Contusions bias cortical thickness estimates after traumatic brain injury: A TRACK-TBI study Authors: Daniel Brennan, Andrea L.C. Schneider, Russell Taki Shinohara, Ramon Diaz-Arrastia, Philip A. Cook, James C. Gee, James J. Gugger Published: 2026 ABSTRACT Highlights•Contusions bias cortical thickness estimates after traumatic brain injury.•Lesion-related bias affects cross-sectional and longitudinal measures.•Lesion masking can attenuate cross-sectional cortical thickness bias.•Lesion-aware preprocessing may improve TBI neuroimaging biomarkers. Introduction Traumatic brain injury (TBI) is increasingly recognized as a chronic disease, with both acute and long-term damage contributing to persistent impairment. (Maas et al., 2022) The pathological sequelae of TBI are complex, often involving an interplay of focal and diffuse injuries. (McKee and Daneshvar, 2015) Focal injuries, such as contusions and hemorrhages, typically manifest early as a consequence of direct biomechanical impact and are usually well-demarcated on conventional clinically-acquired brain imaging modalities. (Harris et al., 2019, Parizel et al., 2020) In contrast, diffuse damage often arises through secondary pathophysiological processes which may evolve over time and is more difficult to detect with conventional imaging. (Parizel et al., 2020, Javeed et al., 2021, Jang, 2020) Given their divergent mechanisms and temporal courses, it is important to consider the independent contributions of these types of injury mechanisms to long-term functional outcomes. Automated morphometry is widely used to quantify structural brain changes following TBI (Brain Morphometric Techniques Applied to the Study of Traumatic Brain Injury. In: Neuromethods. Springer New York;, 2018) and offers the potential to estimate diffuse injury and track progressive neurodegeneration (i.e., atrophy) after injury. (Brennan et al., 2023, Mazaharally et al., 2021, Bigler, 2013, Sassani et al., 2025) However, TBI-related focal lesions present a major challenge for the accuracy of atrophy measurement after TBI by introducing imaging irregularities that are not well modeled by standard image processing pipelines. (Diamond et al., 2020) Most existing methods lack robustness to this TBI-related variability (i.e., heterogeneous anatomical changes across individuals (De Feo et al., 2022)), leading to potential failures in processing steps as fundamental as image registration. (Brett et al., 2001, Piçarra et al., 2023) Focal lesions are also known to induce segmentation and surface reconstruction errors, undermining the reliability of cortical thickness estimates surrounding contused tissue. (Ozzoude et al., 2020, Weiss et al., 2021) As such, surface-based pipelines such as FreeSurfer are particularly vulnerable, (King et al., 2020) as they rely on a continuous cortical ribbon to accurately model the gray matter surface boundaries and calculate cortical morphometry data. In cases where lesions produce discontinuities (e.g., cavities or holes), the surface model calculation can fail completely unless post-hoc corrections are applied (e.g., (Tustison et al., 2021)). Volumetric pipelines such as ANTs Cortical Thickness (Tustison et al., 2021) are theoretically less sensitive to focal deviations from normal tissue shape and appearance as they calculate morphometry directly from the volumetric image itself and do not require a continuous cortical surface model. Despite this, tissue misclassification still occurs, potentially leading to downstream biases in morphometric estimates. Consequently, the presence of contusions may lead to inaccurate assessments of diffuse injury and cortical atrophy post-TBI. The effects of focal lesions on cortical thickness data do appear idiosyncratic, with no clear effect of group-level statistics yet identified. (Bigler et al., 2016) However, given the well-documented heterogeneity of TBI, the frequent exclusion of participants with focal lesions from group level analyses, and the growing emphasis on precision medicine across neurological disorders, it is essential to identify and correct systematic measurement biases introduced by frequent pathologies such as contusions. Using data from the prospective, multi-center Transforming Research and Clinical Knowledge in Traumatic Brain Injury (TRACK-TBI) Study, we aimed to quantify the impact of focal cerebral lesions (i.e., contusions) on cortical thickness estimates and evaluated strategies to mitigate lesion-induced bias. We hypothesized that focal lesions would systematically bias cortical thickness estimates relative to the contralesional homologue at the earliest available timepoint, 2 weeks post-injury, when contusions are visible and exhibit heterogeneous, irregular signal characteristics. We also hypothesized that systematic biases would be present in estimates of longitudinal cortical thickness change from 2 weeks to 6 months post injury as these contusions resolve over time. Methods Study design and population The TRACK-TBI Study (ClinicalTrials.gov #NCT02119182) is a prospective cohort of acute TBI patients who presented to one of eighteen level 1 trauma centers in the United States and who underwent a non-contrast head computed tomography (CT) scan within 24 h of injury. (Yue et al., 2013) Participants were enrolled between February 26, 2014 and August 8, 2018 and underwent brain MRI scans at 2 weeks and 6 months post-injury. TRACK-TBI Study demographic, clinical, and imaging data were collected in accordance