Accuracy of FastSurfer cortical parcellation after lesion filling in patients with moderate-to-severe traumatic brain injury

Lesion characteristics (e.g., types, extents, and locations) vary widely in patients with moderate to severe Traumatic Brain Injury (ms-TBI). Automated tools for parcellating anatomical MRI scans, such as FastSurfer, assume the input brain has normal anatomy. Thus, heterogeneous lesions often cause parcellation errors. In this lesion simulation study, we test whether a recently developed lesion filling tool (Virtual Brain Grafting, VBG) reduces FastSurfer parcellation errors. T1-weighted (T1w) magnetic resonance images (MRIs) were obtained from 140 healthy controls and 14 ms-TBI patients from the ENIGMA TBI working group. Binary masks indexing focal lesions (ranging from small [38mm3] unilateral, to large [164,291mm3] bilateral lesions) in the ms-TBI patients were each projected onto 10 healthy control T1w MRIs, creating a group of 140 synthetically lesioned images (lesions from 14 patients x 10 HCs T1w images). VBG was used to fill lesioned images with simulated healthy brain tissue, creating 140 VBG-filled images. A third set of lesion-free images were created in the same anatomical space, to be used as the “ground truth” against which to compare the lesioned and VBG-filled images. FastSurfer was used to parcellate the cortex. The accuracy of VBG filling was assessed qualitatively through visual inspection. Mean Dice Similarity Coefficients (DSC) and Percent Volume Differences (PVD) between lesioned and lesion-free images were compared to DSC and PVD scores between VBG-filled and lesion-free images, using paired sample t-tests.VBG-filled images predominantly showed anatomically plausible filling. However, several distinct boundaries and visible textural differences were observed, especially for lesions on the cortical ribbon. Near large focal lesions, VBG-filled images appeared more similar than the lesioned images to the ground truth (lesion-free images). Contrary to expectations, parcellations from lesioned images (DSC M = 0.93, SD = 0.03; PVD M = -0.40, SD = 1.7) had significantly higher DSCs [t(111) = 19.5, p < 0.001], and lower PVDs [t(111) = 11.3, p < 0.001] with the parcellations from the lesion-free images, compared to the parcellations from VBG-filled images (DSC M = 0.81, SD = 0.07; PVD M = -9.03, SD = 7.72). In other words, the parcellations from the lesioned images had closer spatial alignment and volume measurements were in greater agreement with the lesion-free (ground truth) images, when compared to VBG-filled images. This unexpected finding might be due to a high frequency of smaller focal lesions in our TBI patients and a possible improvement in FastSurfer parcellations when encountering such lesions. VBG was not developed for lesions caused by TBI, and thus a lesion filling tool is needed that specifically addresses the heterogeneity of lesions observed in ms-TBI.