Recovering SWI-filtered phase data using deep learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Quantitative susceptibility mapping (QSM) is crucial for analyzing MRI data.
• Standard QSM processing relies on filtered phase data, often obtained through complex pipelines.
• Recovering filtered phase directly could streamline QSM computation.

Purpose of the Study:

• To develop a deep neural network (DNN) capable of recovering filtered phase from clinical MR phase images.
• To enable accurate computation of quantitative susceptibility maps (QSMs) from the recovered phase data.

Main Methods:

• Trained 18 deep learning networks on 132 3D MRI volumes to recover filtered phase data.
• Implemented two experiments: inverting filtering operations and fine-tuning networks to recover unfiltered local fields.
• Computed susceptibility maps from recovered fields and compared them to gold-standard reconstructions.

Main Results:

• Susceptibility maps computed from recovered phase achieved comparable accuracy (NRMSE 0.725 ± 0.095) to standard QSM processing (NRMSE 0.732 ± 0.095).
• A network trained on all 13 filtering methods generalized well to unseen filters, maintaining reconstruction accuracy.
• Deep learning effectively recovered filtered phase, demonstrating its utility in QSM workflows.

Conclusions:

• Deep learning provides a feasible method for recovering filtered phase from MR images.
• QSM can be accurately computed from the recovered phase data, offering comparable results to traditional methods.
• This approach enhances the efficiency and accuracy of QSM acquisition and analysis.