create a local deformation transform to simulate compressed spinal cords and apply as augmentation #137
Description
As of release of the new nnunet model, we discovered that the it does not work as good as the previous monai-based model on DWI images (especially on DCM DWI data). It is quite strange because nnunet model works quite well on T1w/T2w compressions but not as good for the DWI contrast.
We could create a data augmentation approach that simulates local deformation on the cord mask, either using standard transforms (like elastic deformation) or create a warping field obtained from a non-linear registration and apply those to the trianing images. We would have to find healthy DWI data and register them to compressed DWI scans.
Another key point is to integrate it into the transforms used by the nnUNet framework. Though it is not straightforward, it is doable. There is a paper that integrates a new transform into the nnunet framework and creates a trainer class to be run directly from the command line.