MoDL-QSM: Model-based deep learning for quantitative susceptibility mapping.

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Quantitative susceptibility mapping (QSM) quantifies tissue susceptibility for brain disease diagnosis.
• The accuracy of QSM is limited by an ill-posed inverse problem and streaking artifacts.
• Existing deep learning methods for QSM face a mismatch between observed and theoretical phases.

Purpose of the Study:

• To propose a novel model-based deep learning architecture, MoDL-QSM, for improved QSM accuracy.
• To address the phase mismatch issue in deep learning-based QSM.
• To leverage the physical model of susceptibility tensor imaging (STI) for enhanced QSM.

Main Methods:

• Developed MoDL-QSM, a deep learning architecture integrated into the STI physical model.
• Embedded convolutional neural networks within the physical model to learn regularization terms.
• Utilized χ33 and phase induced by χ13 and χ23 terms as labels for network training.

Main Results:

• MoDL-QSM demonstrated superior performance compared to recent deep learning QSM methods.
• The proposed model effectively accounts for the relationship between STI-derived phase contrast and acquired single-orientation phase.
• Reduced streaking artifacts and improved accuracy in quantitative susceptibility mapping.

Conclusions:

• MoDL-QSM offers a promising approach for accurate quantitative susceptibility mapping.
• The model-based deep learning strategy effectively resolves phase mismatches in QSM.
• MoDL-QSM shows significant potential for future clinical applications in brain disease quantification.