Performance and Sensitivity of QSM Deep Learning algorithm to Background Field residuals using a Realistic Numerical Head Phantom

Motivation: Residual background fields in QSM reduce the accuracy of susceptibility reconstructions, making it crucial to evaluate the performance of different algorithms in managing these artifacts. Goal(s): Compares the robustness of traditional iterative and Deep Learning-based QSM algorithms in the presence of residual background fields to determine which provides the most reliable reconstructions. Approach: We applied three background field removal methods to simulated data and tested various QSM algorithms, measuring reconstruction error and variance across methods. Results: Iterative methods, especially WH-QSM, outperformed Deep Learning approaches, showing lower error and more consistent results across local field estimations. Across methods, LBV produced the best reconstructions. Impact: This study highlights the limitations of current Deep Learning-based QSM algorithms in handling residual background fields, suggesting a need for improved training strategies. It provides insights into which methods offer more reliable susceptibility maps, guiding future QSM development and applications.