NeXtQSM-A complete deep learning pipeline for data-consistent Quantitative Susceptibility Mapping trained with hybrid...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Deep learning (DL) shows promise for Quantitative Susceptibility Mapping (QSM).
• Current DL QSM methods often lack data consistency, require in vivo training data, or suffer from error propagation due to sequential processing.
• These limitations hinder the clinical translation and reliability of DL-based QSM.

Purpose of the Study:

• To develop a novel, data-consistent, end-to-end deep learning framework for Quantitative Susceptibility Mapping (QSM).
• To overcome the limitations of existing deep learning QSM approaches, including error propagation and reliance on in vivo data.
• To integrate QSM processing steps into a unified, trainable network.

Main Methods:

• Developed a hybrid training data generation method.
• Implemented a variational network combining the QSM model term and a learned regularizer for joint background field correction and dipole inversion.
• Enabled end-to-end training in a data-consistent manner.

Main Results:

• NeXtQSM demonstrates superior performance compared to previous deep learning QSM methods.
• The framework achieves data consistency throughout the QSM processing pipeline.
• Results are shown to be robust and computationally fast.

Conclusions:

• NeXtQSM presents a new deep learning pipeline for Quantitative Susceptibility Mapping.
• The integrated, data-consistent approach overcomes key limitations of prior DL methods.
• This framework offers a robust and efficient solution for QSM computation.