Konsep Inti
A non-parametric Bayesian deep learning framework is proposed to reconstruct high-quality MRI images from undersampled k-space data while providing quantitative measures of uncertainty in the reconstructed images.
Abstrak
The content describes a novel non-parametric Bayesian deep learning approach, called NPB-REC, for magnetic resonance imaging (MRI) reconstruction from undersampled k-space data. The key highlights are:
- NPB-REC employs Stochastic Gradient Langevin Dynamics (SGLD) during training to sample from the posterior distribution of the network parameters, enabling full characterization of the uncertainty in the reconstructed images.
- The proposed method outperforms the baseline End-to-End Variational Network (E2E-VarNet) in terms of reconstruction accuracy, as measured by PSNR and SSIM, particularly at higher acceleration rates.
- NPB-REC exhibits better generalization capabilities against anatomical distribution shifts and undersampling mask distribution shifts compared to the baseline and Monte Carlo Dropout methods.
- The uncertainty measures provided by NPB-REC correlate better with the reconstruction error and can effectively detect out-of-distribution data, facilitating the safe utilization of deep learning-based MRI reconstruction methods in clinical settings.
Statistik
The authors report the following key metrics:
PSNR and SSIM values for the whole image and annotated regions of interest (ROIs) at acceleration rates R=4, R=8, and R=12.
Pearson correlation coefficient between the uncertainty measure and reconstruction error (MSE) for NPB-REC and Monte Carlo Dropout.