The proposed SGM-Net leverages low-quality samples obtained using score-based generative priors to guide a model-driven network, effectively mitigating hallucination artifacts and achieving robust and high-quality MRI reconstruction results.
