Unpaired MRI Super Resolution with Contrastive Learning

TL;DR

This work tackles the scarcity of high-resolution MRI training data by proposing an unpaired MRI super-resolution framework augmented with contrastive learning. By embedding an InfoNCE-based contrastive objective into an unsupervised SR architecture, the method creates robust positive and negative sample pairs from unpaired LR-HR MRI images, enabling effective representation learning with limited HR data. Empirical results on HCP data show notable PSNR gains and competitive SSIM, particularly when HR data are scarce, outperforming several state-of-the-art unsupervised methods under data-constrained conditions. The approach holds practical promise for clinical MRI where large HR training datasets are difficult to obtain, and points to future improvements in sample-pair construction and augmentation strategies.

Abstract

Magnetic resonance imaging (MRI) is crucial for enhancing diagnostic accuracy in clinical settings. However, the inherent long scan time of MRI restricts its widespread applicability. Deep learning-based image super-resolution (SR) methods exhibit promise in improving MRI resolution without additional cost. Due to lacking of aligned high-resolution (HR) and low-resolution (LR) MRI image pairs, unsupervised approaches are widely adopted for SR reconstruction with unpaired MRI images. However, these methods still require a substantial number of HR MRI images for training, which can be difficult to acquire. To this end, we propose an unpaired MRI SR approach that employs contrastive learning to enhance SR performance with limited HR training data. Empirical results presented in this study underscore significant enhancements in the peak signal-to-noise ratio and structural similarity index, even when a paucity of HR images is available. These findings accentuate the potential of our approach in addressing the challenge of limited HR training data, thereby contributing to the advancement of MRI in clinical applications.

Figures (3)

• Figure 1: The diagram of the proposed unsupervised contrastive learning model with unpaired LR-HR MRI images. C, G, D are contrastive loss, generator loss and discriminator loss, respectively. The target $X_t$ and source $Y_s$ are fed into the model together to train the model, followed by three data flows: $Y_s \to f_s \to \widehat{Y_s}$, $Y_s \to f_s \to X_{s\sim t} \to f_{s\sim t\sim s} \to Y_{s\sim t\sim s}$, and $X_t \to f_t \to \widehat{X_t}$. Contrastive loss and generator losses are calculated between the data with the same content, such as $f_{s\sim t\sim s}$ to $f_s$ and $\widehat{Y_s}/Y_{s\sim t\sim s}$ to $Y_s$. The discriminator loss is calculated between the data in the same domain but with different contents, such as $X_{s\sim t}$ to $X_t$ and $f_t$ to $f_s$. For inference, only $X_t$ is fed to the model and follows the data flow of $X_t \to f_t \to \widehat{Y_t}$.
• Figure 2: Comparison in visual effect and error maps with various numbers of training HR images. The visualization shows the super-resolution image in the sagittal plane of the HCP dataset which is downsampled with a scale factor of 2×2×2.
• Figure 3: Performance of models with various numbers of HR images. (a) and (b) are SSIM and PSNR results of ablation on contrastive loss for our method, respectively.