with the National Institutes of Health-National Institute of Neurological Disorders and Stroke TBI common data elements. The TRACK-TBI Study was approved by the institutional review board at each enrolling site, and all participants (or their legally authorized representatives) provided written informed consent. De-identified TRACK-TBI demographic, clinical, and imaging data were obtained from the Federal Interagency Traumatic Brain Injury Research (FITBIR) Informatics System. Of the 2526 TRACK-TBI Study participants with TBI aged >=18 available on FITBIR, we excluded 2048 who did not have structural brain MRI, 390 without contusions on MRI, and two participants with no 6-month structural MRI resulting in N = 86 participants included in our analytic sample. Brain magnetic resonance imaging All subjects were screened for MRI contraindications immediately before the MRI scan using the local standard protocol. MRI contraindications included ferrous metallic implants, pacemakers, defibrillators, or other non-MRI-compatible devices. Sedation was not permitted. Participants were positioned in the scanner with the head aligned at the nasion using foam padding and optional forehead straps to minimize motion. A 3-plane localizer and scanner-specific calibration scan were acquired at the start of each session. Structural imaging included two required sequences: sagittal 3D T1-weighted (MPRAGE for Siemens/Philips; IR-SPGR for GE) and sagittal 3D T2-FLAIR (SPACE for Siemens, VISTA for Philips, or CUBE for GE). Both sequences were acquired in straight sagittal planes with no oblique angulation. Prescriptions were based on 3-plane scouts and required full head coverage, including skull, cerebellum, and nose. Acquisition boxes were adjusted to avoid wrap-around and motion artifacts. Sites were required to scan on TRACK-TBI-qualified 3 T MRI systems using pre-approved protocols. Brain MRI images were acquired at approximately 2 weeks and 6 months post-injury. Contusions were drawn by a board-certified neurologist using co-registered T1 and FLAIR images in the T1 space. As most lesions present at 2 weeks had resolved by 6 months, the lesion analysis was constrained to the 2-week timepoint, with 6-month data used to estimate lesion influences on longitudinal estimation of cortical thickness (i.e., atrophy). Cortical thickness estimation Cortical thickness was estimated from MRIs performed at 2 weeks and 6 months post-injury using longitudinal ANTsNetCT [https://github.com/ftdc-picsl/antsnetct] implementation of ANTsPyNet deep learning longitudinal thickness pipeline. (Tustison et al., 2021, Tustison et al., 2019) Default parameters were used, with the “SyN” template building option enabled. A custom post-processing script was applied (see Data Availability) to derive final cortical thickness maps. An advantage of this pipeline is the deep learning application of tissue segmentation for the single-subject template, which does not require registration-based tissue prior estimation. The Desikan-Killiany-Tourville (DKT) atlas (Klein and Tourville, 2012) was applied to original-space T1 using deep learning tools available via ANTsPyNet, to summarize cortical thickness at the region of interest (ROI) level. Lesion presence was also identified at the ROI level: if any lesion volume was present within any portion of the ROI, the ROI was considered to contain a lesion. The proportion of the ROI containing the lesion (i.e., lesion volume within ROI / total ROI volume) was also calculated. This lesion proportion metric was used in an exploratory sensitivity analysis to assess whether greater lesion burden was associated with larger contusion-related differences in cortical thickness or longitudinal thickness change. Contusion correction Three levels of lesion masking were evaluated, visualized in Fig. 1. First, cortical thickness was estimated with no correction applied (i.e., the default pipeline; “No Masking”). Second, post-hoc correction of the cortical thickness data was evaluated: in this approach the cortical thickness data was zeroed within the lesion mask, retaining only the cortical thickness data outside the lesioned region of the ROI (i.e., Atlas Masking). Third, a tissue probability correction was applied: here, the underlying tissue estimates (gray matter/white matter/cerebrospinal fluid) within the lesion mask were set to 0 with resulting tissue probability map processed as usual by the DiReCT cortical thickness algorithm (Rebsamen et al., 2020) (i.e., Full Masking). The summarized cortical thickness within each DKT ROI was used as the dependent variable in the analysis. Note, that because gray matter segmentation was removed via the lesion mask, cortical thickness was only estimated from non-lesioned cortex in the Full Masking Approach by default.Fig. 1Visual depiction of each lesion masking strategy. A) Focal lesion evident on structural MRI. Contusion core (Red Arrow) and edema (Green Arrow) are identified and masked in subsequent masking strategies. B) Cortical thickness map with lesion included in the estimation (No Masking). Note, the location and relative intensity of the contusion edema has caused misattribution of these voxels as gray matter and estimation of cortical thickness in this area (Green Arrow). C) Desikan-Killiany-Tourville parcellation with lesion mask removed, which excludes data from within the lesion at the level of data summarization (Atlas Masking). (D) Tissue segmentation, cortical thickness and cortical parcellation maps with lesion fully excluded (Full Masking). In this strategy, the lesion mask is removed from tissue segmentation. