Authorea

Anisha Keshavan added We_initially_ran_freesurfer_as__.tex over 8 years ago

Commit id: 7447ff2657ab2996af71f2f9826d18367e355a42

deletions | additions

We initially ran freesurfer as is for the MS patients, but during the QC process we found errors that were very different from the errors on healthy controls, because of the effect of lesions. If we followed the QC procedure of the healthy controls (by discarding scans with noticeably bad GM/WM classification) then too many scans would be thrown out of the analysis. But more importantly, when calculating a scaling factor, its really important that the tissue classification is accurate, otherwise the scaling factor estimate would also be innaccurate. There may be a better segmentation pipeline that handles lesions better, and optimizing that pipeline is set aside until we acquire patient data from the consortium. To clarify this point, and to give more details on the types of errors, I've edited the text: \textit{The accuracy of our scaling factor estimates depends on the accuracy of tissue segmentation, but the lesions in MS specifically impact white matter and pial surface segmentations. Because of the effect of lesions on Freesurfer's tissue classification, all images were manually corrected for lesions on the T1-weighted images by a neurologist after editing by Freesurfer's quality assurance procedure, which included extensive topological white matter corrections, white matter mask edits, and pial edits on images that were not lesion filled. These manual edits altered the white matter surface so that white matter lesions were not misclassified as gray matter or non-brain tissue. The errors in white matter segmentations most typically occurred at the border of white matter and gray matter and around the ventricles. The errors in pial surface segmentations most typically occurred near the eyes (orbitofrontal) and the superior frontal or medial frontal lobes. Images that were still misclassified after thorough edits were removed from the analysis, because segmentations were not accurate enough to produce realistic scaling factor estimates.}