Hi Georg
Since you are comparing manual and automated, I assume you do not really need to incorporate any covariates, so using shapeAnalysisMANCOVA may not be necessary (there is no potential for overfitting in StatNonParamTest).

I could advise a bit better if I have more information. Can you tell me what exactly you want to do? Do you want to compare how shape changes longitudinally (two time points) for manual vs automatically segmented structures? 

So, you would want to get difference vectors at every surface location measuring the change from one timepoint to the next and then analyze those 2 vector maps between 2 groups (manual vs automatic segmentation)?

1) can you post the command line you use with MeshMath to compute the vector difference maps, so we can try to reproduce the error here?
2) Instead of looking at vectors, you could also look at the signed distance maps (univariate). When doing so, I suggest you project the distance/difference to the local surface normal. This is achieved via MeshMath's -magNormDir option (it operates on the vector maps computed with MeshMath).

Re statistical analysis:
If you use option (2) above, you can use other statistical software such as SurfStat, FreeSurfer's tool or FSL's "first_utils --vertexAnalysis": http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FI... The nice thing about the FSL tool is that it should work natively with our vtk surfaces (though we have not yet used it ourselves)

Just FYI, we are working on a replacement for shapeAnalysisMANCOVA that would solve all of these issues.

Best regards
Martin