Robust quantitative susceptibility mapping via approximate message passing with parameter estimation.

Area of Science:

• Medical Imaging
• Computational Neuroscience
• Biophysics

Background:

• Quantitative susceptibility mapping (QSM) is crucial for neuroimaging but faces challenges in parameter selection for dipole inversion due to the lack of clinical ground-truth.
• Accurate parameter determination is vital for reliable susceptibility map recovery in QSM.

Purpose of the Study:

• To develop a probabilistic Bayesian approach for QSM with integrated parameter estimation.
• To incorporate a nonlinear dipole inversion formulation for robust susceptibility map recovery.

Main Methods:

• A Bayesian framework models image wavelet coefficients using a Laplace distribution and measurement noise with a Gaussian-mixture model.
• Approximate message passing (AMP) with built-in parameter estimation (AMP-PE) is employed for joint recovery of susceptibility maps and distribution parameters.
• The proposed AMP-PE method is compared against L1-QSM, FANSI, and MEDI on simulated and in vivo datasets.

Main Results:

• AMP-PE demonstrated superior performance on simulated data, achieving the lowest NRMSE and highest SSIM compared to other methods.
• On in vivo datasets, AMP-PE robustly recovered susceptibility maps using automatically estimated parameters.
• State-of-the-art methods like L1-QSM, FANSI, and MEDI often required manual parameter tuning for optimal results.

Conclusions:

• AMP-PE offers automatic and adaptive parameter estimation for QSM, eliminating the need for subjective manual tuning.
• This approach enhances the reliability and clinical applicability of QSM by providing objective parameter selection.