Abstract:The ability to reconstruct high-quality images from undersampled MRI data is vital in improving MRI temporal resolution and reducing acquisition times. Deep learning methods have been proposed for this task, but the lack of verified methods to quantify the uncertainty in the reconstructed images hampered clinical applicability. We introduce "NPB-REC", a non-parametric fully Bayesian framework, for MRI reconstruction from undersampled data with uncertainty estimation. We use Stochastic Gradient Langevin Dynamics during training to characterize the posterior distribution of the network parameters. This enables us to both improve the quality of the reconstructed images and quantify the uncertainty in the reconstructed images. We demonstrate the efficacy of our approach on a multi-coil MRI dataset from the fastMRI challenge and compare it to the baseline End-to-End Variational Network (E2E-VarNet). Our approach outperforms the baseline in terms of reconstruction accuracy by means of PSNR and SSIM ($34.55$, $0.908$ vs. $33.08$, $0.897$, $p<0.01$, acceleration rate $R=8$) and provides uncertainty measures that correlate better with the reconstruction error (Pearson correlation, $R=0.94$ vs. $R=0.91$). Additionally, our approach exhibits better generalization capabilities against anatomical distribution shifts (PSNR and SSIM of $32.38$, $0.849$ vs. $31.63$, $0.836$, $p<0.01$, training on brain data, inference on knee data, acceleration rate $R=8$). NPB-REC has the potential to facilitate the safe utilization of deep learning-based methods for MRI reconstruction from undersampled data. Code and trained models are available at \url{

What problem does this paper attempt to address?

This paper proposes a non-parametric Bayesian deep learning method called NPB-REC to solve the undersampled magnetic resonance imaging (MRI) reconstruction problem and provide uncertainty estimation. Traditional methods such as parallel imaging and compressed sensing have limitations in handling undersampled MRI data, while deep learning methods improve the quality of image reconstruction but lack reliable quantification of reconstruction uncertainty. The NPB-REC framework utilizes Stochastic Gradient Langevin Dynamics (SGLD) training to characterize the posterior distribution of network parameters, thereby improving both the quality of reconstructed images and the quantification of uncertainty. Compared to the baseline method based on End-to-End Variational Network (E2E-VarNet), NPB-REC demonstrates higher reconstruction accuracy (measured by PSNR and SSIM) and better correlation with uncertainty measurement on a multi-coil MRI dataset. Moreover, NPB-REC exhibits better generalization ability in the presence of adversarial anatomical distribution variations. The main contributions of this paper are as follows: 1. Proposing a non-parametric Bayesian deep learning method for MRI undersampled data reconstruction. 2. Fully describing the posterior distribution of reconstructed images. 3. Providing uncertainty estimation of reconstructed images, which is correlated with reconstruction errors, anatomical distribution variations, and undersampling pattern variations. 4. Experimental results show the additional value of NPB-REC compared to baseline methods. By conducting experiments on the fastMRI challenge dataset, NPB-REC demonstrates its advantages in improving reconstruction quality and uncertainty estimation. The code and trained models have been made publicly available, providing a safer approach for clinical application of MRI reconstruction.

NPB-REC: A Non-parametric Bayesian Deep-learning Approach for Undersampled MRI Reconstruction with Uncertainty Estimation

Deep Learning for Under-Sampled Non-Cartesian ASL MRI Reconstruction

Uncertainty Estimation and Propagation in Accelerated MRI Reconstruction

NPBDREG: Uncertainty Assessment in Diffeomorphic Brain MRI Registration using a Non-parametric Bayesian Deep-Learning Based Approach

Bayesian MRI reconstruction with joint uncertainty estimation using diffusion models

UncertaINR: Uncertainty Quantification of End-to-End Implicit Neural Representations for Computed Tomography

Scan-specific Self-supervised Bayesian Deep Non-linear Inversion for Undersampled MRI Reconstruction

Uncertainty Estimation in Medical Image Denoising with Bayesian Deep Image Prior

Adaptive Local Neighborhood-based Neural Networks for MR Image Reconstruction from Undersampled Data

Enhancing quality and speed in database‐free neural network reconstructions of undersampled MRI with SCAMPI

Enhancing quality and speed in database-free neural network reconstructions of undersampled MRI with SCAMPI

Stable Deep MRI Reconstruction using Generative Priors

Bayesian Reconstruction of Magnetic Resonance Images using Gaussian Processes

Deep learning for undersampled MRI reconstruction

Deep Magnetic Resonance Image Reconstruction: Inverse Problems Meet Neural Networks

Benchmarking 3D multi-coil NC-PDNet MRI reconstruction

ERNAS: An Evolutionary Neural Architecture Search for Magnetic Resonance Image Reconstructions

Estimating Uncertainty in PET Image Reconstruction via Deep Posterior Sampling

Predictive uncertainty in deep learning–based MR image reconstruction using deep ensembles: Evaluation on the fastMRI data set

Accelerated MRI with Un-trained Neural Networks