Improving Data Consistency and Generalizability in Deep Learning-Based QSM Reconstruction

Motivation: QSM reconstruction is a highly ill-posed inverse problem. Machine learning promises to address this problem by incorporating data-driven priors. But existing learning-based QSM reconstructions suffer from data inconsistency and limited generalizability. Goal(s): To improve data consistency and generalizability of learning-based QSM reconstruction. Approach: Data consistency was improved by enforcing measured tissue phase in k-space, leveraging the characteristics of the dipole kernel. Generalizability was improved by augmenting a small COSMOS training dataset with extensive phase and spatial variation from a large single-orientation QSM dataset. Results: The proposed method has been validated on both simulation and in vivo data, producing data-consistent and generalizable QSM maps. Impact: With improved data consistency and generalizability, the proposed method could significantly enhance the reliability of AI-powered QSM reconstruction, making it a practically useful tool to support a wide range of QSM studies in scientific and clinical applications.