I am using Freesurfer version 5.3. I tried to refine the pial surfaces by using  -T2pial option:

recon-all -subjid $SUBJ -T2/SUBJECT_DIR/$SUBJ/mri/orig/T2raw.mgz -T2pial -autorecon3 

This step was helpful - e.g. it removed nicely residuals of venous sinuses, but in some subjects it also removed part of the parietal or frontal cortex, while the white matter is appropriately segmented. 

The T2-space images had the same resolution as T1-MPRAGE (1x1x1mm) and were acquired in the same session (3T Siemens Trio, standard 32 channel coil). 

The problem does not seem to be caused by T2 misregistration :

Question : What could cause this error and how could I extend pial surfaces and get correct final segmentation ? 

Thanks a lot !

PB