Computational protocol: Population-based studies of myocardial hypertrophy: high resolution cardiovascular magnetic resonance atlases improve statistical power

Similar protocols

Protocol publication

[…] To guide image segmentation 20 cardiac atlases were created to provide prior information about the inter-subject variability in cardiac anatomy []. A set of 3DLVSA images from 20 subjects had each voxel manually labeled as LV cavity, myocardium or right ventricle cavity on the end-diastolic image by two readers using freely available software (ITKsnap, National Library of Medicine’s Insight Segmentation and Registration Toolkit, http://www.itk.org) []. Subsequent image segmentation was automated but was initialised using six pre-defined landmarks on each target image. A multi-atlas PatchMatch algorithm [] was used to find correspondences between "patches" of neighbouring voxels (5 × 5 × 5 mm) within the atlases and target images (Figure  ). Each selected atlas patch was given a weighting according to its similarity and distance to the target patch. Labels from all the atlas patches were then combined to produce a final segmentation. Lastly, the mean shape of all atlases in the pool was co-registered to the segmentation to ensure each spatial coordinate in the 3D model was consistent between all subjects. Endocardial and epicardial mesh surfaces, with 8122 and 10696 points respectively, were reconstructed using the marching cubes algorithm [,]. Analysis was performed using numerical computing software (Matlab, Natick, MA) on a workstation (Xeon quad-core 2.4 GHz with 8GB of random access memory; Intel, Santa Clara, CA).All data were represented in a standard coordinate system and visualized on a 3D myocardial template created from the mean shape of the 20 atlases. Wall thickness was measured perpendicular to a midwall plane equidistant to the endocardial and epicardial surfaces. The volume of the voxels labeled as left ventricle cavity and myocardium were used to calculate LV end-diastolic volumes and mass which were then indexed to body surface area (LVEDVI and LVMI respectively). Myocardial density was assumed to be 1.05 g/mL [].To assess the accuracy of segmentation each of the 20 atlases was segmented using the other 19 and the results compared to its own manually labeled atlas. To enable an unbiased evaluation, the 2D LVSA images were compared to the corresponding manually-labeled atlas down-sampled to the equivalent spatial resolution. For the 3D and 2D LVSA data the distance between each epicardial and endocardial point on the segmentations and its corresponding point on the labeled atlas was measured. The accuracy of segmentation was also assessed by the proportion of concordant voxels in the target images and the corresponding manually labeled atlas. This agreement was expressed using the Dice similarity coefficient where 0 indicates no overlap and a value of 1 indicates perfect agreement []. [...] Data was analysed using R version 3.0.1 [] and SigmaPlot (Systat Software, San Jose, CA). Normally distributed data were reported as mean ± one standard deviation (SD) or otherwise as median and interquartile range (IQR). Comparison between methods was made using Bland-Altman plots []. CRs were compared using a two-sided paired t-test and reported with 95% confidence interval (CI) for the difference of the mean. Image quality scores were compared with the Wilcoxon signed-rank test. Test-retest reliability was assessed using an intraclass correlation coefficient (ICC) with a two-way random model for absolute agreement []. Voxelwise comparisons between the 2D and 3D techniques were made using the Wilcoxon signed rank test. The sample size required for automated segmentation of 2D and 3D techniques to detect a 1 mm difference in wall thickness at each point across the myocardium was calculated from the voxelwise interstudy variances (see Appendix) []. A P value <0.05 was considered significant and Bonferroni correction was made for multiple comparisons in all voxelwise tests. […]

Pipeline specifications