From datasets shared within the ENIGMA adult and pediatric ms-TBI working groups, we selected T1w MRI scans from 140 HCs (10–60 years, M = 28.68 years, SD = 13.05 years, 70 males) and 14 TBI patients in the chronic stage of injury (> 6 months after injury) (10–60 years, M = 25.6 years, SD = 16.14 years, 7 males). TBI severity was defined using either (1) the Mayo classification system (33) or a Glasgow Coma Scale score (34) at the time of hospital admission; (2) loss of consciousness for 30 min or greater; (3) post-traumatic amnesia >24 h (35); and (4) positive findings of gross injury on MRIs as per evaluation by a neuroradiologist (PB).

Supplementary Table 1 provides an overview of the original data collection sites. Informed consent was provided by participants in accordance with local ethics guidelines, and the privacy rights of participants have been observed, with data analysis occurring under the approval of Deakin University Ethics number 2023–267. TBI participants were excluded if they had any of the following: (1) previous TBIs; (2) previous diagnoses of neurological disorders; (3) current or previous substance abuse issues; and (4) contraindications for MRI (e.g., ferromagnetic implants). Ten HC subjects were age- and sex-matched to each TBI patient (age within 2 years; M = 0.4 years, SD = 1.13 yrs). Patients with TBI were selected to represent the heterogeneity of TBI lesions (Figure 1), including lesion size, location, and laterality (36). Specifically, 11 patients with TBI had focal bilateral lesions, two patients had focal unilateral lesions (one left hemisphere and one right hemisphere), and one patient had a single small lesion in the right corpus callosum. The number of lesion clusters in each TBI subject ranged from 1 to 14 (M = 7.1, SD = 3.42), while total lesion volume ranged from 38mm³ to 164,291 mm³ (M = 30,274.41, SD = 38,582.98). The 14 lesion profiles are shown visually in Figure 1. Moreover, lesion descriptions observed on T1w MRIs are provided by a neuroradiologist (PB) (Table 1).

T1w MRI images were collated from 15 different cohorts shared by investigators within the ENIGMA adult and pediatric ms-TBI working groups. An overview of each cohort’s scanning parameters can be found in Supplementary Table 1, and the processing pipeline can be found in Figure 2. Raw T1w MRI images underwent a visual quality check in SPM12¹ to ensure images were free from pronounced artifacts, ringing, or blurring. Four HC images were excluded at this stage, and replacement images were chosen. After orientation was uniformly set across all images, the origin was manually placed at the anterior commissure. All images were then rescaled, resulting in consistent dimensions of 256 × 256 × 256 (field of view) and a voxel size of 1 mm³ across the images. HC images underwent an initial whole-brain parcellation (Desikan–Killiany–Tourville (DKT) atlas) using FastSurfer’s recon-surf pipeline (v1.0.0, e4ed6f7) and were subsequently visually inspected by ED. No major errors were identified, enabling all HC images to be included.

All lesion masks were drawn in T1w native space using FSLeyes version 0.27.3 in FSL version 6.0.1.² Lesions were delineated by ED, who was trained in lesion identification by neuroradiologist PB, with masks subsequently reviewed by PB. Delineation followed an in-house systematic search method and lesion identification protocol (37), based on previous protocols using T1w MRI scans only (38, 39). Abnormalities resulting in tissue loss, such as regions of encephalomalacia and damage from surgical drainage tracts, were included, while enlarged ventricles were not. The mask included non-lesion MRI abnormalities such as hyperintensities occurring in proximity to the skull (e.g., from surgical craniotomy clips), because they are known to disrupt automated algorithms during registration and segmentation (40). To reflect the inclusion of these non-lesion abnormalities, this study will refer to repair masks instead of lesion masks throughout the manuscript.

The VBG lesion filling workflow described in detail in Radwan et al. (25) provides two methods, one for unilateral (uVBG) and one for bilateral (bVBG) lesions. As the majority of TBI subjects in this sample had bilateral lesions, and the bVBG workflow exhibits greater accuracy in unilateral lesions compared to uVBG, we chose to run bVBG (25). HD-BET is the name of a brain extraction tool, an expanded version of the name was not provided in the original paper introducing the tool (41), before undergoing brain extraction using HD-BET (42). The pre-processed input brain image is then warped to VBG’s single-subject normative template brain in MNI space using cost-function masking. Following this, the repair mask is subtracted from the brain mask before undergoing segmentation to produce tissue probability maps (TPMs) (43).

