Quantitative susceptibility mapping using multi-channel convolutional neural networks with dipole-adaptive...

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Quantitative susceptibility mapping (QSM) is crucial for assessing brain tissue composition (iron, myelin, calcium) in neurological diseases.
• QSM reconstruction faces challenges due to the ill-posed nature of the dipole kernel, particularly its zero-frequency response.
• Deep learning (DL) shows promise for QSM but often neglects the dipole kernel's inherent properties.

Purpose of the Study:

• To introduce a novel deep learning method, DIAM-CNN, that incorporates dipole kernel characteristics for improved QSM reconstruction.
• To address the limitations of existing DL methods in QSM by accounting for the dipole kernel's frequency response.
• To enhance the accuracy and reduce artifacts in QSM for better clinical applications.

Main Methods:

• Proposed a dipole kernel-adaptive multi-channel convolutional neural network (DIAM-CNN).
• DIAM-CNN processes tissue field components (high-fidelity and low-fidelity) derived from dipole kernel thresholding.
• Utilized a multichannel 3D Unet architecture, trained and validated using COSMOS-derived QSM maps.

Main Results:

• DIAM-CNN demonstrated superior image quality compared to conventional methods (MEDI, iLSQR) and another DL method (QSMnet) in healthy volunteers.
• Quantitative metrics including HFEN, PSNR, NRMSE, and SSIM confirmed DIAM-CNN's improved performance.
• Experiments with simulated hemorrhagic lesions showed DIAM-CNN produced significantly fewer shadow artifacts.

Conclusions:

• Integrating dipole kernel-specific knowledge into network design substantially improves deep learning-based QSM reconstruction.
• DIAM-CNN offers a promising advancement for accurate and artifact-reduced QSM, aiding in the diagnosis and monitoring of brain diseases.
• This adaptive approach highlights the potential for physics-informed deep learning in medical imaging.