Abstract: Magnetic resonance fingerprint is an efficient multi-parameter quantitative imaging technology, but the traditional reliance on signal dictionaries for parameter quantization has problems such as large discretization errors and low matching efficiency. In view of the limitations of existing supervised learning methods that rely on pseudo-labels and lack physical interpretability, this paper proposes a self-supervised parameter quantization method that integrates imaging physical models and manifold structure modeling. This method obtains reliable unlabeled constraints through self-supervised physical consistency learning driven by Bloch equations, and combines knowledge distillation driven by manifold structure to migrate the features of long frames to short frame models, realizing the joint optimization of physical constraints and structural priors, thereby simultaneously improving the accuracy and efficiency under unlabeled conditions. Experiments have verified the advantages of this method in terms of accuracy and robustness, and provide new ideas for achieving efficient and reliable magnetic resonance fingerprint parameter estimation.

Key words: magnetic resonance fingerprint imaging, self-supervised learning, parameter quantization, imaging physical model, manifold structure