The second step in bVBG involves flipping the pre-processed, warped input brain image along the right–left axis and performing further iterative deformation to match the template brain (41). Next, an initial donor brain image (a TPM-based T1w image) is created by combining the inverse warped template, prior TPMs, and the subject’s brain image (with the repair mask excluded). In bVBG, this synthetic brain is used to derive the initial filled brain, which is then segmented with Atropos (43).

In the final step of lesion filling, the initially filled brain is warped back to native space and sharpened with ANTs (41) to create the final donor brain image. Using a 2-mm FWHM smoothed and dilated mask, the repair mask is subtracted from the initial filled brain image before being inserted into the recipient image. To generate a realistic lesion-free whole head T1w image, the skull and noise are added back in, and the initial transformation is reversed to re-generate the image in native space. In this study, dilation of the mask was not performed to ensure the regions outside of the mask were unaffected by any processes of VBG.

Lesion simulation was performed using the same steps as described in a previous study (25) (see Figure 2). Using one HC image, one TBI image, and the repair mask as input, lesion simulation generates three images: (1) a “lesion-free” image (created by registering the HC to the TBI native image space), which acts as a ground truth; (2) a “lesioned” image (created by inserting the TBI lesion onto the lesion-free image); and (3) a “VBG-filled” image (created by performing VBG lesion filling on the lesioned image). Each set of three images—lesion-free, lesioned, and VBG-filled—created from the same input images is identical in all regions outside of the lesion, enabling FastSurfer parcellations from the lesion-free images to be used as the ground truth for qualitative and quantitative analysis.

TBI images and their corresponding binary repair masks (all in native space) were repaired by VBG in TBI native space. The lesion-free images were generated using ANTs (41) non-linear warping of each set of 10 HC T1w brain images onto their matched repaired TBI image. These lesion-free images can resemble increases in general atrophy or enlarged ventricles commonly associated with TBI (44) but are not influenced by the presence of a lesion.

The lesioned images were created by matching the intensity histogram of the original TBI T1w brain image to the lesion-free images. The repair masks were smoothed (2-mm FWHM 3D Gaussian kernel) before being used to extract the histogram-matched lesioned area and insert it in the corresponding location on the lesion-free images. This process generates synthetically lesioned images that are identical to the lesion-free images in all regions outside of the repair mask.

To generate the VBG-filled images, the lesioned images were processed using bVBG (25). After lesion simulation, the VBG-filled images were observed to have visible global textural differences and were then identified to have intensity histograms noticeably different from the lesioned and lesion-free images. All VBG-filled images therefore underwent intensity matching using MRtrix3 (45) mrhistmatch-scale using the lesion-free ground truth images as the reference image (see Supplementary materials for details of histogram matching and intensity histogram plots, Supplementary Figure 1).

Cortical parcellation of all three groups of brain images, lesion-free (N = 140), lesioned (N = 140), and VBG-filled (N = 140), was conducted using FastSurfer’s (v1.0.0, e4ed6f7) recon-surf pipeline (20). In line with previous studies (17, 18) we selected 62 cortical regions (31 pairs across hemispheres) of interest (ROIs) from the Desikan–Killiany–Tourville atlas (46). Each recon-surf run was allocated 2.5 h of runtime. Runs that quit with an error unrelated to time were considered to have failed. Each image was provided two attempts to achieve successful parcellation, with run time and computing allocations kept consistent across both runs. Portions of the parcellation that fell within the repair mask were removed from the lesion-free, lesioned, and VBG-filled images.

