[Paper Review] Scan-specific Self-supervised Bayesian Deep Non-linear Inversion for Undersampled MRI Reconstruction

Andrew P. Leynes, Nikhil Deveshwar+2 more|arXiv (Cornell University)|Mar 1, 2022

Advanced MRI Techniques and Applications4 citations

TL;DR

This paper proposes a scan-specific, self-supervised Bayesian deep non-linear inversion (DNLINV) method for undersampled MRI reconstruction that eliminates the need for fully-sampled data or auto-calibration signals. By leveraging deep image priors and approximate Bayesian inference, DNLINV achieves state-of-the-art reconstruction quality across diverse anatomies, contrasts, and sampling patterns without requiring external datasets or calibration scans.

ABSTRACT

Magnetic resonance imaging is subject to slow acquisition times due to the inherent limitations in data sampling. Recently, supervised deep learning has emerged as a promising technique for reconstructing sub-sampled MRI. However, supervised deep learning requires a large dataset of fully-sampled data. Although unsupervised or self-supervised deep learning methods have emerged to address the limitations of supervised deep learning approaches, they still require a database of images. In contrast, scan-specific deep learning methods learn and reconstruct using only the sub-sampled data from a single scan. Here, we introduce Scan-Specific Self-Supervised Bayesian Deep Non-Linear Inversion (DNLINV) that does not require an auto calibration scan region. DNLINV utilizes a deep image prior-type generative modeling approach and relies on approximate Bayesian inference to regularize the deep convolutional neural network. We demonstrate our approach on several anatomies, contrasts, and sampling patterns and show improved performance over existing approaches in scan-specific calibrationless parallel imaging and compressed sensing.

Motivation & Objective

• To address the limitation of supervised deep learning in MRI reconstruction, which requires large datasets of fully-sampled scans.
• To overcome the dependency on auto-calibration signals (ACS) in parallel imaging and compressed sensing methods.
• To develop a self-supervised, scan-specific method that reconstructs undersampled MRI using only the sub-sampled data from a single scan.
• To improve image quality and generalization across diverse anatomical structures, contrasts, and sampling patterns without external data.
• To integrate deep image priors with approximate Bayesian inference for robust regularization in deep neural networks.

Proposed method

• The method employs a deep convolutional neural network (CNN) as a generative model, initialized with a deep image prior to encode image structure without pretraining.
• It formulates the MRI reconstruction problem as a Bayesian inference task, using approximate Bayesian inference to regularize the network and reduce overfitting to undersampled data.
• The approach is self-supervised, meaning it trains solely on the sub-sampled k-space data from a single scan, without requiring fully-sampled reference data.
• A variational inference framework is used to approximate the posterior distribution over image parameters, enabling uncertainty quantification and improved generalization.
• The method is designed to be scan-specific, adapting to the unique characteristics of each individual MRI acquisition without requiring a database of prior scans.
• It is applicable to both parallel imaging and compressed sensing sampling patterns, demonstrating flexibility across different k-space trajectories.

Experimental results

Research questions

• RQ1 Can a self-supervised deep learning method achieve high-quality MRI reconstruction without requiring fully-sampled data or auto-calibration signals?
• RQ2 How does a scan-specific deep image prior combined with Bayesian regularization improve reconstruction fidelity in undersampled MRI?
• RQ3 To what extent can a method trained solely on sub-sampled k-space data generalize across different anatomies, contrasts, and sampling patterns?
• RQ4 How does the integration of approximate Bayesian inference enhance robustness and uncertainty estimation in deep MRI reconstruction?
• RQ5 Does the proposed DNLINV method outperform existing scan-specific and self-supervised approaches in terms of image quality and reconstruction accuracy?

Key findings

• The proposed DNLINV method achieves superior image reconstruction quality compared to existing scan-specific and self-supervised methods across multiple anatomical regions and contrasts.
• The method demonstrates consistent performance improvements in both parallel imaging and compressed sensing settings, even without access to fully-sampled data or ACS regions.
• By eliminating the need for external datasets or calibration scans, the method enables practical deployment in clinical settings where fully-sampled data is unavailable.
• The integration of deep image priors and Bayesian inference leads to enhanced generalization and reduced overfitting, particularly in low-sampling regimes.
• The approach enables uncertainty-aware reconstruction, providing probabilistic estimates of image quality and confidence in the output.
• The method achieves state-of-the-art performance in terms of PSNR and SSIM on benchmark datasets, with quantitative improvements over prior scan-specific approaches.