Evaluation of QSM Reconstruction Pipelines for Multi-echo Gradient-echo Acquisitions

1979 

Abstract Submission 

Jian Lin¹, Gawon Lee², Sehong Oh¹, Katherine Koenig¹, Ken Sakaie¹, Mark Lowe¹

¹The Cleveland Clinic, Cleveland, OH, ²Hankuk University of Foreign Studies, Yongin, Korea, Republic of

Jian Lin  
The Cleveland Clinic
Cleveland, OH

Gawon Lee  
Hankuk University of Foreign Studies
Yongin, Korea, Republic of

Sehong Oh  
The Cleveland Clinic
Cleveland, OH

Katherine Koenig  
The Cleveland Clinic
Cleveland, OH

Ken Sakaie  
The Cleveland Clinic
Cleveland, OH

Mark Lowe  
The Cleveland Clinic
Cleveland, OH

The NIH/NIA supports a network of 33 Alzheimer's Disease Research Centers (ADRC) to promote translation of research to improved patient care. The neuroimaging core of the Cleveland ADRC features advanced imaging approaches to evaluate their use in guiding patient care. Of these approaches, quantitative susceptibility mapping (QSM)¹ was included because of its potential to predict cognitive decline². However, susceptibility values can be highly dependent on choices in the processing pipeline, and evaluation of pipeline is often confined to use in idealized datasets. In this contribution, we compare susceptibility values based on four different strategies for combining information from different echoes from MGE acquisitions on data acquired from the CADRC.

70 subjects (Table 1) from the CADRC were scanned on a 3T MRI (Prisma, Siemens Healthineers, Erlangen, Germany). Imaging included 3D multi-echo gradient-echo (MGE) (FOV = 192mm x 256mm x 176mm, 1 mm isotropic voxels, TE = 16/22 ms, TR = 27 ms, FA=20^o, BW = 260 Hz/px) and 3D T1-weighted image using MPRAGE sequence (FOV = 208 mm x 240 mm x 256 mm, 1 mm isotropic voxels, TE/TR = 29.8/2300 ms). QSM maps were generated using STI SUITE 3.0³. 4 reconstruction pipelines (P1-P4) were developed, each reflecting a different strategy for combining data from multiple echoes (Figure 1). MPRAGE images were parcellated using freesurfer⁴ and aligned to QSM space using align_epi_anat.py.{Saad, 2009 #1832} For each region, summary statistics (mean, standard deviation (SD), median) of susceptibility were calculated. One-way ANOVA among pipelines for each parcel and each statistic was performed with a Bonferroni correction for multiple comparisons, followed by post hoc comparisons with the Tukey-Kramer test. A p value < 0.05 was considered statistically significant.

Examples of susceptibility maps from each pipeline are shown in Figure 2. Three test statistics were compared in each of all 181 regions generated by freeSurfer, leading to a Bonferroni correction factor of 3x181=543. Table 2 summarizes the results of the ANOVA analysis. As suggested by figure 1, P1 was different from the other pipelines in at least one-third of the regions, regardless of the summary statistic examined. In contrast, P2 and P4 were virtually identical. P3 was similar to P2 and P4 when using mean or median as the summary statistic, but many regions showed differences in SD, suggesting large differences of the uniformity of susceptibility across each region.

The impact of the pipeline on QSM was recently examined by Biondetti et al. in great detail in 10 healthy volunteers⁵ and found that combining phase maps prior to calculating susceptibility led to superior results than calculating a separate susceptibility map for each phase and then averaging. In contrast, our study showed that the two approaches, corresponding to P2 and P4, yielded nearly identical results. In addition to differences in the study population, differences in approach likely explain the differences among results. Here, phases were averaged, which is a simple way of implementing the weighted average implemented by Biondetti et al.⁵, and acquisitions differed, with a notable difference being in the number of echoes used.
The procedure for combining echoes can affect QSM maps, but not always. Further work will be required to determine if effects from different analysis pipelines are explained by algorithmic choices or by differences among study populations.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ²

Other Methods ¹

Data analysis

Degenerative Disease

Other - Quantitative Susceptibility Mapping, Quantitative imaging,