was another busy week, but a productive one. My main goals for this past week were focused on:

  • Making a first attempt at implementing a SyN-compatible MI metric for DIPY.

  • Improving our benchmarking strategy.

First MI Implementation#

After an in-depth study of the Mutual Information implementation in ITK, this time going well beyond the interface level, I was able to understand how ITK manages to offer one shared metric structure for both parameter-based global registration and diffeomorphic, displacement-field-based registration.

The key idea lies behind the LocalSupport configuration. ITK considers two paths when computing the derivatives of the metric with respect to the parameters being optimized:

  • For global transforms, such as affine transforms, it can directly compute the derivative of the joint histogram with respect to a small parameter vector.

  • For dense displacement fields, it does something smarter and avoids storing a full derivative tensor for all parameters. This is important because not every displacement vector in the field affects every bin in the joint histogram. Instead, ITK stores only the local contributions of the displacement parameters that actually affect the relevant histogram bins.

Inspired by this strategy, I built a first implementation, simple and not yet as efficient as it could be, in a new double-loop-MI branch on my DIPY fork.

This first version follows a two-pass strategy. It loops over the image once to build the joint histogram and compute the MI weights associated with the histogram bins. It then loops over the image a second time to compute the dense update field directly, using the moving-image gradient, the Parzen-window derivative, and the previously computed MI bin weights.

For now, I tested it against ITK’s implementation using toy examples built from smooth shifted Gaussian images and running one step under identical initial conditions. Up to now, both frameworks return nearly the same update direction, with the first local tests showing a cosine similarity close to 0.999 between the DIPY and ITK derivative fields.

This does not yet mean that the implementation is ready to be merged, but it does represent a strong start.

Benchmarking#

I also studied Omar Ocegueda’s past experiments repository, which was previously used for registration benchmarking. Inspired by the evaluation strategy used there, I decided to add label-overlap-based metrics to my benchmarking pipeline.

In my case, I started using SynthSeg-generated labels over OASIS 2 data. The idea is to save the generated anatomical label maps during preprocessing, apply the same prealignment and SyN warps to the moving labels using nearest-neighbor interpolation, and then compare them against the fixed labels.

For this, I added Dice and Jaccard overlap scores to the benchmark results, including both whole-brain overlap and mean per-label overlap.

New PR#

Finally, I opened a PR for removing holes and islands from SynthSeg masks.

What’s Next?#

For the next week, I have the following goals:

  • Test the MI implementation with real examples and compare it against ANTs.

  • Inspired by Omar’s repository, build a cluster-based workflow for faster, more efficient, and more reproducible remote benchmarking.

  • Complete the recently opened PR after discussing the details with my mentors Serge and Jong Sung.

Find Me Online#

Thank you for reading once again!

First Coding Week: Benchmarking and MI Study   Third Coding Week: MI testing and Benchmarking
On this page