Learned Proximal Networks for Quantitative Susceptibility Mapping.

Area of Science:

• Medical Imaging
• Computational Neuroscience
• Biophysics

Background:

• Quantitative Susceptibility Mapping (QSM) reconstructs magnetic susceptibility from MR phase data.
• Traditional QSM methods face challenges like artifacts and high acquisition costs.
• Deep learning, particularly CNNs, shows promise for medical image reconstruction but often lacks interpretability.

Purpose of the Study:

• To develop an interpretable deep learning framework for QSM.
• To address the ill-posed dipole inversion problem in QSM.
• To create a flexible QSM method adaptable to varying numbers of phase measurements.

Main Methods:

• Introduced a Learned Proximal Convolutional Neural Network (LP-CNN) using iterative proximal gradient descent.
• The LP-CNN learns an implicit regularizer, decoupling the forward operator from data-driven parameters.
• The framework handles an arbitrary number of phase input measurements without re-training.

Main Results:

• LP-CNN achieved state-of-the-art reconstruction results compared to traditional and other deep learning QSM methods.
• The method demonstrated improved artifact reduction and smoothing compared to conventional techniques.
• The LP-CNN framework offers greater flexibility in QSM reconstruction.

Conclusions:

• The LP-CNN provides a powerful and flexible deep learning solution for QSM.
• This approach combines the benefits of data-driven learning and interpretable iterative reconstruction.
• The LP-CNN advances the field of medical image reconstruction for quantitative susceptibility mapping.