Generation affine/non-linear transformed datasets

[EvanBeal]

This PR doesn’t aim to be merged to the main branch but is opened to highlight the branch that contains the files used to generate affine and non-linear transformed datasets.

Description of the process

Affine and non-linear transformed datasets are generated by transforming the T2w image of each subject of the Spine Generic dataset.

For the generation of affine transformed datasets: The magnitude of the transformations (rotation, scaling, translation) are randomly and independently selected (within a certain range) for each T2w image, the aim being to observe the registration accuracy on a broader spectrum of data (and potentially determine the limits of the registration models) by adding these linear transformations.

For the generation of a non-linear transformed dataset: The dataset aims to mimic strong variations in the position of a subject between the acquisitions of the different modalities. A synthetic warping field is produced for each T2w image, using a similar approach than for the generation of unregistered label maps in the Synthmorph method. The parameters used to generate the warping fields have been selected to produce realistic or stronger non-linear deformations. Considering the parameters present in the config file for training a registration model, the warping fields have been generated using a vel_res of [16, 32, 64] and a vel_std of [2.5, 5, 5]. The dataset is generated to observe how well the registration methods/models generalize on largely displaced data with non-linear relationships.