Adaptive regularization weight selection for compressed sensing MRI reconstruction

Proper regularization weights are crucial for the reconstruction quality of compressed sensing (CS) MRI. This work aims to develop an automatic and adaptive regularization weights selection method for CS reconstruction METHODS: A statistical model based on Bayesian theory is designed, incorporating prior information about the Gaussian distribution of incoherent noise and the Laplacian distribution of wavelet coefficients in the wavelet transform domain. Using the variance of coefficients and noise, the adaptive regularization weight for achieving optimal reconstruction quality in each iteration step is obtained through a maximum a posteriori estimator. The adaptive regularization weights vary across different subjects, slices, iterations, and wavelet sub-bands RESULTS: The efficacy of the proposed method was demonstrated through retrospective and prospective studies. Compared to reconstruction results using optimal fixed regularization weights and sparsity-adaptive composite recovery method (SCoRe), the proposed method successfully reduces reconstruction errors and effectively recovers original signals from noise-like incoherent artifacts in the wavelet transform domain. It also saves weight selection time when searching for optimal fixed regularization weights CONCLUSION: We propose an adaptive regularization weights selection method for CS-MRI reconstruction. It provides optimal regularization weights for different subjects, slices, and iterations without requiring manual intervention.