FastSurfer parcellation accuracy after lesion filling in moderate-to-severe traumatic brain injury

Focal lesions in T1-weighted (T1-w) magnetic resonance images (MRIs) of patients with moderate-to-severe traumatic brain injury (ms-TBI) can introduce errors during image processing. We tested whether errors in FastSurfer cortical parcellation could be reduced using lesion filling (virtual brain grafting (VBG)).

T1-w MRIs from 140 healthy controls and 14 ms-TBI patients were shared within the ENIGMA TBI working group. A "ground truth" set of 140 lesion-free images was created by registering 10 healthy controls (HCs) onto each of 14 ms-TBI images. Masks indexing focal lesions (small [38 mm³] unilateral to large [164,291 mm³] bilateral) were projected onto lesion-free images, creating 140 synthetically lesioned images. Lesioned images underwent VBG filling to replace lesioned regions with simulated healthy brain tissue, creating 140 VBG-filled images. To calculate parcellation accuracy, paired sample t-tests of mean Dice similarity coefficients (DSCs) and percent volume differences (PVDs) for lesioned and VBG-filled images were compared to lesion-free images.

Parcellations from lesioned images (DSC M = 0.93, SD = 0.03; PVD M = -0.40, SD = 1.7) unexpectedly had significantly higher DSCs [t(111) = 19.5, p < 0.001] and lower PVDs [t(111) = 11.3, p < 0.001] than VBG-filled images (DSC M = 0.81, SD = 0.07; PVD M = -9.03, SD = 7.72).

Parcellations from lesioned images were more accurate (than VBG-filled images) than lesion-free ground truth images. While likely due to a high frequency of smaller focal lesions in our sample, these results could suggest that FastSurfer parcellation may be robust in the presence of such lesions.