xQSM: quantitative susceptibility mapping with octave convolutional and noise-regularized neural networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neuroscience

Background:

• Quantitative susceptibility mapping (QSM) is a powerful MRI technique for visualizing magnetic susceptibility. The reconstruction process for QSM is inherently challenging due to the ill-posed nature of dipole inversion.
• Existing deep learning and conventional methods for QSM reconstruction face limitations in accuracy and speed.

Purpose of the Study:

• To develop and evaluate a novel deep learning-based method, xQSM, for improved QSM image reconstruction.
• To compare the performance of xQSM against other state-of-the-art deep learning and conventional QSM reconstruction techniques.

Main Methods:

• A U-net architecture was employed for the xQSM method, incorporating noise regularization and modified octave convolutional layers.
• The xQSM network was trained using both synthetic and in vivo MRI datasets.
• Performance was evaluated using digital simulations and in vivo experiments, comparing metrics like accuracy, structural similarity, and reconstruction time.

Main Results:

• The xQSM method, particularly when trained with in vivo data, demonstrated superior reconstruction accuracy compared to other deep learning methods (DeepQSM, QSMnet+) and conventional methods (MEDI, iLSQR).
• Significant improvements were observed in globus pallidus susceptibility estimation accuracy, with xQSM outperforming other methods in both simulations and in vivo acquisitions.
• xQSM exhibited enhanced linearity across susceptibility scales and robust generalization to varying spatial resolutions, while drastically reducing reconstruction time from minutes to seconds.

Conclusions:

• The developed xQSM method represents a significant advancement in QSM reconstruction, offering improved accuracy and efficiency.
• xQSM shows strong potential for clinical applications due to its enhanced performance and reduced processing time.
• This deep learning approach provides a more robust and faster solution for quantitative susceptibility mapping.