[MINC-development] mincresample using grids with large steps and input volumes with small steps

[Matthijs van Eede]

Greetings MINC-experts,

When you resample a file with relatively small voxel sizes (60 micron or 
smaller) and the deformation grid that is being used to resample this 
file has significantly larger voxels, some resampling issues can occur. 
This is because the tolerance for determining where a point came from is 
based on the step sizes of the deformation grid, and not the input 
volume. (This is related to the ftol variable in the function 
grid_inverse_transform_point in grid_transforms.c).

There are some changes in minc-2.1.00, and the ftol is set to a lower 
value than would have been the case before, but we still wanted to base 
this tolerance value on the input file. In order to do this, I forked 
Andrew's git repository of minc, and implemented changes to get the 
information of step sizes all the way into that 
grid_inverse_transform_point function. (It's a long way :-))

My approach was as follows. I did not want to break any possible 
compatibility with other pieces of code that use functions defined in 
volume_io, and wanted to use a "default argument" for the functions that 
are involved in getting the step sizes of the input volume. In c that is 
not directly supported, so I did the following:

Original function:

VIOAPI  void  grid_inverse_transform_point(
     General_transform   *transform,
     Real                x,
     Real                y,
     Real                z,
     Real                *x_transformed,
     Real                *y_transformed,
     Real                *z_transformed )
{
...
...
}

And changed that in the following two function:

VIOAPI  void  grid_inverse_transform_point_with_input_steps(
     General_transform   *transform,
     Real                x,
     Real                y,
     Real                z,
     Real                *input_volume_steps,
     Real                *x_transformed,
     Real                *y_transformed,
     Real                *z_transformed )
{
...
...
}

and

VIOAPI  void  grid_inverse_transform_point(
     General_transform   *transform,
     Real                x,
     Real                y,
     Real                z,
     Real                *x_transformed,
     Real                *y_transformed,
     Real                *z_transformed )
{
   grid_inverse_transform_point_with_input_steps(transform, x, y, z, NULL,
                                            x_transformed, 
y_transformed, z_transformed );
}

And applied this to a couple of functions in order to reach the function 
in mincresample where there is still information about the input volume 
around. I'd be happy to hear any suggestions/comments/remarks about this 
approach, and if you think this is a proper way to deal with this issue. 
The forked git branch is here:

https://github.com/mcvaneede/minc

Thank you in advance,
Matthijs