AbstractPurposeConventional quantitative susceptibility mapping (QSM) methods rely on simplified physical models that assume isotropic and homogeneous tissue properties, leading to artifacts and inaccuracies in biological tissues. This study aims to develop and evaluate DEEPOLE, a deep learning–based method that incorporates macroscopically nondipolar Larmor frequency shifts into QSM to enhance the quality and accuracy of susceptibility maps.MethodsDEEPOLE integrates the QUASAR model into a deep convolutional neural network to account for frequency contributions neglected by conventional QSM. We trained DEEPOLE using synthesized data reflecting realistic power spectrum distributions. Its performance was evaluated against traditional QSM algorithms—including deep learning QSM, QUASAR (quantitative susceptibility and residual mapping), morphology‐enabled dipole inversion (MEDI), fast nonlinear susceptibility inversion (FANSI), and superfast dipole inversion (SDI)—using realistic digital brain models with and without microstructure effects, as well as in vivo human brain data. Quantitative assessments focused on susceptibility estimation accuracy, artifact reduction, and anatomical consistency.ResultsIn digital brain models, DEEPOLE outperformed conventional QSM methods by producing susceptibility maps with fewer artifacts and greater quantitative accuracy, especially in regions affected by microstructure effects. In vivo, DEEPOLE generated more anatomically consistent susceptibility maps and mitigated artifacts such as inhomogeneities and streaking, providing improved susceptibility estimates in deep gray matter and white matter.ConclusionIncorporating macroscopically nondipolar Larmor frequency shifts into QSM through DEEPOLE improves the quality and accuracy of susceptibility maps. This methodological advancement enhances the reliability of susceptibility measurements, particularly in studies of neurodegenerative and demyelinating conditions where macroscopically nondipolar contributions are substantial.