lookup_algorithm.txt

- Store pointers to voxel neighbourhoods std::vector< std::vector< std::vector< std::vector<*Voxel> > > >

- multiply x coord by inverse of voxel size.
- add negative offset of each vector (i.e. if x goes from -2*voxel_size to 5*voxel_size, then the x_offset would be 2.0)
- cast the value to size_t
- check to see if new coord is out of bounds (i.e. negative or greater than vector size), 
  - if so add enough new elements to the vectors to reach it.
- Repeat for y 
- For z repeat the look up procedure
  - If elements are missing, new neighbourhoods and new voxels need to be initialised. The generated image should be 
    made big enough to avoid this happening often.
  - Copy the segment to each of the voxels in the neighbourhood as rows in its segment matrix.

NB: This procedure could be performed in multiple threads if a lock was placed on the lookup/increment of the number of 
segments for each voxel (and the resulting reallocation of the voxels vector and segment matrix if required however this
should be rare as long as they are initialised large enough).
  
These voxels can then be processed independently in separate threads and their gradients, hessians collected back into 
the master matrices when they return.
  
Side note, all the transformations can be pregenerated for the Fourier tract conversion in one conversion matrix. An index
to the constructed rows of this matrix (and the tangent matrix) can then be passed to the voxels alongside the segment 
and the intensity value.

Base intensity and SNR should be used to prescale the reference image so that direct comparisons can be made with the 
generated intensities. The acs values of the tractlets can also be scaled I suppose.