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Visual depiction of each lesion masking strategy. A) Focal lesion evident on structural MRI. Contusion core (Red Arrow) and edema (Green Arrow) are identified and masked in subsequent masking strategies. B) Cortical thickness map with lesion included in the estimation (No Masking). Note, the location and relative intensity of the contusion edema has caused misattribution of these voxels as gray matter and estimation of cortical thickness in this area (Green Arrow). C) Desikan-Killiany-Tourville parcellation with lesion mask removed, which excludes data from within the lesion at the level of data summarization (Atlas Masking). (D) Tissue segmentation, cortical thickness and cortical parcellation maps with lesion fully excluded (Full Masking). In this strategy, the lesion mask is removed from tissue segmentation. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Statistical analysis In subsequent analyses, each bilateral cortical region defined by the DKT atlas was treated as a matched pair comprised of the left and right homologs. When bilateral lesions were present within a given region for a participant, observations for that region/participant were excluded (102/473 region-by-participant observations excluded). The remaining data comprised 371 lesion-contralesion ROI pairs, with participants contributing one or more paired observations across regions. To ensure stable estimation, only regions with at least 10 valid lesion-contralesion pairs were retained for modeling, resulting in 14 of 27 unique lesioned ROIs. Two complementary analyses were performed to quantify lesion-induced bias in cross-sectional cortical thickness and longitudinal cortical thinning. First, independent models were fit for each region using the LMER package (Bates et al., 2015) in R, with statistical significance estimated via the Satterthwaite method (Luke, 2017, Giesbrecht and Burns, 1985, Alex and Cornelius, 1996) for degrees of freedom. Multiple comparison correction across the 14 independent region-wise models was applied via Benjamini-Hochberg false-discovery rate (FDR) correction. (Benjamini and Hochberg, 1995). Second, to evaluate the overall magnitude of lesion-related bias and the effectiveness of the lesion-correction strategies, the lesion-presence regression coefficients (β terms) estimated from each ROI model were compared across masking strategies. To assess whether these coefficients systematically differed from zero across all evaluated ROI, a nonparametric bootstrap analysis was conducted by resampling participant-level data with replacement to generate empirical distributions of the mean lesion-associated bias within each lesion masking strategy. Bias was considered significant when the resulting 95% empirical confidence interval did not include 0, representing a systematic lesion-related deviation in cortical thickness estimates compared with contralesional homologue. Estimating lesion-induced bias in cortical thickness cross-sectionally at 2 weeks post-injury To estimate lesion-induced bias in cortical thickness at 2 weeks post-injury, we compared cortical thickness between lesioned and contra-lesional hemispheres within each bilateral ROI pair using the form: Cortical_Thickness=Lesion + Hemisphere + Age + Sex + (1 | Subject) where “Cortical_Thickness” represents the estimated mean thickness within each ROI, “Lesion” is an indicator variable representing the presence or absence of a lesion in that observation, “Hemisphere” is the observation’s hemisphere regardless of lesion, and “Subject” is included as a random intercept to account for within-participant correlations between the lesion/contralesion pairs. The regression coefficient for “Lesion” quantifies the lesion-associated bias in cortical thickness relative to the contralesional hemisphere; i.e., a significantly positive or negative value represents over- or under-estimation of cortical thickness in the lesioned hemisphere compared with the contra-lesional homologue. For the bootstrap analysis, identical models were refit to 10,000 bootstrapped participant-level samples, preserving bilateral pairing, to derive empirical distributions of the mean lesion coefficient for inference. Estimating lesion-induced bias in longitudinal cortical thickness change between 2 weeks and 6 months post-injury We hypothesized that the initial bias in cortical thickness in lesioned ROIs at 2 weeks would lead to artifactual cortical thickness change compared with the contralesional hemisphere (e.g., inflated estimates of progressive cortical thinning). For each ROI and participant, we estimated the slope of cortical thickness change expressed as daily rate of change: Cortical_Thickness_SlopeROI=(Thickness6mo − Thickness2wk)/(Days6mo-Days2wk) These per-ROI slopes were used as the dependent variable in analogous mixed-effects models: Cortical_Thickness_Slope=Lesion + Hemisphere + Age + Sex + (1 | Subject) The regression coefficient for “Lesion” again quantifies the lesion-associated bias in cortical thickness change-slope relative to that of the contralesional hemisphere; i.e., a significantly positive or negative value represents over- or under-estimation of the rate of cortical thickness change in the lesioned hemisphere compared with the contra-lesional homologue. As in the cross-sectional analyses, the described model was refit to 10,000 bootstrapped participant-level samples, generating empirical distributions of the lesion-effect coefficients used to quantify overall longitudinal bias across masking strategies. Results Participant and focal lesion characteristics Demographic and clinical characteristics of the included TRACK-TBI participants are summarized in Table 1. The analytic sample comprised 86 participants with a mean age of 39.1 years (SD = 16.2), of whom 25.6% were female. 