Hi Georg

Re within subject: Yes, both StatNonParamTestPDM and shapeAnalysisMANCOVA cannot model within subject designs (it would treat the data as separate groups rather than matched/paired data). We used the difference vector fields (using the signed difference mapped to the surface normal) as you are thinking about for our analysis in E. Maltbie, K. Bhatt, B. Paniagua, R. G. Smith, M. M. Graves, M. W. Mosconi, S. Peterson, S. White, J. Blocher, M. El-Sayed, H. C. Hazlett, and M. Styner, "Asymmetric bias in user guided segmentations of brain structures.," NeuroImage, vol. 59, no. 2, pp. 1315–1323, Jan. 2012.

Re MeshMath: MeshMath offers a mix of functions that have been accumulated over the years. Some of these functions are only offered for the older meta surface format. Our current tools will yield vtk surfaces that cannot be directly used with those functions.
The difference function is one of those older functions. So in order to use it, you need to convert the surface data from vtk to meta. In our distribution there are converters called Meta2VTK and VTK2Meta. Use these to convert all your vtk surface to meta and then use those meta surfaces with the "-subtract" option.

Re alignment: I don't like that you needed to push that alignment button in ShapePopulationViewer. This means the manual and the automatic data are not in the same space. You can align these via SPHARM & rigid procrustes (--regTemplate), but better would be to make the data reside in the same space. How are the two segmentations different? Does the automatic segmentation perform a registration, does it resample the data, that is not done prior to the manual segmentation?

Martin