Viljami Sairanen MSc, Presenter: Nothing to Disclose

Linda Kuusela, Abstract Co-Author: Nothing to Disclose

Sampsa Vanhatalo, Abstract Co-Author: Nothing to Disclose

Sauli E. Savolainen PhD, Abstract Co-Author: Nothing to Disclose

Tract-Based spatial statistics (TBSS) is commonly used to search for differences or changes in white matter structures between groups. TBSS uses diffusion tensor (DT) data –derived Fractional Anistropy (FA) values that are usually calculated using linear least squares (LLS) tensor fit. Since FA values can be sensitive to the DT fitting algorithm used, this study was set out to evaluate the impact of LLS and non-linear least squares (NLLS) DT fitting routines in TBSS pipeline.

Diffusion weighted images, based on a healthy volunteer 3T MRI data, were used as a base to generate 40 synthetic patients. Rician noise was added to images to generate two groups with Signal-Noise-Ratio of 80 dB (SNR80) and 40 dB (SNR40). We created simulated left hemisphere brain lesions in both groups by decreasing voxel intensity in three brain regions by 10, 20 and 30% respectively. DT’s were calculated for each patient using both fitting routines to obtain FA maps. The effect of fitting routines on TBSS analysis was assessed by searching for significant (Monte Carlo P-value < 0.01) differences in individual voxels in FA-skeleton from both hemispheres. In an ideal case, the TBSS would return all simulated voxels in the respective hemisphere only. The number of false positive findings due to noise was determined from the right naïve hemisphere, and it was subtracted from the number of voxels found on the left modified hemisphere. TBSS output was then evaluated as the ratio of significant voxels from the total size of the FA-skeleton in the same area.

In the SNR40 group, the number of identified modified voxels in three brain regions was markedly higher after NLLS compared to LLS method: 68% vs 17% (NLLS vs LLS; 10% signal drop brain regions), 86% vs 62% (20% signal drop brain regions), and 85% vs. 31% (30% signal drop brain regions). The difference was also seen in the SNR80 group in all three regions: 38% vs 20% (NLLS vs LLS), 73% vs 65%, and 52% vs. 32%.

Our observations show that TBSS pipeline based on FA values derived from the NLLS method is able to identify a much higher proportion of true changes than the conventional LLS–based method.

The challenges with DT fitting in obtaining an anatomically reliable FA map presents a significant confounder in TBSS. Our work indicates that NLLS can improve the reliability of TBSS analysis.

Sairanen, V, Kuusela, L, Vanhatalo, S, Savolainen, S, Comparing Non-linear and Linear Least Square Diffusion Tensor Fitting Algorithms on the Tract-based Spatial Statistics Workflow.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14013448.html