Quantitative Susceptibility Mapping of Human Brain at 3T: A Multisite Reproducibility Study

Discussion

Quantitative susceptibility mapping provides an important opportunity for measuring pathologic changes in iron content to improve the diagnosis of brain diseases.²⁵ Before one fully uses the potential of QSM, it should be understood that settings of the parameters in the data-preprocessing procedures and the susceptibility-mapping algorithms could both lead to variations in the quantitative results. Therefore, it is essential to establish the precision level of QSM in terms of both within-site and cross-site reproducibility before wide application of this technique to clinical studies.

The results from our study show that while within-site and cross-site yield similar reproducibility levels in general, the caudate nucleus, putamen, and globus pallidus show significantly higher precision in susceptibility values (average variations on the order of 0.005 ppm) than the red nucleus and substantia nigra (average variations on the order of 0.01 ppm) when computed by using iterative QSM. This regional dependency of QSM precision is anticipated to arise from the influences of air-tissue interfaces. In fact, the shimming condition is expected to alter in different scans particularly near the air-tissue interfaces, which in turn, could cause inconsistencies in background field removal via the PDF method. Moreover, the PDF approach, though relatively robust compared with the high-pass-filtering approach, is known to exhibit difficulties in the object boundaries, which also correspond to regions adjacent to the air cavities.¹⁸ Our experimental observation of the somewhat larger cross-site discrepancy in the frontal region, as shown in Fig 4, is also consistent with the inference that boundaries of large susceptibility changes remain the major challenging regions for QSM. The red nucleus and substantia nigra are 2 examples of tissues at the skull base level, hence exhibiting relatively lower precision than other regions investigated in this study.

One recent study reported aging-related increases in susceptibility values in healthy human subjects, with stronger statistical significance found in the putamen, globus pallidus, and red nucleus than in the substantia nigra, and insignificant changes in the caudate nucleus,⁷ all in very good agreement with findings from our study. In the study by Bilgic et al,⁷ the analysis was performed on 2 groups of subjects whose age ranges were far separated (21–29 years for the young population versus 64–86 years for the elderly population, respectively), as opposed to our preliminary investigation in which the subject age range was relatively narrow (23–37 years). Nevertheless, the ranges of the numeric values of the magnetic susceptibility found for these brain regions in our data are consistent with those in previous investigations.^7,26 More important, with the levels of precision achieved in our study, future examinations of the aging-related susceptibility changes on the basis of regression analysis as opposed to categoric analysis should be highly feasible.

In the phantom experiment, the slope between measured and expected susceptibility values in Fig 1 is 0.91, which reveals 9% of the inaccuracy in QSM. The inaccuracy of QSM in this study was found to be at a level similar to that found in another study (Fig 2 in de Rochefort et al, slope = 0.93).¹ Although the slope is found to deviate from the ideal value of unity, the R² value of 0.9999 between measured and expected susceptibility values demonstrates the high linearity of susceptibility mapping within the range of 0–2.04 ppm. Thus, the precision and reproducibility are not affected, even if the nonunity slope currently remains an unsolved issue if absolute accuracy is required.

Comparison of iterative and deterministic QSM shows that both algorithms are capable of solving the inverse problem of susceptibility mapping with good consistency, at least for region-of-interest-based analysis. However, deterministic QSM may be more susceptible to streaking artifacts at regions surrounding strong susceptibility changes, as shown in the phantom images. This susceptibility could also explain the somewhat inferior reproducibility of deterministic QSM to iterative QSM in the caudate nucleus. An examination of equations 1 and 2 suggests that the inclusion of the edge mask from the magnitude image in the regularization term is the major cause for the difference in the performance for the 2 QSM algorithms. Most interesting, on human brain susceptibility maps obtained in vivo by using deterministic QSM, streaking artifacts are not as visually prominent as on the phantom images, where the possible presence of the streaking artifacts may have been masked by the texture of brain tissues. While for these human data, the comparative superiority of the 2 QSM algorithms cannot be assessed because no ground truth could be obtained for validation, causing difficulty in determining which of the 2 QSM methods gives a lower level of streaking artifacts in Fig 3, the phantom results are certainly in favor of iterative QSM. Consequently, although the deterministic QSM is found to be at least 700 times faster than iterative QSM and thus opening an opportunity for on-line reconstruction of susceptibility maps at the MR imaging console, for detailed quantitative analysis such as aging-related changes, the iterative QSM approach is still recommended.

In data preprocessing, the purpose of background field removal is to extract local magnetic field variation generated by brain tissues, which is subsequently used for QSM. The high-pass filtering method is not preferred for quantitative susceptibility mapping, even though it is one of the most frequently used methods for background field removal in MR imaging (eg, in susceptibility-weighted imaging). One of the major limitations of applying high-pass filtering in QSM is that part of the local field information could be mistakenly filtered out, because the distributions of spatial frequency in local field and in background field cannot be cleanly separated by a single high-pass-filtering kernel size. Therefore, new methods, including the PDF¹⁸ and the sophisticated harmonic artifact reduction for phase data (SHARP)²² methods, have been proposed for background field removal in QSM. The idea of both the PDF and the SHARP methods is to decompose local field and background field by using their physical properties. While the PDF method uses a projection theorem, the SHARP method applies the Laplace equation to separate background field. In previous studies, both the PDF and the SHARP methods have been demonstrated to be superior to the high-pass-filtering method for background field removal in QSM.^18,22 From numeric simulation, we have also confirmed that local field obtained by the PDF method preserves subtle field changes from brain tissues better than that obtained by the high-pass-filtering method (data not shown). We, therefore, think that both the PDF and the SHARP methods are adequate for background field removal in QSM. The comparison between the PDF and the SHARP methods, however, is beyond the scope of this study.

This study is limited by the absence of flow compensation to reduce flow-related phase accumulations in vessels and the ventricles. Because our study used commercial 3D multiecho gradient-echo sequences on the tested scanners, full flow compensation for all echoes was unavailable at the time of this study. This scenario may have affected the reference susceptibility values calculated for the flowing CSF in the lateral ventricles. Nonetheless, this minor weakness influences only the absolute susceptibility values reported for the 5 ROIs and is relatively unimportant to our investigation focused on within-site and cross-site precision assessment. Some investigators used the splenium as a reference⁷ instead of using CSF.²⁶ Yet in our experience, the regional variations of susceptibility in the splenium are no smaller than those found for the lateral ventricles. This potential limitation of uncertain susceptibility values for the CSF could be alleviated with pulse sequence modifications to include flow compensation for all echoes.^17,27 As a last note, the statistical sensitivity of cross-site imprecisions may be somewhat limited by the derivation of SDs by using only 3 numbers. However, this limitation is because we could only access the three 3T scanners. Moreover, even with this statistical limitation, results from this study could still serve as a reference for the imprecision level to be expected in future multisite QSM studies.