Hi & welcome Simona!
How should I interpret these connections?
As connections that both start and end within the same parcel; that’s basically it 
A random compilation of thoughts:
-
It would actually be highly unusual if we were to place a relatively crude parcellation atop a cortical ribbon that looks basically continuous, and somehow perfectly align the borders between parcels such that there are no connections whatsoever that both start and end within the same parcel. Bear in mind here that tractogram generation and connectome construction are two separate steps.
-
If you picture these histograms (as well as those in my article and the biological justification enclosed within), there are a massive number of very short WM fibres. If the lengths of these very short fibres are shorter than the spatial extent of the parcels, then it’s inevitable that there will be more intra-parcel connections than inter-parcel connections.
-
Particularly with HCP data, the default minimum streamline length during tracking will be much shorter in MRtrix3 (twice the voxel size, or 2.5mm) compared to other software packages (~ 20-30mm).
-
Are all of them “anatomically reliable” or should I somehow threshold the results?
You’ll find that a lot of global connectome metrics either mask these values out before calculation, or just ignore them outright. So it depends on exactly what kind of analysis you’re doing.
-
Even though I’d advocate that such connections do in fact exist, chances are we’re not able to get a good quantitative estimate of how many such connections there are. Not only due to the streamline minimum length criterion, which is an ill-posed threshold from the outset given that’s where so many biological fibres actually exist, but also because there’s most likely a continuum of different connection lengths, from ~ mm down to sum-mm, and even shorter connections that are entirely within the GM but nevertheless along the plane of the cortical ribbon.
¯\_(ツ)_/¯
Rob