Quantitative susceptibility mapping in magnetically inhomogeneous tissues.

Area of Science:

• Medical Imaging
• Neuroimaging
• Computational Biology

Background:

• Conventional quantitative susceptibility mapping (QSM) methods utilize simplified physical models, leading to artifacts and inaccuracies in biological tissues due to assumptions of isotropic and homogeneous properties.
• Accurate susceptibility mapping is crucial for understanding various neurological conditions.

Purpose of the Study:

• To develop and evaluate DEEPOLE, a deep learning-based QSM method that incorporates macroscopically nondipolar Larmor frequency shifts.
• To enhance the quality and accuracy of susceptibility maps compared to traditional QSM algorithms.

Main Methods:

• DEEPOLE integrates the QUASAR model within a deep convolutional neural network to capture neglected frequency contributions.
• The method was trained on synthesized data and validated against established QSM algorithms (deep learning QSM, QUASAR, MEDI, FANSI, SDI) using digital brain models and in vivo human brain data.

Main Results:

• DEEPOLE demonstrated superior performance over conventional QSM methods in digital brain models, yielding fewer artifacts and higher quantitative accuracy, particularly with microstructure effects.
• In vivo, DEEPOLE produced more anatomically consistent susceptibility maps, reduced artifacts like inhomogeneities and streaking, and improved susceptibility estimates in deep gray and white matter.

Conclusions:

• Incorporating macroscopically nondipolar Larmor frequency shifts via DEEPOLE significantly improves susceptibility map quality and accuracy.
• This advancement enhances the reliability of susceptibility measurements, especially for neurodegenerative and demyelinating conditions where such contributions are substantial.