Analysis of diffusion tensor measurements of the human cervical spinal cord based on semiautomatic segmentation of the white and gray matter

BackgroundPurposeSegmentation of the gray and white matter (GM, WM) of the human spinal cordin MRI images as well as the analysis of spinal corddiffusivity are challenging. When appropriately segmented, diffusion tensor imaging (DTI) of the spinal cordmight be beneficial in the diagnosis and prognosis of several diseases. To evaluate the applicability of a semiautomatic algorithm provided by ITK-SNAP in classification mode (CLASS) for segmenting cervical spinal cordGM, WM in MRI images and analyzing DTI parameters. Study TypeSubjectsProspective. Twenty healthy volunteers. SequencesAssessment1.5T, turbo spin echo, fast field echo, single-shot echo planar imaging. Three raters segmented the tissues by manual, CLASS, and atlas-based methods ( Spinal CordToolbox, SCT) on T-2-weighted and DTI images. Masks were quantified by similarity and distance metrics, then analyzed for repeatability and mutual comparability. Masks created over T-2 images were registered into diffusion space and fractional anisotropy(FA) values were statistically evaluated for dependency on method, rater, or tissue. Statistical TestsResultst-test, analysis of variance (ANOVA), coefficient of variation, Dice coefficient, Hausdorff distance. CLASS segmentation reached better agreement with manual segmentation than did SCT (P<0.001). Intra- and interobserver repeatability of SCT was better for GM and WM (both P<0.001) but comparable with CLASS in entire spinal cord segmentation(P=0.17 and P=0.07, respectively). While FA values of whole spinal cordwere not influenced by choice of segmentation method, both semiautomatic methods yielded lower FA values (P<0.005) for GM than did the manual technique (mean differences 0.02 and 0.04 for SCT and CLASS, respectively). Repeatability of FA values for all methods was sufficient, with mostly less than 2% variance.