ProxiMO: Proximal Multi-operator Networks for Quantitative Susceptibility Mapping.

Area of Science:

• Medical Imaging
• Computational Neuroscience
• Machine Learning

Background:

• Quantitative Susceptibility Mapping (QSM) reconstructs magnetic susceptibility from MRI phase data.
• Accurate QSM typically requires multi-orientation data acquisition, which is time-consuming.
• Existing deep learning methods for QSM often rely on supervised training with multi-orientation data.

Purpose of the Study:

• To develop an unsupervised deep learning approach for QSM reconstruction using single-orientation MRI data.
• To enable efficient and accurate QSM without the need for multiple orientations or ground truth data.
• To introduce a semi-supervised variant for further performance enhancement.

Main Methods:

• ProxiMO (Proximal Multi-Operator) combines Learned Proximal Convolutional Neural Networks (LP-CNN) with multi-operator imaging (MOI).
• The method facilitates unsupervised training of LP-CNNs for QSM on single-phase data.
• A semi-supervised variant was developed to improve reconstruction performance.

Main Results:

• ProxiMO successfully trains on single-orientation measurement data without ground truth reconstructions.
• The unsupervised approach demonstrates significant advantages for QSM reconstruction.
• Experiments on multicenter datasets show the superiority of the proposed ProxiMO method.

Conclusions:

• Unsupervised learning offers an efficient alternative for QSM reconstruction.
• ProxiMO provides a robust and superior method for QSM, particularly with single-orientation data.
• The developed approach advances the field of medical image analysis and computational imaging.