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

Magnetic source separation (MSS) in quantitative susceptibility mapping (QSM) provides a powerful tool to disentangle paramagnetic and diamagnetic contributions, enabling more accurate quantification of brain iron and myelin alterations. However, existing MSS approaches typically depend on approximations derived from reversible transverse relaxation (R₂^'=R₂^∗-R₂) or extrapolate from a limited number of brain regions to perform whole-brain separation. Furthermore, current deep learning-based methods often require extensive and high-quality labels, which are difficult to obtain. To address these limitations, we propose an unsupervised MSS framework guided by prior information and constrained by physics-informed loss functions to improve separation fidelity. The proposed model directly processes whole-brain QSM and R₂^∗ data, infers intermediate parameters, and reconstructs the spatial distributions of paramagnetic and diamagnetic sources via biophysical modeling. Experimental results show that the method achieves high structural similarity (SSIM = 0.9945 for paramagnetic and 0.9942 for diamagnetic components) and a low normalized mean square error (0.11) relative to the original QSM, demonstrating robust and consistent source decomposition performance. Code is available at https://github.com/TyrionJ/PI-uMSS.