77 participants had a Glasgow Coma Scale (GCS) score of 13–15 and 7 participants had a GCS score < 13; GCS data were unavailable for 2 participants. Mean time from injury to MRI at the 2-week time point was 14.4 days (SD = 2.0), and the mean time from injury to MRI at the 6-month time point was 184.5 days (SD = 8.5). The distribution of focal lesions at the ROI level is reported in Table 2. Summary of additional injury characteristics, including mechanism of injury and lesion characterization, are provided in Supplementary Table 1.Table 1Cohort Demographics.CharacteristicValueN (participants)86Age, mean (SD), years39.1 (16.2)Female, %25.6%Glasgow Coma Scale = 13–15*77Glasgow Coma Scale < 13*7Time from injury, mean (SD), days − 2 weeks14.4 (2.0)Time from injury, mean (SD), days − 6 months184.5 (8.5)Demographic and clinical characteristics of the TRACK-TBI sample. Values are reported as mean (SD), median, or percentage as appropriate. Time from injury represents the mean number of days elapsed between injury and scan acquisition at the 2-week and 6-month study visits. * GCS was not available for two participants.Table 2Lesion count per ROI.ROINROINMiddletemporal43Parsopercularis9Superiortemporal41Caudalmiddlefrontal8Inferiortemporal38Transversetemporal8Lateralorbitofrontal28Supramarginal8Parsorbitalis23Lateraloccipital6Rostralmiddlefrontal23Inferiorparietal6Superiorfrontal17Parahippocampal5Medialorbitofrontal16Rostralanteriorcingulate5Parstriangularis15Superiorparietal5Entorhinal14Lingual4Precentral13Precuneus2Fusiform11Cuneus1Insula11Paracentral1Postcentral10Number of observations available for each ROI across the final sample. Values reflect the count of lesioned ROIs contributing to the analyses. ROIs with ≥ 10 observations are shown in the left panel and are included in the analyses. Regions with fewer observations are presented in the right panel and are excluded from subsequent analyses. Table 1Cohort Demographics.CharacteristicValueN (participants)86Age, mean (SD), years39.1 (16.2)Female, %25.6%Glasgow Coma Scale = 13–15*77Glasgow Coma Scale < 13*7Time from injury, mean (SD), days − 2 weeks14.4 (2.0)Time from injury, mean (SD), days − 6 months184.5 (8.5)Demographic and clinical characteristics of the TRACK-TBI sample. Values are reported as mean (SD), median, or percentage as appropriate. Time from injury represents the mean number of days elapsed between injury and scan acquisition at the 2-week and 6-month study visits. * GCS was not available for two participants. Cohort Demographics. Demographic and clinical characteristics of the TRACK-TBI sample. Values are reported as mean (SD), median, or percentage as appropriate. Time from injury represents the mean number of days elapsed between injury and scan acquisition at the 2-week and 6-month study visits. * GCS was not available for two participants. Table 2Lesion count per ROI.ROINROINMiddletemporal43Parsopercularis9Superiortemporal41Caudalmiddlefrontal8Inferiortemporal38Transversetemporal8Lateralorbitofrontal28Supramarginal8Parsorbitalis23Lateraloccipital6Rostralmiddlefrontal23Inferiorparietal6Superiorfrontal17Parahippocampal5Medialorbitofrontal16Rostralanteriorcingulate5Parstriangularis15Superiorparietal5Entorhinal14Lingual4Precentral13Precuneus2Fusiform11Cuneus1Insula11Paracentral1Postcentral10Number of observations available for each ROI across the final sample. Values reflect the count of lesioned ROIs contributing to the analyses. ROIs with ≥ 10 observations are shown in the left panel and are included in the analyses. Regions with fewer observations are presented in the right panel and are excluded from subsequent analyses. Lesion count per ROI. Number of observations available for each ROI across the final sample. Values reflect the count of lesioned ROIs contributing to the analyses. ROIs with ≥ 10 observations are shown in the left panel and are included in the analyses. Regions with fewer observations are presented in the right panel and are excluded from subsequent analyses. Lesion-induced bias in cortical thickness at 2 weeks post-injury The mean absolute difference in cortical thickness between hemispheres with no lesions at 2 weeks post-injury was 0.01 mm (SE = 0.0175). While the brain is not symmetrical, we were unable to detect any systematic hemispheric asymmetry and expected no absolute difference between lesioned ROIs and the corresponding ROI in the contralesional hemisphere. Any difference between the hemispheres was interpreted as lesion-induced bias. Following multiple-comparisons correction, significant lesion-associated bias was observed in no-masking in six cortical regions: medial orbitofrontal, middle temporal, pars triangularis, postcentral, superior frontal, and superior temporal cortices (Table 3). Across all ROIs, a systematic overestimation of cortical thickness was observed, as indicated by the bootstrap analysis (median β = 0.22, mean β = 0.22, 95% CI = [0.12, 0.29]). This bias persisted under the atlas-masking strategy (median β = 0.17, mean β = 0.17, 95% CI = [0.10, 0.24]). In contrast, the full lesion-masking approach effectively eliminated the overall bias across ROIs (median β = -0.013, mean β = -0.06, 95% CI = [-0.13, 0.01]). These results are summarized in Table 4 and visualized in Fig. 2.Table 3ROI-wise Bias Estimates in CT values at two-weeks post-injury.ROIβLesionset-valuep-valueq-valueMiddletemporal0.240.092.830.0070.027Superiortemporal0.190.072.750.0090.027Inferiortemporal0.100.100.960.3430.37Lateralorbitofrontal0.150.081.830.0780.119Parsorbitalis0.160.131.220.2280.266Rostralmiddlefrontal0.230.112.070.050.088Superiorfrontal0.210.063.580.0030.027Medialorbitofrontal0.430.152.930.010.027Parstriangularis0.510.182.90.0120.027Entorhinal0.040.210.210.8340.834Precentral0.210.082.540.0260.052Fusiform0.400.271.490.1540.196Insula0.430.231.910.0850.119Postcentral0.320.103.160.0120.027Note. ROI-wise linear mixed-effects model results for lesion-associated bias in cortical thickness at 2 weeks post-injury. Significant results after FDR correction (q-value < 0.05) are shown in bold. βLesion represents fixed effect coefficient associated with the lesion effect in the linear mixed effects model for each ROI. se = standard error.Table 4Global Bias Estimates (across all ROIs) in CT values at two-weeks post-injury.Masking StrategyMean Beta (mm)Empirical 95% CI LowerEmpirical 95% CI UpperNo Masking0.220.120.29Atlas Masking0.170.100.24Full Masking−0.06−0.130.01Note. Summary of the empirical distribution of systematic lesion-associated bias across masking strategies based on a nonparametric bootstrap procedure. 95% confidence interval (CI) for the empirically estimated distribution of the mean beta for lesion-associated bias regression term. Units are millimeter (mm) cortical thickness.Fig. 2Observed ROI-wise lesion-associated bias (β coefficients) in cortical thickness across masking strategies at 2 weeks post-injury. Distributions shown for the No Masking, Atlas Masking, and Full Masking strategies. Values closer to zero indicate reduced lesion-induced bias in cortical thickness estimates. Beta values greater than 0 represent overestimation of cortical thickness relative to the contralesional hemisphere; values less than 0 represent underestimation. Table 3ROI-wise Bias Estimates in CT values at two-weeks post-injury.ROIβLesionset-valuep-valueq-valueMiddletemporal0.240.092.830.0070.027Superiortemporal0.190.072.750.0090.027Inferiortemporal0.100.100.960.3430.37Lateralorbitofrontal0.150.081.830.0780.119Parsorbitalis0.160.131.220.2280.266Rostralmiddlefrontal0.230.112.070.050.088Superiorfrontal0.210.063.580.0030.027Medialorbitofrontal0.430.152.930.010.027Parstriangularis0.510.182.90.0120.027Entorhinal0.040.210.210.8340.834Precentral0.210.082.540.0260.052Fusiform0.400.271.490.1540.196Insula0.430.231.910.0850.119Postcentral0.320.103.160.0120.027Note. ROI-wise linear mixed-effects model results for lesion-associated bias in cortical thickness at 2 weeks post-injury. Significant results after FDR correction (q-value < 0.05) are shown in bold. βLesion represents fixed effect coefficient associated with the lesion effect in the linear mixed effects model for each ROI. se = standard error. ROI-wise Bias Estimates in CT values at two-weeks post-injury. Note. ROI-wise linear mixed-effects model results for lesion-associated bias in cortical thickness at 2 weeks post-injury. Significant results after FDR correction (q-value < 0.05) are shown in bold. βLesion represents fixed effect coefficient associated with the lesion effect in the linear mixed effects model for each ROI. se = standard error. Table 4Global Bias Estimates (across all ROIs) in CT values at two-weeks post-injury.Masking StrategyMean Beta (mm)Empirical 95% CI LowerEmpirical 95% CI UpperNo Masking0.220.120.29Atlas Masking0.170.100.24Full Masking−0.06−0.130.01Note. Summary of the empirical distribution of systematic lesion-associated bias across masking strategies based on a nonparametric bootstrap procedure. 95% confidence interval (CI) for the empirically estimated distribution of the mean beta for lesion-associated bias regression term. Units are millimeter (mm) cortical thickness. Global Bias Estimates (across all ROIs) in CT values at two-weeks post-injury. Note. Summary of the empirical distribution of systematic lesion-associated bias across masking strategies based on a nonparametric bootstrap procedure. 95% confidence interval (CI) for the empirically estimated distribution of the mean beta for lesion-associated bias regression term. Units are millimeter (mm) cortical thickness. Observed ROI-wise lesion-associated bias (β coefficients) in cortical thickness across masking strategies at 2 weeks post-injury. Distributions shown for the No Masking, Atlas Masking, and Full Masking strategies. Values closer to zero indicate reduced lesion-induced bias in cortical thickness estimates. Beta values greater than 0 represent overestimation of cortical thickness relative to the contralesional hemisphere; values less than 0 represent underestimation. Lesion-Induced bias in cortical thickness change from 2 weeks to 6 months post-injury In the unmasked data, significant differences in the longitudinal cortical thickness slope were observed between the lesional and contralesional hemispheres in the fusiform, superior temporal, superior frontal, middle temporal, entorhinal, and inferior temporal cortices (Table 5). There was a significant underestimation of the change slope across all evaluated regions using the unmasked data (median observed β = -0.00084, bootstrapped mean β = -0.00078, 95% CI = [-0.00101, −0.00055]), and the atlas-masked data (median observed β = -0.00050, bootstrapped mean β = -0.00064, 95% CI = [-0.00088, −0.00040]). There was a significant overestimation of the change slope in the fully masked data (median observed β = 0.00048, bootstrapped mean β = 0.00062, 95% CI = [0.00033, 0.00088]). These results are summarized in Table 6 and visualized in Fig. 3.Table 5ROI-wise Bias Estimates in Longitudinal CT.ROIβLesionset-valuep-valueq-valueMiddletemporal−0.000970.000278−3.500.0010.004Superiortemporal−0.001040.000273−3.80< 0.0010.003Inferiortemporal−0.000720.000285−2.530.0160.039Lateralorbitofrontal−0.000740.000337−2.190.0370.066Parsorbitalis−0.000660.000521−1.270.2170.304Rostralmiddlefrontal−0.000570.000514−1.100.2820.359Superiorfrontal−0.000790.000191−4.120.0010.004Medialorbitofrontal−0.000930.000393−2.360.0260.052Parstriangularis−0.000890.000925−0.960.3440.401Entorhinal−0.002450.000891−2.750.0170.039Precentral−0.000020.000554−0.040.9720.972Fusiform−0.002710.000586−4.63< 0.0010.003Insula−0.002190.001188−1.840.0830.129Postcentral0.000640.0011450.560.5870.633Note. ROI-wise linear mixed-effects model results for lesion-associated bias in longitudinal cortical thickness change slope (mm/day). Significant results after FDR correction are shown in bold. βLesion represents fixed effect coefficient associated with the longitudinal lesion effect in the linear mixed effects model for each ROI. se = standard error.Table 6Global Bias Estimates (across all ROIs) in Longitudinal CT.Masking StrategyMean Beta (mm/day)Empirical 95% CI LowerEmpirical 95% CI UpperNo Masking−0.00078−0.00101−0.00055Atlas Masking−0.00064−0.00088−0.00040Full Masking0.000620.000330.00088Note. Summary of the empirical distribution of systematic lesion-associated bias across masking strategies based on a nonparametric bootstrap procedure. 95% confidence interval (CI) for the empirically estimated distribution of the mean beta for lesion-associated bias regression term. Units are mm (cortical thickness)/day.Fig. 3Observed ROI-wise lesion-associated bias (β coefficients) in longitudinal cortical thickness change slope (mm/day) across masking strategies. Distributions shown for the No Masking, Atlas Masking, and Full Masking strategies. Values closer to zero indicate reduced lesion-induced bias in estimated cortical change slope. Positive values represent overestimation of regional cortical thickness change slope relative to the homologue; negative values represent underestimation. Table 5ROI-wise Bias Estimates in Longitudinal CT.ROIβLesionset-valuep-valueq-valueMiddletemporal−0.000970.000278−3.500.0010.004Superiortemporal−0.001040.000273−3.80< 0.0010.003Inferiortemporal−0.000720.000285−2.530.0160.039Lateralorbitofrontal−0.000740.000337−2.190.0370.066Parsorbitalis−0.000660.000521−1.270.2170.304Rostralmiddlefrontal−0.000570.000514−1.100.2820.359Superiorfrontal−0.000790.000191−4.120.0010.004Medialorbitofrontal−0.000930.000393−2.360.0260.052Parstriangularis−0.000890.000925−0.960.3440.401Entorhinal−0.002450.000891−2.750.0170.039Precentral−0.000020.000554−0.040.9720.972Fusiform−0.002710.000586−4.63< 0.0010.003Insula−0.002190.001188−1.840.0830.129Postcentral0.000640.0011450.560.5870.633Note. ROI-wise linear mixed-effects model results for lesion-associated bias in longitudinal cortical thickness change slope (mm/day). Significant results after FDR correction are shown in bold. βLesion represents fixed effect coefficient associated with the longitudinal lesion effect in the linear mixed effects model for each ROI. se = standard error. ROI-wise Bias Estimates in Longitudinal CT. Note. ROI-wise linear mixed-effects model results for lesion-associated bias in longitudinal cortical thickness change slope (mm/day). Significant results after FDR correction are shown in bold. βLesion represents fixed effect coefficient associated with the longitudinal lesion effect in the linear mixed effects model for each ROI. se = standard error. Table 6Global Bias Estimates (across all ROIs) in Longitudinal CT.Masking StrategyMean Beta (mm/day)Empirical 95% CI LowerEmpirical 95% CI UpperNo Masking−0.00078−0.00101−0.00055Atlas Masking−0.00064−0.00088−0.00040Full Masking0.000620.000330.00088Note. Summary of the empirical distribution of systematic lesion-associated bias across masking strategies based on a nonparametric bootstrap procedure. 95% confidence interval (CI) for the empirically estimated distribution of the mean beta for lesion-associated bias regression term. Units are mm (cortical thickness)/day. Global Bias Estimates (across all ROIs) in Longitudinal CT. Note. Summary of the empirical distribution of systematic lesion-associated bias across masking strategies based on a nonparametric bootstrap procedure. 