χ-sepnet: Deep Neural Network for Magnetic Susceptibility Source Separation.

Area of Science:

• Neuroimaging
• Medical Physics
• Artificial Intelligence in Medicine

Background:

• Quantitative Susceptibility Mapping (QSM) methods like magnetic susceptibility source separation (χ-separation) estimate brain's paramagnetic and diamagnetic sources.
• Conventional QSM suffers from streaking artifacts and requires time-consuming data acquisition for accurate susceptibility mapping.
• Existing methods face challenges in resolving susceptibility distributions and require extensive imaging protocols.

Purpose of the Study:

• To develop novel deep learning-based pipelines for artifact reduction and improved accuracy in susceptibility source separation.
• To propose two pipelines: χ-sepnet- $$ (using GRE and spin-echo data) and χ-sepnet- $$ (using GRE data only).
• To evaluate the performance of these pipelines against conventional methods in healthy subjects and multiple sclerosis patients.

Main Methods:

• Development of a deep learning network, χ-sepnet, trained on streaking artifact-free data from multiple head orientations.
• Implementation of two pipelines: χ-sepnet- $$ and χ-sepnet- $$ , utilizing multi-echo GRE and optional multi-echo spin-echo data.
• Qualitative and quantitative assessments in healthy subjects and analysis of lesion characteristics in multiple sclerosis patients.

Main Results:

• The proposed deep learning pipelines generated high-quality χ-separation maps with substantially reduced artifacts compared to conventional methods.
• χ-sepnet- $$ and χ-sepnet- $$ demonstrated superior performance in quantitative analysis, outperforming regularization-based reconstruction.
• Accurate classification of multiple sclerosis lesions into paramagnetic and diamagnetic subtypes (99.6% and 98.4% concordance, respectively).

Conclusions:

• Deep learning-based susceptibility source separation pipelines (χ-sepnet) offer significant improvements in image quality and artifact reduction.
• The χ-sepnet- $$ pipeline, requiring only multi-echo GRE data, shows great potential for widespread clinical and scientific applications.
• Further evaluation is warranted for diverse diseases and pathological conditions to fully establish clinical utility.