I like the proposed solution of implementing this by making vtkITKArchetypeImageSeriesScalarReader smarter.
I have just a small comment: it would be a bit nicer to make the DICOM reader selector a string attribute (PreferredDICOMReader, PreferredDICOMIOMethod, PreferredIOClass, …), instead of a Boolean flag (UseGDCMImageIO), as in the future we will probably want to have more options (for example, different modes for using DCMTK or GDCM, auto-select between different readers, use other toolkits) without changing the interface.
It would be useful to have a way to load using DCMTK, even when GDCM returns something. For this, the DICOMScalarVolumePlugin could always offer two loadables for a series: a GDCM-based and a DCMTK-based. By default, if GDCM can interpret the image, the GDCM-based loadable would have higher confidence (so it would be selected by default).