Participants
Participants were recruited from the outpatient clinic of the neuropsychology unit at the Department of Psychiatry and the Department of Neurosurgery, Kyoto University Hospital.
Inclusion criteria were: 1) age more than 18 years, 2) an injury sustained through significant trauma; 3) a brain MRI or CT scan showing possible diffuse pathology (without large focal lesions (>10 mm3); 4) the injury occurred ≥ 6 months before the study; 5) ability to give informed consent for participation; 6) ability to undergo MRI. Exclusion criteria were: 1) history of another TBI with altered consciousness; 2) history of drug or alcohol abuse; 3) history of neurological or psychiatric disorder before TBI onset; 4) contraindications to MRI (e.g., implanted metal, claustrophobia). Neuropsychiatrists (UK, TM) specialized in the neuropsychiatric aspects of TBI assessed patients MRI and CT scans and confirmed the information concerning the clinical history and residual symptoms related to the inclusion and exclusion criteria mentioned above.
Twenty-six patients with TBI (20 males, mean age of 40.15 years, standard deviation [SD] 14.93) were recruited to the study. According to the Glasgow Coma Scale (GCS) or Japan Coma Scale (JCS; a measure of the severity of impaired consciousness used in Japan), 5 patients (19.2%) had mild TBI, 2 (7.7%) had moderate TBI, and 19 (73.1%) had severe TBI. The relationship between the JCS score and severity of injury has been explored previously [33]. For comparison purposes, twenty-six age- and sex matched healthy controls (20 males, mean age of 40.08 years, SD 12.82) were recruited to the study.
Imaging acquisition
For DTI, diffusion-weighted volumes were acquired on a 3.0-T whole body scanner (MAGNETOM Tim Trio; Siemens, Erlangen, Germany) using a 40-mT/m gradient and a receiver-only eight-channel phased-array head coil. The scanning parameters were: echo time (TE) = 96 ms, repetition time (TR) = 10 500 ms, matrix = 96 × 96, field of view = 192 × 192 mm, 70 contiguous axial slices of 2.0-mm thickness, 81 non-collinear axis motion-probing gradients, and b = 1500 s/mm2. The b = 0 images were acquired before each set of nine diffusion-weighted images, thus giving 90 volumes in total.
Image processing
DTI data were processed using the FMRIB Software Library (FSL) version 5.2 (FSL; Oxford Centre for Functional MRI of the Brain; www.fmrib.ox.ac.uk/fsl) [34,35]. DTI images were registered to the b = 0 image by affine transformations to minimize distortion due to head motion and eddy currents [35]. Images were then brain-extracted using the Brain Extraction Tool [36]. Using the DTIFIT program, the diffusion tensors were calculated for whole brain volumes and fractional anisotropy (FA) maps, axial diffusivity maps (AD [λ1]), radial diffusivity maps (RD [(λ2+λ3)/2]), and mean diffusivity maps (MD [(λ1+ λ2+λ3)/3]) were generated [37- 40]. MD, AD and RD are in units of mm2/s, whereas FA is unitless. Tract-Based Spatial Statistics (TBSS) was used for voxel-wise analysis of the DTI indices maps [40]. TBSS is a fully automated whole-brain analysis technique for applying voxel-wise statistics to diffusion indices while minimizing the effects of misalignment that may occur using a conventional voxel-based analysis method [40].
The TBSS procedure included nonlinear registration of all subjects FA images into the common FMRIB58_FA template space [39]. TBSS performs a non-linear registration to align each FA image to every other one and then calculates the amount of warping needed for the images to be aligned. The most representative image is determined as the one needing the least warping for all other images to be aligned to it [41]. The aligned FA images were then averaged to create a 4D mean FA image, which was then subjected to thinning to create a mean FA skeleton representing the center of all white matter tracts, thereby removing partial-volume confounds. The FA skeleton was then thresholded at an FA value of 0.2 to limit the effects of poor alignment across subjects, to exclude areas with extremely low mean FA, and to ensure that grey matter and CSF voxels were excluded from the skeleton. The same non-linear transformation steps were applied to the MD, AD, and RD maps. For statistical analysis, the randomize tool in FSL was used to conduct a non-parametric permutation-based statistics using the Threshold-Free Cluster Enhancement (TFCE) method with 10,000 permutations to investigate group differences in FA, MD, AD and RD.Voxelwise maps were thresholded at p < 0.05 and corrected for multiple comparisons with family-wise error rate (FWE). The significant white matter clusters were identified with reference to the atlas tool JHU ICBM-DTI-81 white matter labels. To qualitatively assess structural difference between TBI patients and healthy controls groups, mean DTI indices for each subject were extracted by averaging FA, MD, AD, and RD for the significant white matter clusters using fslstats tool in FSL.
Dimensionality reduction and feature extraction
PCA was applied to FA, MD, AD, and RD skeletonized maps, in which each voxel represented a variable in the cross validation training dataset, and the same transformation was then applied to the test dataset. A voxel-wise approach was used to combine FA, MD, AD, and RD into one dataset named “ALL”, and PCA was applied to this ALL dataset in the same manner.
Support Vector Machines
A support vector machine (SVM) was used to perform ML analysis using the PCs of the DTI indices. For each DTI index, an SVM classification task was trained to distinguish TBI patients from healthy controls. PCs of the skeletonized maps of each DTI index and the ALL dataset were evaluated. Five-fold cross validation was used to yield an unbiased assessment of the classification method and prevent overestimation. A linear kernel SVM was chosen as a classifier and the hyperparameter (C) of the linear kernel SVM was fixed to 1.0. To evaluate the different classification tasks, the mean accuracy rate and its SD was calculated for the five-fold cross validation for each classification task (Fig.1). To validate the robustness of the classification results, 1000 times permutation tests were conducted to assess the statistical significance of the classification accuracy scores. To further estimate the performance of the different classification task, receiver operating characteristic curves (ROC) were plotted for each classifier and the areas under the curves (AUC) were obtained. The AUC quantifies the overall ability of the classifier to distinguish between the TBI and the healthy subjects. The ML analysis including feature extraction was conducted in Python using the Python library scikit-learn.