@onurulgen As mentioned earlier - I am finally able to allocate some time to NiftyReg.

Looking at differences between this branch and master, it appears some discrepancies happen when using lncc as a measure of similarity between the two branches. I will add relevant tests to pinpoint the potential error(s).

@onurulgen When you have a chance, could you go through the list of what still needs to be done before we finalise this branch and merge it back?
Please add it as a list of tasks and we can get cracking.

@onurulgen As mentioned earlier - I am finally able to allocate some time to NiftyReg.

Looking at differences between this branch and master, it appears some discrepancies happen when using lncc as a measure of similarity between the two branches. I will add relevant tests to pinpoint the potential error(s).

Fixed in e6855af

The symmetric scheme isn't supported for these classes:

• reg_nmi_gpu
• reg_ssd_gpu
• reg_optimiser_gpu
• reg_conjugateGradient_gpu

CUDA implementations needed:

• CudaCompute::GetDefFieldFromVelocityGrid()
• CudaCompute::ApproxLinearEnergy()
• CudaCompute::ApproxLinearEnergyGradient()
• CudaCompute::SmoothGradient()
• CudaCompute::ConvolveVoxelBasedMeasureGradient()
• CudaCompute::ExponentiateGradient()
• CudaCompute::SymmetriseVelocityFields()
• CudaCompute::UpdateVelocityField()
• reg_lncc_gpu
• reg_optimiser_gpu::Perturbation()
• CudaCompute::GetLandmarkDistance()
• CudaCompute::LandmarkDistanceGradient()
• CudaCompute::GetApproximatedGradient()
• CudaCompute::BchUpdate()
• reg_mindssc_gpu
• reg_mind_gpu
• reg_kld_gpu
• reg_dti_gpu

Incomplete implementations:

• Fix reg_spline_getDeformationField_gpu to accept composition
• Fix reg_getImageGradient_gpu to accept interpolation and timepoints

@onurulgen The block matching test added above fails in Release., but runs fine in Debug. There might be a trouble with CPU or CUDA block matching. I'll add some unit tests for this as well (but unlikely this week).

Morning @onurulgen,

I was writing a test for the NMI gradient computation and found out that measures have to be initialised with a F3DContent. I think that NMI (and the other measures) should use the base Content instead. As we discussed last week, the measures computations do not need to know about the transformation model as all is based on a voxel-based approach. Would you be able to change this throughout? Happy to have a chat later today if this is not clear or you don't agree. Thanks

We can't put into Content since it fills AladinContent with unnecessary stuff. I'll reorganise Content classes as