95% confidence interval (CI) for the empirically estimated distribution of the mean beta for lesion-associated bias regression term. Units are mm (cortical thickness)/day. Observed ROI-wise lesion-associated bias (β coefficients) in longitudinal cortical thickness change slope (mm/day) across masking strategies. Distributions shown for the No Masking, Atlas Masking, and Full Masking strategies. Values closer to zero indicate reduced lesion-induced bias in estimated cortical change slope. Positive values represent overestimation of regional cortical thickness change slope relative to the homologue; negative values represent underestimation. Exploratory analysis between lesion proportion and lesion-induced bias In an exploratory analysis, we evaluated whether regional lesion burden was associated with the magnitude of lesion-induced bias in cortical thickness. In the No Masking approach, greater lesion proportion was associated with larger 2-week cortical thickness differences (Spearman ρ = 0.49, p < 0.001) and more negative longitudinal cortical thickness change-slope differences (ρ = -0.38, p < 0.001), suggesting that lesion burden contributes to the magnitude of measurement bias when lesions are not explicitly accounted for. In contrast, the relationship between lesion proportion and 2-week cortical thickness bias was attenuated after Full Masking (ρ = 0.11, p = 0.31), consistent with reduced lesion-related bias using lesion-aware processing. However, because relatively few ROIs had high lesion burden, particularly lesion proportions greater than 0.5, these analyses were not sufficient to define a lesion-volume threshold above which cortical thickness estimation becomes unreliable. Discussion Robust MRI processing in the presence of focal pathology remains a critical challenge in study of neurotrauma. (Deutscher et al., 2025, Wilde et al., 2021) By identifying a systematic bias in the estimation of cortical thickness in the presence of focal lesions, a common pathology following TBI, this study highlights a key limitation of current morphometric methods and underscores the need for more robust processing approaches. Using the contralesional hemisphere as an internal reference, we found that cortical thickness in lesioned ROI was overestimated by an average of 0.21 mm at 2 weeks post-injury. There was regional variation in the magnitude of this effect across the brain, with only a subset of ROI reaching statistical significance for this lesional bias despite a consistent directional effect across all ROI. A lesion-associated bias was also observed in estimations of longitudinal cortical thickness change, with a significant subset of ROI showing a larger degree of cortical thickness change compared with the contralesional hemisphere. We evaluated two practical correction strategies to address this bias. Excluding lesion voxels from the tissue probability maps prior to cortical thickness estimation (Full Masking) effectively eliminated detectable cross-sectional biases in cortical thickness, zero-centering the overall distribution of differences between lesional and contra-lesional data. In contrast, masking lesions only from the cortical atlas (Atlas Masking) did not adequately correct this bias. Thus, approaches that do not consider these abnormalities at the level of data estimation (i.e., in early/upstream preprocessing steps rather than post-hoc) risk propagating biases into perilesional data, distorting both the spatial distribution and longitudinal trajectory of cortical thickness changes. While significant differences between lesional and contralesional estimates of longitudinal change remain in the full-masking approach, multiple considerations should be taken into account when interpreting these findings. First, in the evaluations of cortical thickness bias at 2 weeks post-injury, it was observed that cortical thickness estimates were inflated in the presence of the lesion, while estimates of cortical thickness change (using the 2-week data as baseline) are significantly more negative in the longitudinal evaluation. These observations are consistent with the hypothesis that cortical thickness increases related to lesion presence and tissue misclassification may lead to erroneously greater estimates (larger negative magnitude) of cortical thinning as the lesion resolves or these image abnormalities change over time. In contrast, while significant ROI-wise variation remains in the fully-masked data at 2 weeks post-injury, the baseline 2-week cortical thickness data are not significantly biased; this suggests that full masking may reduce lesion-related inflation of baseline cortical thickness estimates and partially mitigate downstream bias in longitudinal change, though residual regional differences remain. The observed systematic biases have broad implications for neuroimaging studies of brain injury. Cortical thickness is widely used as a marker of diffuse neurodegeneration following TBI, yet our findings highlight a potential source of anti-conservative bias introduced by co-occurring focal lesions. Without appropriate correction, apparent longitudinal cortical thinning induced by lesion resolution may be misinterpreted as ongoing neurodegeneration, leading to erroneous conclusions about the extent and progression of atrophy and potentially obscuring treatment effects or inflating the degree of disease progression. These issues are magnified with increasing injury burden, as both the extent of focal damage and the degree of progressive degeneration are expected to be greater in more severe injury, substantially elevating the risk of misattribution. Several strategies have been proposed to replace abnormal tissue and facilitate MRI analysis across neurological populations. Virtual Brain Grafting, for example, has been applied in stroke populations by reconstructing large lesions with contralateral homologous tissue. (Radwan et al., 2021, Bey et al., 2024) In TBI, where bilateral damage is common, the contralateral hemisphere may also be damaged, complicating the application of such methods; indeed, we identified about 22% of all potential lesioned ROI to have bilateral lesion presence. More