To compare the spatial accuracy of FastSurfer’s parcellations, ANTs DiceandMinDistSum (8) was used to calculate the Dice similarity coefficient (DSC) for each cortical ROI in the lesioned and VBG-filled images, relative to that same ROI in the lesion-free images. DSC is a measure of spatial similarity between two images or regions, ranging from 0 (no similarity) to 1 (perfect overlap). The calculation of DSC for the VBG-filled image is expressed in formula 1 (where VBG-filled is substituted with lesioned as needed):

Volume of each cortical ROI in all images was calculated using ANTs LabelOverlapMeasures (8). The percent volume difference (PVD) was calculated by comparing the ROI volume of the lesioned and VBG-filled images to that same ROI volume in the ground truth lesion-free images. This can be expressed in formula 2 for the VBG-filled image (which can be substituted by the lesioned image as needed): PVD = ( VBG _ Filled _ ROI _ Vol − Lesion _ Free _ ROI _ Vol ) Lesion _ Free _ ROI _ Vol × 100

Images with a mean percent volume difference score closer to zero have volumes more similar to the ground truth compared to those with larger values. Negative values indicate that lesioned and VBG-filled images are smaller than the ground truth volumes, while positive values indicate volumes greater than the ground truth.

All statistical analyses were performed in R version 4.2.2 (2022-10-31 ucrt), within RStudio 2022.12.0 + 353. To compare the parcellation accuracy of lesioned and VBG-filled images with respect to the ground truth images, the DSC and PVD were calculated for all cortical regions and then averaged to generate a mean DSC and PVD value for each individual image. Paired samples t-tests were then used to investigate group differences between the mean DSC and PVD values for the lesioned and VBG-filled images when compared to the lesion-free images. An exploratory sensitivity analysis was conducted using Spearman correlations to investigate the impact of lesion volume and intensity distribution RMSE on parcellation accuracy (DSC and PVD) for image type, and the association between lesion volume and RMSE for image type. All results were considered significant at a p ≤ 0.05.

After visual inspection of the lesion simulation, the lesioned images generated from TBI03 and TBI13 were not successfully simulated (Supplementary Figure 2 and Supplementary Table 2 for examples). Therefore, all simulated sets of images generated from TBI03 and TBI13 were removed from subsequent analysis.

The remaining 360 brain images were run through the FastSurfer recon-surf pipeline. The majority of images (N = 350) achieved successful FastSurfer parcellation on the first attempt. Of the images generated from TBI05 (largest total lesion volume, 16,4291 mm³), FastSurfer failed to successfully complete parcellation in four of the VBG-filled images and six of the lesioned images. After a second attempt, all four VBG-filled images completed parcellation successfully, while parcellation for the six lesioned images failed once again. Those six lesioned images, along with their corresponding lesion-free and VBG-filled images, were excluded from group comparisons. The final analysis included 114 simulated sets with 342 individual images.

Visual inspection of VBG lesion filling revealed predominantly anatomically plausible filling. WM lesions appeared to be filled more coherently, with smooth boundaries making them almost undetectable without prior knowledge of their location, while the filling of cortical gray matter exhibited more distinct boundaries and visible textural differences (Supplementary Table 3). Inspection of the raw FastSurfer parcellation of lesioned images revealed misclassification errors where, when cortical regions were missing due to the presence of a large lesion, the missing tissue had been incorrectly assigned to nearby normal-appearing tissue (Supplementary Table 1 and Supplementary Figure 3). Figure 3 depicts examples where, in the regions neighboring a lesion, parcellations produced after VBG filling were in line with our expectations, appearing visually closer to the lesion-free ground truth images compared to the lesioned images.

Paired-samples t-tests were conducted to compare the mean Dice similarity coefficient (DSC) and mean percent volume difference (PVD) scores between lesioned (DSC M = 0.93, SD = 0.03; PVD M = −0.40, SD = 1.7) and VBG-filled images (DSC M = 0.81, SD = 0.07; PVD M = −9.03, SD = 7.72). Checks for the assumption of normality identified two images as statistical outliers, as their values were beyond 1.5 times the interquartile range from the quartiles. These two images and their corresponding images in that simulated set were excluded.

