Pipelines using classical/iterative registration methods

[EvanBeal]

This PR doesn’t aim to be merged to the main branch but is opened to highlight the branch that contains the different pipelines for classical/iterative registration (not deep learning-based). These pipelines are (will be) used to determine baseline registration results enabling a direct comparison with the deep learning-based pipelines developed.

Pipelines description

Classical registration has been tested using ANTs and (axial) slice-by-slice regularized registration (SliceReg) through the sct_register_multimodal command line of SCT. Different registration strategies were designed to enable a large comparison with the developed approach:

• IMD or Syn - ANTs image based deformable registration. This approach is similar to the deep-learning based registration models as the registration is image based and includes only deformable registration. Mutual information is used as a similarity metric and with a non-linear symmetric normalization (SyN) algorithm to compute the transformation. The images are down-sampled by a factor of 2 to allow bigger deformations and faster computations and a smooth factor of 0.5 mm is used.

• IMRSR or RigSlicereg - ANTs image based rigid registration + image based (axial) slice-by-slice regularized registration. This approach is better adapted than classical affine or rigid methods for the registration of the SC due to the articulated nature of the spine. The mutual information similarity metric is used for the rigid registration and SliceReg algorithm. 

• IMRSRD or RigSliceregSyn – ANTs image based rigid registration + image based SliceReg + ANTs image based deformable registration. This approach is somewhat similar to what is sought to be achieved with the cascade approach of using two deep learning-based deformable registration models in succession, first bringing closer the moving and fixed images and then refining the registration. Similar parameters than in IMD and IMRSR are used for the deformable, rigid and slice-by-slice regularized registration.

• IMRSRD-priorSC or RigSliceregSynMask - ANTs image based rigid registration + image based SliceReg + ANTs image based deformable registration focusing on the SC using a mask to optimize the similarity metric. This registration pipeline should represent the most accurate image based registration results of the SC that can be obtained using iterative methods. The mask is obtained by dilating the SC segmentation using a mathematical morphology operation with a ball-shaped structuring element of radius 8 voxels. Similar parameters than in IMD and IMRSR are used for the deformable, rigid and slice-by-slice regularized registration.

• SEGreg or SCSegReg - SC segmentation based registration (upper baseline). This approach should represent the best registration results that could be obtained as the overlapping of the SC segmentations will be maximized and the registration results are assessed by computing the SC segmentations overlapping of the pair of images. It consists of a first affine (translation + rotation + scaling) registration step followed by a slice-by-slice regularized registration and a deformable registration using mean squares similarity metric.