/TransMorph_TVF

Unsupervised Learning of Diffeomorphic Image Registration via TransMorph (PyTorch)

Primary LanguagePythonMIT LicenseMIT

Unsupervised Learning of Diffeomorphic Image Registration via TransMorph

We propose a learning-based framework for unsupervised and end-to-end learning of diffeomorphic image registration. Specifically, the proposed network learns to produce and integrate time-dependent velocity fields in an LDDMM setting.

This repository contains the source code for two models, TM-TVFLDDMM and TM-TVF, from our paper: "Unsupervised Learning of Diffeomorphic Image Registration via TransMorph"

Modeling time-dependent velocity fields using TransMorph:


Skip-connections were omitted for visualization.

Smoother transformation without imposing a diffeomorphism:

Diffeomorphic registration

Forward:
 

Inverse:
 

Inversion and composition:

State-of-the-art performance:

Click on the Model Weights to start downloading the pre-trained weights.
We also provided the Tensorboard training log for each model. To visualize loss and validation curves, run:
Tensorboard --logdir=*training log file name* in terminal. Note: This requires Tensorboard installation (pip install tensorboard).

2021 MICCAI Learn2Reg challenge Task 03:

Validation set results

Ranking Model Dice SDlogJ HdDist95 Pretrained Weights Tensorboard Log
1 TM-TVF 0.8706 ± 0.0154 0.0998 1.3903 Model Weights (1.72GB) Tensorboard Training Log (1.52GB)
2 TM-Large 0.8623 ± 0.0144 0.1276 1.4315 - -
3 TransMorph (TM) 0.8575 ± 0.0145 0.1253 1.4594 - -
4 TransMorph-TVF_LDDMM 0.833 ± 0.016 0.090 1.630 Model Weights (1.71GB) Tensorboard Training Log (1.58GB)

Test set results (results obtained from Learn2Reg challenge organizers)

Ranking Model Dice SDlogJ HdDist95
1 TM-TVF 0.8241 ± 0.1516 0.0905 ± 0.0054 1.6329 ± 0.4358
2 TM-Large 0.8196 ± 0.1497 0.1244 ± 0.0148 1.6564 ± 1.7368
3 TM 0.8162 ± 0.1541 0.1242 ± 0.0136 1.6920 ± 1.7587
4 LapIRN 0.82 0.07 1.67
5 ConvexAdam 0.81 0.07 1.63
...