Meta-Learning Enabled Score-Based Generative Model for 1.5T-Like Image Reconstruction from 0.5T MRI

Zhuo-Xu Cui,Congcong Liu,Chentao Cao,Yuanyuan Liu,Jing Cheng,Qingyong Zhu,Yanjie Zhu,Haifeng Wang,Dong Liang

2023-05-04

Abstract:Magnetic resonance imaging (MRI) is known to have reduced signal-to-noise ratios (SNR) at lower field strengths, leading to signal degradation when producing a low-field MRI image from a high-field one. Therefore, reconstructing a high-field-like image from a low-field MRI is a complex problem due to the ill-posed nature of the task. Additionally, obtaining paired low-field and high-field MR images is often not practical. We theoretically uncovered that the combination of these challenges renders conventional deep learning methods that directly learn the mapping from a low-field MR image to a high-field MR image unsuitable. To overcome these challenges, we introduce a novel meta-learning approach that employs a teacher-student mechanism. Firstly, an optimal-transport-driven teacher learns the degradation process from high-field to low-field MR images and generates pseudo-paired high-field and low-field MRI images. Then, a score-based student solves the inverse problem of reconstructing a high-field-like MR image from a low-field MRI within the framework of iterative regularization, by learning the joint distribution of pseudo-paired images to act as a regularizer. Experimental results on real low-field MRI data demonstrate that our proposed method outperforms state-of-the-art unpaired learning methods.

Image and Video Processing,Computer Vision and Pattern Recognition,Machine Learning

What problem does this paper attempt to address?

This paper attempts to solve the problem of reconstructing high - quality high - field - strength (1.5T) MRI images from low - field - strength (0.5T) magnetic resonance imaging (MRI). Specifically, the main challenges include:

1. **Reduction in signal - to - noise ratio (SNR) and resolution**: As the magnetic field strength decreases, the SNR and resolution of MRI images also decrease, resulting in a decline in image quality during the transition from high - field - strength to low - field - strength. Therefore, reconstructing high - field - strength images from low - field - strength MRI is an ill - posed inverse problem.

2. **Lack of paired data**: Obtaining paired low - field - strength and high - field - strength MRI images is impractical, which makes it difficult for traditional deep - learning methods to be directly applied to this scenario.

3. **Low coil redundancy**: Low - field - strength MRI devices typically use fewer receiving coils (such as four channels), which further exacerbates the ill - posed nature of the reconstruction problem.

To overcome these challenges, the authors propose a new method based on meta - learning. This method uses a teacher - student mechanism to generate pseudo - paired data and regularizes by learning the joint distribution of these pseudo - paired data, thereby solving the problem of reconstructing high - field - strength images from low - field - strength MRI. The specific steps are as follows:

- **Teacher model**: The teacher model driven by optimal transport (OT) learns the degradation process from high - field - strength to low - field - strength and generates pseudo - paired high - field - strength and low - field - strength MRI images.

- **Student model**: Through the score - based student model, the joint distribution of the pseudo - paired data is learned within the framework of iterative regularization to solve the inverse problem of reconstructing high - field - strength images.

Experimental results show that this method is superior to existing unpaired - learning methods in terms of reconstruction contrast, SNR, and detail - structure visualization.

### Formula summary

- **Degradation model**:
\[
x_{0.5}=f(x_{1.5})
\]
where \(x_{0.5}\) and \(x_{1.5}\) are 0.5T and 1.5T MRI images respectively, and \(f\) is an unknown non - linear degradation process.

- **k - space data acquisition model**:
\[
y_{0.5}=A x_{0.5}+n
\]
where \(y_{0.5}\) is the undersampled 0.5T k - space data, \(A\) is the encoding matrix, and \(n\) is Gaussian measurement noise.

- **Optimal transport problem**:
\[
T_{\alpha\rightarrow\beta}=\arg\min_T\left\{\sum_{i = 1}^N c(T(x_{1.5}^i), x_{1.5}^i):T^{\sharp}\alpha=\beta\right\}
\]
where \(\alpha\) and \(\beta\) are the distributions of 1.5T and 0.5T MRI images respectively.

- **Posterior probability model**:
\[
\log p(\tilde{x}_{1.5},\tilde{x}_{0.5}|y_{0.5})=\log p(y_{0.5}|\tilde{x}_{0.5})+\log p(\tilde{x}_{0.5},\tilde{x}_{1.5})+C
\]

Through these formulas and methods, the authors have successfully solved the key problem of reconstructing high - field - strength images from low - field - strength MRI.