ISDU-QSMNet: Iteration Specific Denoising With Unshared Weights for Improved QSM Reconstruction.

Area of Science:

• Medical Imaging
• Computational Neuroscience
• Artificial Intelligence

Background:

• Quantitative susceptibility mapping (QSM) is crucial for estimating tissue magnetic susceptibility from MRI phase data.
• Solving the inverse problem in QSM is computationally challenging and requires robust reconstruction methods.
• Existing deep learning approaches for QSM have limitations in robustness and training efficiency.

Purpose of the Study:

• To introduce ISDU-QSMNet, an end-to-end model-based deep learning framework for QSM reconstruction.
• To improve the accuracy, robustness, and training efficiency of QSM reconstruction.
• To evaluate ISDU-QSMNet against existing model-based and pure deep learning QSM methods.

Main Methods:

• Developed ISDU-QSMNet, incorporating unshared denoiser weights and random subset sampling for training.
• Evaluated the framework on 94 imaging volumes with varying acquisition parameters.
• Compared performance against LPCNN, SpiNet-QSM, QSMnet, DeepQSM, and xQSM under full and limited training data scenarios.

Main Results:

• ISDU-QSMNet demonstrated substantial improvements in QSM reconstruction, reducing high-frequency error norm (HFEN) by up to 3.5% with full training data.
• In limited training data scenarios, ISDU-QSMNet matched the performance of state-of-the-art model-based deep learning methods.
• The proposed approach showed strong generalization capabilities across different acquisition parameters and in ROI analysis.

Conclusions:

• ISDU-QSMNet offers a powerful, robust, and training-efficient solution for QSM reconstruction.
• The novel deep learning framework enhances QSM accuracy and effectively handles diverse datasets.
• ISDU-QSMNet represents a significant advancement in model-based deep learning for quantitative susceptibility mapping.