Structural-Mr Informed Qsm Reconstruction Using Deep Image Prior

Quantitative Susceptibility Mapping (QSM) is an emerging quantitative magnetic resonance imaging (MRI) technique that aims to provide a spatial map of tissue magnetic susceptibility. Obtaining a large dataset of high-quality single-orientation QSM data remains a challenge and thus limits the applicability of supervised learning-based schemes. This work proposes a hybrid model that accounts for both data-fidelity and an image enhancement stage using an unsupervised neural network inspired by the deep image prior framework. Further, we use patient-specific anatomical information from a structural MRI scan as prior to stabilize the training process. Our reconstructed images with noisy phase data improves over conventional denoisers that do not require training data.