recently, deep learning-based in-painting approaches have been introduced to reconstruct irregular or otherwise damaged tissue (e.g.,Kang et al., 2021, Almansour et al., 2021). For example, FastSurfer-LIT (Pollak et al., 2025) has demonstrated the ability to reconstruct pre-operative cortical thickness estimates from post-operative MRI scans in glioma patients by infilling recessed tissue areas following temporo-mesial resection surgery. Such approaches have not yet been systematically applied and evaluated in TBI, and their validity remains uncertain due to the heterogeneous nature of brain injuries. Our proposed method makes far fewer assumptions about structural imaging data, leveraging data exclusively from non-lesioned areas, providing a more robust strategy for assessing biases related to focal abnormalities and providing a baseline correction strategy against which more advanced lesion-correction approaches may be compared. The importance of accurate tissue classification extends beyond morphometry. BOLD functional MRI and perfusion imaging techniques, such as arterial spin labeling, are primarily interpreted within gray matter and are susceptible to partial volume effects. (Chappell et al., 2021, Dukart and Bertolino, 2014) Particularly in the context of neurodegeneration or focal injury, these modalities may be adversely impacted by inaccurate gray matter estimation. As neuroimaging increasingly becomes an essential tool for precision medicine, particularly in TBI, ensuring the accuracy and reliability of image processing pipelines in the context of neurotrauma is critical. The approach described here provides a practical and flexible solution that enhances the accuracy of cortical thickness measurements and their downstream interpretations in the setting of focal lesions. Importantly, this method establishes a transparent and computationally lightweight baseline upon which more sophisticated image interpolation and bias-correction strategies may be compared. Broader adoption of such practices will be essential to ensure the reproducibility and clinical relevance of neuroimaging biomarkers in TBI and related conditions. Limitations Several limitations should be considered when interpreting these findings. First, the lesion distribution and cortical thickness measurements described in this study were derived from a TBI study sample and may not fully generalize to broader neurological populations. Second, we were unable to utilize a “ground-truth” measurement of cortical thickness from which lesion bias could be calculated; instead, we used an internal control of the contralesional hemisphere. While our data suggest there is no expected hemispheric bias outside of lesioned ROIs, effects on the DiReCT cortical thickness near lesions could be introduced by spatial smoothing or regularization, which could be spatially heterogeneous and difficult to estimate. Indeed, evaluation of these methods relied on unilateral lesions, whereas bilateral lesions were excluded. This selection criterion may limit the generalizability of our findings to brains with more extensive damage or injury patterns that commonly produce bilateral lesions. Although lesion burden was associated with measurement bias in exploratory analyses, relatively few regions had lesions occupying more than 50% of the ROI. This limited our ability to define a lesion-burden threshold above which contusion correction is necessary. Finally, the method relies on the assumption that perilesional cortical thickness is representative of the entire cortical region, such that the mean thickness estimated from the preserved tissue reflects what would have been observed in the absence of a lesion. Given the disproportionate vulnerability of gyral crests to focal damage, (McGinn and Povlishock, 2016) this assumption may be violated. While the proposed lesion-correction strategy appears to reduce lesion-induced bias in cortical thickness estimates, future studies in neurotrauma should incorporate sensitivity analyses to assess the robustness of these corrections (e.g., excluding lesioned participants/regions altogether) to guard against potential unrecognized sources of bias. Conclusions In summary, focal lesions represent a significant yet addressable source of bias in cortical thickness estimation and longitudinal morphometric analysis. Implementing lesion-aware processing steps will be essential for improving the accuracy, interpretability, and clinical relevance of neuroimaging studies in TBI and other neurologic conditions. CRediT authorship contribution statement Daniel Brennan: Writing – original draft, Visualization, Investigation, Formal analysis, Conceptualization. Andrea L.C. Schneider: Writing – review & editing, Supervision, Resources, Methodology, Formal analysis, Conceptualization. Russell Taki Shinohara: Writing – review & editing, Validation, Supervision, Methodology, Investigation, Formal analysis, Conceptualization. Ramon Diaz-Arrastia: Writing – review & editing, Resources, Investigation, Conceptualization. Philip A. Cook: Writing – review & editing, Supervision, Software, Methodology, Investigation, Conceptualization. James C. Gee: Writing – review & editing, Software, Funding acquisition, Formal analysis, Conceptualization. James J. Gugger: Writing – review & editing, Supervision, Resources, Methodology, Investigation, Funding acquisition, Formal analysis, Data curation, Conceptualization. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.