PI-uMSS: Prior information-based unsupervised magnetic source separation in quantitative susceptibility mapping.

Area of Science:

• Neuroimaging
• Biophysics
• Medical Physics

Background:

• Magnetic source separation (MSS) in quantitative susceptibility mapping (QSM) is crucial for quantifying brain iron and myelin.
• Existing MSS methods rely on approximations or limited regional data and struggle with whole-brain analysis.
• Current deep learning approaches for MSS demand extensive, high-quality labeled datasets, which are challenging to acquire.

Purpose of the Study:

• To develop an unsupervised MSS framework for whole-brain quantitative susceptibility mapping.
• To improve the fidelity of separating paramagnetic and diamagnetic contributions in the brain.
• To overcome limitations of existing MSS methods, including reliance on approximations and data scarcity.

Main Methods:

• Proposed an unsupervised MSS framework utilizing prior information and physics-informed loss functions.
• Directly processed whole-brain QSM and R₂^∗ data.
• Inferred intermediate biophysical parameters to reconstruct spatial distributions of paramagnetic and diamagnetic sources.

Main Results:

• Achieved high structural similarity (SSIM) for both paramagnetic (0.9945) and diamagnetic (0.9942) components.
• Demonstrated a low normalized mean square error (0.11) compared to the original QSM.
• Showcased robust and consistent source decomposition performance across the whole brain.

Conclusions:

• The proposed unsupervised MSS framework effectively separates paramagnetic and diamagnetic sources in QSM.
• The method offers improved accuracy and robustness for whole-brain analysis without requiring extensive labels.
• This advancement has significant implications for quantifying brain iron and myelin alterations in neuroimaging research.