A latent code based multi-variable modulation network for susceptibility mapping.

Area of Science:

• Medical Imaging
• Computational Neuroscience
• Biophysics

Background:

• Quantitative susceptibility mapping (QSM) provides crucial tissue susceptibility information for clinical use.
• Current deep learning methods for QSM face limitations in data distribution diversity, impacting clinical generalization.
• Solving the ill-posed inverse problem in QSM is essential for accurate results.

Purpose of the Study:

• To develop a novel deep learning network, LCMnet, for improved QSM reconstruction.
• To enhance model generalization and stability in clinical QSM applications.
• To address the limitations of existing deep learning approaches in handling diverse data distributions.

Main Methods:

• Proposed a Latent Code based Multi-Variable modulation network (LCMnet) for QSM reconstruction.
• Incorporated a modulation module using field map, magnitude image, and initial susceptibility.
• Utilized an encoder-decoder framework to learn latent codes and a cross-fusion block for multi-level feature integration.

Main Results:

• LCMnet demonstrated outstanding performance in accurate susceptibility measurement.
• The method achieved excellent generalization across various datasets, including in vivo human brain, challenge, clinical, and synthetic data.
• The proposed network enhanced stability and accelerated training convergence.

Conclusions:

• LCMnet offers a robust and generalizable solution for quantitative susceptibility mapping.
• The novel modulation approach effectively integrates multiple data variables for superior QSM reconstruction.
• This technique shows significant potential for advancing clinical applications of QSM.