A data-driven deep learning pipeline for quantitative susceptibility mapping (QSM).

Area of Science:

• Medical Imaging
• Deep Learning
• Quantitative Susceptibility Mapping (QSM)

Background:

• Quantitative Susceptibility Mapping (QSM) is crucial for analyzing magnetic resonance imaging (MRI) data.
• Existing deep learning models for QSM often lack accuracy and robustness.
• There is a need for improved QSM quantification methods, especially for clinical applications.

Purpose of the Study:

• To develop and validate a data-driven optimization approach to enhance deep learning-based QSM quantification accuracy.
• To introduce a novel QSM pipeline (POCSnet1 and POCSnet2) integrating projections onto convex set (POCS) models.
• To compare the performance of the proposed pipeline against established QSM methods.

Main Methods:

• Developed a deep learning QSM pipeline using two POCS models (POCSnet1 for background field removal, POCSnet2 for dipole inversion).
• Employed unrolled V-Net architecture with iterative data-driven optimization to enforce data fidelity.
• Trained models using simulated phantom data and evaluated on synthetic, COSMOS, Parkinson's disease, and small-vessel disease cohorts.

Main Results:

• POCSnet1 demonstrated superior performance to naïve V-Net models on COSMOS data (NRMSE 23.7% vs. 62.7%).
• POCSnet2 achieved lower NRMSE (58.1%) and HFEN (56.7%) compared to DLL2, FINE, and autoQSM on COSMOS data.
• The proposed pipeline showed high consistency with VSHARP+STAR-QSM on clinical data and improved cerebral microbleed detection sensitivity (100% vs. 92%).

Conclusions:

• Data-driven optimization significantly improves the accuracy of deep learning-based QSM quantification.
• The proposed POCS-based deep learning pipeline offers a robust and accurate solution for QSM.
• This method holds promise for enhanced diagnostic capabilities in neurological disease research.