A paired-samples t-test comparing the spatial similarity of lesioned and VBG-filled images relative to the ground truth revealed a statistically significant difference in DSC between image types [t(111) = 19.5, p < 0.001, d = 1.49 (SD = 0.08)]. Specifically, the lesioned images produced a higher dice score (95% CI [0.11, 0.14]) compared to the VBG-filled parcellation images. In other words, the parcellations of the lesioned images were more closely aligned with the ground truth images (i.e., lesion-free parcellations) than the parcellations of the lesion-filled images (see Figure 4A).

A paired-samples t-test comparing mean percent volume difference (PVD) of the lesioned and VBG-filled images relative to the ground truth volumes revealed a significant difference in PVD between image types [t(111) = 11.3, p < 0.001, d = 1.22 (SD = 7.07)]. Specifically, the average PVD was 5.08% (95% CI [7.12, 10.1]) lower in the lesioned compared to the VBG-filled images. In respect to the ground truth parcellations, images that had undergone VBG lesion filling exhibited a significant estimation of regional cortical volumes when compared to lesioned images that had not undergone filling. Taken together, the lesioned images showed both a higher spatial similarity with and a lower PVD from the ground truth lesion-free images compared to the VBG-filled images (see Figure 4B).

An exploratory sensitivity analysis conducting Spearman correlations in the lesioned images revealed significant negative associations between lesion volume and both DSC [ρ(112) = −0.56, p < 0.001] and PVD [ρ(112) = −0.33, p < 0.001], suggesting that the parcellation accuracy in these images decreased as lesion volume increased. In contrast, no significant correlations were observed for VBG-filled images and parcellation accuracy (DSC: ρ(112) = 0.14, p = 0.15; PVD: ρ(112) = 0.17, p = 0.078). With respect to RMSE, lesioned images showed a significant negative association with both DSC [ρ(114) = −0.62, p < 0.001] and PVD [ρ(114) = −0.36, p < 0.001]. Similarly, for VBG-filled images, RMSE was strongly negatively correlated with DSC [ρ(114) = −0.81, p < 0.001] and PVD [ρ(114) = −0.22, p = 0.020]. These findings indicate that as histogram misalignment (higher RMSE) increased, parcellation accuracy decreased. Finally, investigating the relationship between lesion volume and RMSE showed that increasing lesion volume was associated with higher RMSE between intensity distributions for lesioned images relative to lesion-free images (ρ = 0.88, p < 0.001, n = 114). By contrast, RMSE for VBG-filled images was weakly but significantly negatively correlated with lesion volume (ρ = −0.25, p = 0.008, n = 114), suggesting that histogram similarity to lesion-free images improved slightly with larger lesion sizes (see Supplementary Figures 4–8).

The results of this work must be carefully considered with respect to the limitations of this study. First, our sample composition, together with methodological aspects of the lesion simulation design, may have reduced the presence of lesion-induced error in the lesioned images. Specifically, the final sample contained a higher proportion of smaller lesions (<25 cm³) relative to larger lesions (>50 cm³), with a mean lesion volume of 29 cm³. While the range of lesion sizes included in this study is representative of ms-TBI lesions observed in other cohorts (48, 49), the relatively small number of cases with very large lesions (>86 cm³) limited our ability to robustly detect lesion-induced parcellation errors and to evaluate whether VBG might yield measurable improvements under conditions of extensive lesion burden. In addition, it is worth considering whether the lesion simulation procedure, involving multiple processing steps to extract and smoothly reinsert lesions, could have resulted in simulated abnormalities with lower signal disruption than those observed in native ms-TBI scans. Together, these factors likely reduced the overall level of lesion-induced parcellation error present in the dataset, thereby minimizing the potential improvements attributable to lesion filling (25).

Another important limitation of the present work is that the effect of lesion location (or proximity) on parcellation accuracy could not be investigated. A previous study in other clinical populations with unilateral lesions (e.g., stroke, glioma) has reported that brain regions in closer proximity to lesions are likely to exhibit lower parcellation accuracy (25). However, to the best of our knowledge, there is no validated measure of distance that can provide a metric encompassing multiple distances to lesions of varying sizes, which represents a methodological gap in the field. As observed in our qualitative analyses (see Figure 3 and Supplementary Figure 4), it is likely that VBG-filled images may have shown regional improvements in parcellations closer to the lesions, which could not be quantitatively captured in the design of this study. Developing and validating such metrics will be an important issue to address in future studies.

