Axonal injury is the histopathological hallmark of traumatic brain injury (TBI)¹. Early ex vivo studies noticed that axonal changes frequently occur at sites where axons change their anatomical course, such as over blood vessels and within the gray/white matter (GM/WM) interface, where another factor that predisposes them to injury is the change in tissue density^2-4. Additionally, recently it was established that the histopathological hallmark of chronic traumatic encephalopathy (which is thought to be caused by repetitive TBI) are tau deposits at the depths of cerebral sulci⁵. Our goal, therefore, was to investigate whether changes consistent with injury at the GM/WM interface can be imaged in vivo. Since conventional diffusion tensor imaging (DTI) is not sensitive to crossing fiber damage, as at the GM/WM interface, we used proton MR spectroscopic imaging (¹H-MRSI), through quantification of the neuronal marker N-acetyl-aspartate (NAA), as well as of creatine (Cr), choline (Cho) and myo-inositol (mI) for energy and glial status. The strategy was to compare the degree of injury amongst voxels with different GM and WM partial volume, on a continuum from “pure” GM to “pure” WM. The hypothesis was that the amount of injury within WM voxels with small GM partial volume would be larger than the injury within “pure” WM voxels.Subjects: Fifteen symptomatic mild TBI patients (Table 1) and 12 age- and gender-matched controls. Data acquisition: T1-weighted MRI (MP-RAGE), T2-weighted MRI (FLAIR), B₀ shimming, 10×8×4.5 cm (AP×LR×IS)=360 cm^{3 1}H-MRS VOI (PRESS TE/TR=35/1800 ms), encoded to 480 voxels, each 1.0×1.0×0.75 cm³ (Fig. 1). Metabolite quantification: Absolute metabolite amounts were obtained using phantom replacement with correction for T1 and T2 relaxation time differences. Only voxels with metabolite Cramer Rao lower bounds<20% and 4<linewidths<13 Hz were retained. Segmentation: Global GM, WM and cerebro-spinal fluid (CSF) masks were obtained from the MP-RAGE using SPM. Mask thresholding: The masks were co-registered with the ¹H-MRSI matrix, yielding their volume in every ¹H-MRSI voxel (Fig. 2). Any voxels containing more than 10% CSF were excluded. Voxels were then grouped into 10 bins, each with a different fraction of WM, corresponding to a gradient from “pure” GM to “pure” WM. The bins were: 0-9% WM (“pure” GM, i.e. 81%<GM<100%, considering possible 0%<CSF<10%), 10-19% WM, etc. up to 90-100% WM (“pure” WM). For each of the bins, concentrations of each metabolite were compared between patients and controls with unequal variance t tests with voxel count as a weighting factor (i.e., giving greater weight to data values with higher voxel counts).The metabolite concentrations of patients and controls at each of the 10 bins are shown in Fig. 3. There were statistically significant differences only for NAA: within the 20-29% bin, as well as in all bins over 40-49%, as shown in Fig. 4. The regression of the mean NAA group difference to bin number showed a significant non-linear association (p=0.017 for the quadratic term). The regression model to predict the mean difference as a quadratic function of bin number was given by the equation: predicted mean NAA difference = -0.196 + 0.271*(bin number) - 0.0178*(bin number)exp(2) (Fig. 4). This regression equation explained 79.3% of the variance in the group mean NAA differences across bins and implied that the mean group difference increases as a function of bin number from a global minimum within bin 0-9% WM (“pure” GM) to a projected maximum within the 60-69% WM bin (at 60.6%) and decreases as a function of increasing bin number (WM fraction) thereafter.Our hypothesis was supported by the findings. First, the bin with the largest numerical difference between patients and controls had 80-89% WM. Second, the bin predicted by the fitted model function to contain the maximum difference was of 60-69% WM. Both observations indicate larger amount of injury within voxels with small partial GM content, compared to “pure” WM voxels. A limitation of the study is that it was impossible to confirm that within the mixed-tissue voxels, all WM eventually terminates within the GM voxel. This source of noise, however, is unavoidable, given the inherent course spatial resolution of ¹H-MRS, and was addressed by the large number of analyzed voxels. Our findings need confirmation in other cohorts, with ¹H-MRS, as well as with advanced DTI, sensitive to crossing fiber geometry.We present evidence consistent with neuronal damage at the site of the GM/WM interface in mild TBI. To our knowledge, this is the first attempt to image and differentiate this injury in vivo. This has important implications for injury detection in TBI, potentially revealing an injury locale not well characterized previously.