Fourth Coding Week: Halfway there!#
Hello everyone!
I am now halfway through the coding period, and things are going well. This week was challenging, since I was very busy with work and also ran into some issues that I will explain later. Still, my work during the fourth coding week focused on:
Experimenting with benchmarking against ANTs using the NCC metric and monomodal data, tuning and introducing changes to improve the quantitative performance of DIPY’s framework.
Testing whether my MI implementation from last week can work with multimodal data.
Updating the SynthSeg brain-mask PR.
Studying the possibility of updating SynthSeg’s behaviour in DIPY, since it currently crops input data, which can introduce cropping artifacts in the output labels and masks.
Benchmarking#
To evaluate the registration pipeline beyond isolated examples, I ran a larger OASIS-2 monomodal benchmark using the CC metric. The first experiment used 100 registration pairs and confirmed the trend observed in the previous week: ANTs still performed slightly better than the current DIPY configuration.
Method |
NCC |
NMI |
Label Dice |
Label Jaccard |
|---|---|---|---|---|
Baseline |
0.9366 |
1.1833 |
0.9234 |
0.8598 |
ANTs |
0.9506 |
1.2008 |
0.9236 |
0.8600 |
DIPY |
0.9471 |
1.1942 |
0.9219 |
0.8572 |
Since running the full benchmark repeatedly is expensive, I used the 10-pair
benchmark from the previous week as a faster proxy to test possible changes
before launching longer experiments. I explored several implementation paths:
matching the nominal update-field smoothing scale between DIPY and ANTs,
replacing ScaleSpace with IsotropicScaleSpace, revising the scale-space
intensity normalization strategy, and testing a custom implementation for
update-field smoothing.
Across these experiments, the largest changes were observed when modifying the smoothing applied to the update fields at each optimization step. This happened both when tuning the relevant smoothing parameters and when replacing the smoothing algorithm itself. In contrast, the scale-space and normalization changes produced much smaller quantitative differences in the proxy benchmark.
However, part of this process was affected by a benchmark interpretation issue: I initially focused on individual outputs instead of the aggregated benchmark summaries by mistake. As a result, some intermediate decisions were guided by misleading evidence.
Despite this issue, after selecting a set of changes, I was still able to obtain promising results with the updated DIPY configuration. The corresponding changes are available in the ants-syn-improvements branch. For 100 pairs, the updated DIPY configuration slightly outperformed ANTs in the intensity-based metrics.
Method |
NCC |
NMI |
Label Dice |
Label Jaccard |
|---|---|---|---|---|
Baseline |
0.9366 |
1.1833 |
0.9234 |
0.8598 |
ANTs |
0.9506 |
1.2008 |
0.9236 |
0.8600 |
DIPY |
0.9508 |
1.2011 |
0.9219 |
0.8571 |
Overall, these results suggest that the explored changes can reduce the image-similarity gap between DIPY and ANTs. The next step should be to isolate the smallest change that consistently improves DIPY before proposing a minimal PR.
MI on multimodal data#
To assess the behaviour of my double-loop-MI
implementation on multimodal data, I ran a preliminary registration experiment
using the data available through read_syn_data(). This example contains a T1
image and a B0 image, making it a useful first test case for mutual information
registration.
My original goal for this week was to begin implementing the single-loop MI metric for DIPY’s SyN framework. I did not have enough time to start that implementation, but I was able to test the current double-loop implementation on this multimodal pair and compare it against ANTs MI.
Method |
NCC |
NMI |
Dice |
Jaccard |
|---|---|---|---|---|
Baseline |
-0.4415 |
1.0782 |
0.9597 |
0.9225 |
ANTs MI |
-0.5965 |
1.1198 |
0.9801 |
0.9611 |
DIPY MI |
-0.5900 |
1.1224 |
0.9796 |
0.9601 |
Although this is only a single-pair experiment, the results are encouraging. The current DIPY MI implementation achieved the best NMI value and produced overlap metrics very close to ANTs MI. This suggests that the implementation behaves reasonably on a multimodal example, even though more extensive testing will be needed before drawing stronger conclusions.
SynthSeg PR#
The PR #4052 was updated to address
the review comments. The SynthSeg.predict API now follows a consistent return
contract with three items: the prediction, the label dictionary, and an optional
brain mask. When return_masks=False, the mask is returned as None.
The return_prob and return_masks options are also compatible, so users can
request probability maps and a brain mask at the same time. The related workflows
and tests were updated to follow this fixed return structure, and redundant mask
post-processing was removed from the SynthSeg workflow.
SynthSeg Input Size#
Last week, I found that the fixed input size currently used for SynthSeg is
unnecessary and potentially harmful. The current preprocessing crops or pads
all inputs to a fixed 192 x 192 x 192 shape, but SynthSeg only requires each
spatial dimension to be compatible with the network downsampling operations,
meaning that each dimension should be a multiple of 32.
Based on this, I opened the flexible-input-synthseg branch, where each input image is padded to the smallest valid shape whose dimensions are multiples of 32. This avoids imposing an unnecessarily small field of view and reduces the risk of removing relevant anatomy before segmentation.
A more efficient solution would be to compute a deterministic foreground mask from the input image, estimate the smallest bounding box containing the relevant anatomy, and then apply padding or cropping around that region. This would make the preprocessing both memory-efficient and anatomy-aware. However, the current version already solves the main issue by preserving the complete input anatomy while still satisfying the dimensional constraints required by the model.
Next Week’s Work#
Next week, I expect to have more time to dedicate to GSoC. My main goals are:
Further analyze which of the proposed registration changes are actually beneficial, with special attention to update-field smoothing, since it showed the strongest effect on the benchmark results.
Run additional experiments on a different dataset, or on inter-session OASIS-2 pairs, to check whether the observed improvements generalize beyond the current setup.
Test other monomodal metrics to better separate metric-related differences from optimizer-related differences.
Prepare a minimal PR with the changes that are clearly shown to improve NCC-based registration accuracy.
Start implementing the single-loop version of the MI metric.
Build a merged benchmarking branch to evaluate MI as the SyN similarity metric.
As a secondary task, continue investigating flexible SynthSeg input sizes and discuss whether this should be opened as a separate PR.
Find Me Online#
GitHub: TomasGuija
LinkedIn: Tomás Guija Valiente
Thank you for reading!