Synopsis

The current regularization in morphology-enabled dipole inversion (MEDI) does not take into account orientation information in morphology between QSM and its corresponding magnitude image. In this abstract, we consider such orientation information to enhance structural coherence between the two images. In doing so, we achieve better image quality as well as higher RMSE (root mean square error) and HFEN (high frequency error norm) with respect to COSMOS and $$$\chi_{33}$$$.

Purpose

Theory

1) Regularized Field-to-Susceptibility Inversion with Isotropic Weighting

In a spatially continuous setting, the problem of regularized dipole inversion is written as follows:

$$ \text{Find } \hat \chi \text{ such that } \hat \chi \in \operatorname{Argmin}\left\{\chi \in BV(\Omega; \mathbb{R}): R(\chi)+ \frac{\lambda}{2}\int_\Omega |w(d*\chi-b)|^2\right\},$$

where the susceptibility distribution $$$\chi:(\Omega \subset \mathbb{R}^3) \to \mathbb{R}$$$ is estimated from the measured local field $$$b:\Omega \to \mathbb{R}$$$. The space of bounded variation, $$$BV(\Omega;\mathbb{R})$$$, allows the existence of jumps modeling sharp edges. The map $$$w:L^1 \to L^1$$$ compensates for the non-uniform phase noise (SNR weighting). The regularization term $$$R(\chi)$$$ is designed in such a way that it effectively penalizes streaking artifacts originated from the dipole-incompatible part of the field (1). Incorporating morphology information extracted from the corresponding magnitude image, the regularization has been typically designed in the following fashion (2):

$$R_{iso}(\chi) := \int_\Omega ||M_G\nabla \chi||_1 = \int_\Omega \left\| \left( \begin{array}{ccc} 1-\delta_{11} & 0 & 0 \\ 0 & 1-\delta_{22} & 0 \\ 0 & 0 & 1-\delta_{33} \\ \end{array} \right) \left( \begin{array}{c} \partial_x \chi \\ \partial_y \chi \\ \partial_z \chi \\ \end{array} \right) \right\|_1,$$

where $$$\delta_{ii}$$$ is an edge indicator function: $$$\delta_{11}(r)=1$$$ if there exists an edge along the x direction; otherwise 0. Likewise, $$$\delta_{22}(r)$$$ and $$$\delta_{33}(r)$$$ are edge indicators along the y and z directions. Therefore, for an edge indicator function determined by the magnitude image, $$$R_{iso}$$$ penalizes only the regions where tissue structure (edges) is not expected.

2) Regularization via Anisotropic Weighting

Taking into account orientation information, we expect further improvements in QSM. We modify the regularization term such that it penalizes the perpendicular components of the gradient of susceptibility distribution with respect to the direction of the edges of magnitude image as follows:

$$R_{aniso}(\chi) := \int_\Omega ||P_{\perp\xi}\nabla \chi||_1 = \int_\Omega \left\|\left(I-\frac{\xi\xi^\top}{\xi^\top\xi} \right)\nabla\chi\right\|_1 = \int_\Omega \left\| \left( \begin{array}{ccc} 1-\xi_1^2 & -\xi_1\xi_2 & -\xi_1\xi_3 \\ -\xi_2\xi_1 & 1-\xi_2^2 & -\xi_2\xi_3 \\ -\xi_3\xi_1 & \xi_3\xi_2 & 1-\xi_3^2 \\ \end{array} \right) \left( \begin{array}{c} \partial_x \chi \\ \partial_y \chi \\ \partial_z \chi \\ \end{array} \right) \right\|_1,$$

where $$$\xi$$$ is the gradient of magnitude image, i.e., $$$\xi=\nabla\text{Mag}$$$. In other words, we replace the isotropic weighting $$$M_G$$$ with the orthogonal projector $$$P_{\perp\xi}$$$. We illustrate this in Fig. 1.

Methods

We used two in vivo brain MRI dataset that include COSMOS and as reference:

1. QSM 2016 Reconstruction Challenge

- http://www.neuroimaging.at/qsm2016/pages/qsm-challenge.php

- Both $$$\chi_{33}$$$ and COSMOS are available.

2. Cornell MRI research lab

- http://weill.cornell.edu/mri/pages/qsmreview.html

- COSMOS is available.

We refer the reader to the webpages for more details on data acquisition and phase processing. Computed susceptibility maps were validated against the gold standard ($$$\chi_{33}$$$ or COSMOS) using global metrics: RMSE (root mean squared error), HFEN (high frequency error norm), and SSIM (structure similarity index).

Results

Discussion

As can be seen in Figs. 3 and 4, anisotropic weighting achieved improvements in image quality as well as two quantitative measures – RMSE (root mean square error) and HFEN (high frequency error norm). Visually, this is observed from both axial and sagittal planes in Fig. 3. Notice that black/white dots in isotropic weighting were substantially removed in anisotropic weighting with respect to COSMOS. Although the difference between the two images in FIG. 4 is subtle, similar improvements were observed in anisotropic weighting.

The performance of anisotropic weighting largely depends on a predefined edge indicator function. For the experimental results reported herein, we used the same edge indicator as the isotropic weighting and assigned values in the range of [-1, 1] rather than 1 on the location of edges based on the definition of $$$R_{aniso}(\chi)$$$ described in the Theory section. Therefore, further improvements are expected if the location of edges is optimally estimated.

Anisotropic weighting, on the other hand, has lower values of SSIM (structure similarity index) than isotropic weighting although exhibiting less amount of spurious noise as opposed to isotropic weighting. This will be future research directions.

Conclusion

Acknowledgements

References

(1) Choi J. K, Zhou L, Seo J. K, Wang Y. Dipole-Incompatible Field Data Causes Streaking Artifacts in the Field-to-Susceptibility Inverse Problem and Implications for QSM Algorithm Design. QSM Workshop 2016.

(2) Liu J, Liu T, de Rochefort L, Ledoux J, Khalidov I, Chen W, Tsiouris AJ, Wisnieff C, Spincemaille P, Prince MR, Wang Y. Morphology enabled dipole inversion for quantitative susceptibility mapping using structural consistency between the magnitude image and the susceptibility map. Neuroimage 2012;59(3):2560-2568