The findings from this study point to several critical areas for future research. Future studies utilizing a much larger, statistically well-powered sample that encompasses a balanced spectrum of lesion sizes will be essential for disentangling how lesion inpainting and parcellation tools perform across multiple complex lesion characteristics (e.g., volume, location, and laterality). To expand upon the lesion simulation paradigm used in this study, we suggest that future work should use manually drawn parcellations as a gold standard ground truth.

Using manual gold standard parcellations would enable direct comparison between the accuracy of different parcellation tools such as FastSurfer, FreeSurfer, and multi-atlas label propagation with expectation–maximization (MALP-EM) (10, 20, 50). The choice of parcellation tool and corresponding atlas has been shown to significantly affect downstream analyses (51). Quantitatively assessing the strengths and limitations of several different parcellation tools with respect to performance across a variety of lesion characteristics could ultimately produce evidence-backed recommendations to guide both clinicians and researchers as to which tools are most appropriate for use in their clinical populations of interest.

Large-scale studies also hold the key to improving our understanding of the varied levels of lesion-induced errors and which inpainting tools are most effective for reducing these errors. The field of lesion inpainting is rapidly advancing, driven by the increased adoption of advanced deep learning architectures. Recent inpainting tools, such as the iterative frameworks of diffusion denoising models (DDM), have shown promising results in dealing with large lesions with complex boundaries crossing multiple tissue types (52, 53). Additionally, some automated tools are capable of performing end-to-end lesion segmentation and inpainting (54), which would improve the time-consuming process of manually segmenting lesion masks, which was a required input for VBG (25). As additional inpainting tools are developed for use in bilateral lesions, future studies will need to assess the accuracy of inpainting as measured by the reconstruction accuracy within the lesioned region (i.e., how closely aligned the inpainted region is with the corresponding ground truth region). Reconstruction accuracy has been identified as an important factor that can have impacts on downstream analyses such as functional MRI (55). Building a large-scale ms-TBI dataset could additionally support the training of inpainting tools optimized for use in TBI, a critical step toward inclusive pipelines, moving away from the exclusion of ms-TBI images due to extensive lesions and subsequent poor-quality parcellations.

On the basis of our current findings, it is evident that researchers and clinicians need to apply caution when using VBG to fill ms-TBI lesions. Our results suggest that VBG filling is most useful for enabling successful parcellation in images with extensive lesions (~190 cm³), specifically in instances where FastSurfer might otherwise fail to produce a parcellation output. Importantly, any studies utilizing VBG for ms-TBI images should also process control images using the VBG pipeline to ensure consistent pre-processing steps and mitigate potential intensity distribution changes and parcellation boundary shifts across image types. This study provides promising results with respect to FastSurfer parcellation accuracy in images with smaller lesions (<20 cm³). Based on the results in our study, we tentatively suggest that lesion inpainting may not be required in studies with samples restricted to small lesions and employing FastSurfer parcellation. However, to ensure that local lesion-induced errors do not bias subsequent downstream analyses in such studies, the FastSurfer parcellation images should undergo a rigorous, systematic parcellation quality checking procedure optimized for lesioned images, such as ENIGMA’s advanced guide for parcellation error identification (EAGLE-I) (56).

By assessing the accuracy of both VBG and FastSurfer in ms-TBI images, this study provides a foundation for moving toward evidence-backed decision-making to ensure the automated tools used during neuroimaging analysis in ms-TBI are valid for the unique pathology present in each sample. Improving the confidence of tool selection and the efficiency and accuracy with which quantitative brain metrics are calculated could ultimately lead to the clinical use of quantitative brain measures. Single-subject profiles of pathology, for example, provide a way of showing how an individual’s brain differs from what is normally expected by using the parcellation to define regions of interest and then examining the connections between them (37). Having this kind of personalized profile could help identify subgroups of patients with similar patterns of disruption, which, in turn, may support more tailored treatment planning and clearer predictions about long-term outcomes.