New LAMDI method sharpens brain MRI without manual parameter tuning

Quantitative Susceptibility Mapping (QSM) reveals tissue magnetic properties from MRI phase data, but existing dipole inversion methods often require manual regularization tuning and can produce blocky, over-smoothed results. A team led by Shuai Huang from Emory University introduces LAMDI (Laplace-Mixture Dipole Inversion) to solve this. The method extends the approximate message passing with parameter estimation (AMP-PE) framework by replacing a single Laplace prior with a two-component Laplace mixture prior. This better captures the heavy-tailed gradient distribution of brain susceptibility maps, automatically adjusting regularization without user intervention.

Tested on a public multi-orientation QSM dataset, LAMDI achieved normalized root mean square error (NRMSE) and structural similarity index (SSIM) comparable to the single-Laplace AMP-PE-L1, while substantially reducing high-frequency error norm (HFEN) — indicating superior preservation of fine anatomical structures. Compared to standard methods like FANSI and MEDI, LAMDI remained competitive under reference-based tuning but did not require any reference maps or manual parameter selection. The automatic parameter estimation makes LAMDI a practical, accessible tool for clinical MRI workflows, potentially improving diagnosis of conditions such as neurodegeneration where subtle tissue susceptibility changes matter.