Quantitative susceptibility mapping through model-based deep image prior (MoDIP).

Area of Science:

• Medical Imaging
• Computational Neuroscience
• Machine Learning

Background:

• Supervised learning methods for Quantitative Susceptibility Mapping (QSM) struggle with generalization due to varying scan parameters.
• Existing methods face limitations in accurately solving dipole inversion across diverse datasets.

Purpose of the Study:

• To develop a novel, training-free, unsupervised method for QSM dipole inversion that overcomes the generalization limitations of supervised approaches.
• To introduce MoDIP (Model-based Deep Image Prior) as a robust solution for QSM.

Main Methods:

• Proposed MoDIP, a training-free unsupervised method combining an untrained network for implicit image prior and a Data Fidelity Optimization (DFO) module.
• The DFO module enforces the physical model of QSM dipole inversion during optimization.
• The untrained network converges to an interim state, providing regularization for image reconstruction.

Main Results:

• MoDIP demonstrated excellent generalizability for QSM dipole inversion across different scan parameters.
• Achieved over 32% accuracy improvement compared to supervised deep learning methods, particularly in pathological brain QSM.
• Showcased 33% greater computational efficiency and 4x speed improvement over conventional DIP-based methods.
• Enabled 3D high-resolution image reconstruction in under 4.5 minutes.

Conclusions:

• MoDIP offers a robust and efficient solution for QSM dipole inversion, addressing the generalization challenges of supervised methods.
• The training-free, model-based approach provides superior accuracy and speed for medical image reconstruction.
• MoDIP's performance highlights the potential of unsupervised learning in advanced neuroimaging applications.