Last week I finished with the implementation of the VoxelMorph framework and trained it on the raw sample data (150 T1-weighted images). The data was not skull stripped and affine registered as indicated in the paper. Naturally, the results were bad.
This week I performed the necessary preprocessing: removed the skulls in the volumes using the brain segmentation that was provided, and affine registered the volumes to the static image, using the DIPY tutorial: Affine Registration in 3D. The architecture, the loss functions, and the hyperparameters are the same as those described in the paper.
Link to Colab notebook - https://colab.research.google.com/drive/1-lPuD4vRCeihKYMQKDDXLYq1dxm-3tJR?usp=sharing
 |
 |
|---|
After training the model with several settings, what I observed is that the model only works for small deformations, so affine registration is very important. Also, I guess the results can be further improved by training on a larger dataset.
- Balakrishnan, G., Zhao, A., Sabuncu, M.R., Guttag, J. and Dalca, A.V., 2018. An unsupervised learning model for deformable medical image registration. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 9252–9260).
- Jaderberg, M., Simonyan, K. and Zisserman, A., 2015. Spatial transformer networks. In Advances in neural information processing systems (pp. 2017–2025).
- TensorFlow implementation of spatial transformer networks. https://github.com/tensorflow/models/tree/master/research/transformer
- Spatial Transformer Networks by Kushagra Bhatnagar. https://link.medium.com/0b2OrmqVO5