Chosing the QSM parameters

[Nicolas MINIER]

Hello everyone !

As a newbie in QSM, I would like to know how to correctly chose the different QSM parameters (I don't quite understand them fully)...

- Lambda
- Edge/percentage
- SMV
- max_iter vs. cg_max_iter
- tol_norm_ratio
...

Thank you all !

Nicolas MINIER.

[Zhe Liu]

Dear Nicolas,

Thank you for your interest in QSM! 

Below is explanation for terms in the MEDI code:

-Lambda

Lambda stands for the balance between data fidelity and regularization strength. It is still an open-ended question how to choose lambda for QSM. Possible ways might include L-curve and Morozov’s discrepancy principle (with knowledge of noise level). For human brain QSM, our empirical choice is lambda = 1000.

-Edge/percentage

This is used in generating edge mask for L1 regularization. It tells how many voxels can be regarded as "edge" in the ROI. For human brain QSM, our empirical choice is 0.9. For phantom (such as gadolinium-filled balloons), a lower value might be favored.

-SMV

This is an approximation for Laplacian operator, which further removes the background field in the presumed local field.

-max_iter and cg_max_iter

We use a Gauss-Newton solver with Conjugate Gradient. It's formulated as two layers of loops (top layer: Newton loop which linearizes the objective function; second layer: CG loop which solves the linearized problem). max_iter is the maximum iteration (typically 10) for Newton loop, and cg_max_iter (typically 100) for CG loop.

-tol_norm_ratio

This is another stopping criterion along with max_iter. It means that the Newton loop stops whenever the relative update of the solution ||dx||_2/||x||_2 is smaller than tol_norm_ratio (typically 0.1)

Also I strongly recommend these works, which extensively study how different parameters might affect QSM:

Liu, T., et al. (2012). "Accuracy of the morphology enabled dipole inversion (MEDI) algorithm for quantitative susceptibility mapping in MRI." Medical Imaging, IEEE Transactions on(99): 1-1.

Liu, J., et al. (2012). "Morphology enabled dipole inversion for quantitative susceptibility mapping using structural consistency between the magnitude image and the susceptibility map." Neuroimage 59(3): 2560-2568.

Best,

Zhe Liu

--
You received this message because you are subscribed to the Google Groups "medi-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to medi-users+...@googlegroups.com.
To post to this group, send email to medi-...@googlegroups.com.
Visit this group at https://groups.google.com/group/medi-users.
To view this discussion on the web visit https://groups.google.com/d/msgid/medi-users/5e5c0eb7-4f2d-44c5-be11-7ad760dedec4%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

刘喆
Zhe Liu

PhD student
Biomedical Engineering Department
Cornell University
Ithaca, NY, United States, 14850

Phone: (607)-280-8639

[Nicolas MINIER]

Well... thanks a lot !