TagGen: Diffusion-based Generative Model for Cardiac MR Tagging Super Resolution (RF_FR_373)

Cardiac MR tagging is recognized as the gold standard for assessing regional myocardial deformation (1, 2). To incorporate MR tagging (3, 4) into an efficient clinical workflow, there are multiple approaches developed for acceleration such as parallel imaging (5), model-based methods (6) and deep learning-based algorithms (7). However, the relatively lower SNR and sensitivity of tagged signals in MR tagging limit its compatibility with high acceleration rates. Recent advancements in deep generative models have shown promise for super-resolution (SR) tasks in cardiac MR imaging (7, 8). Diffusion models have demonstrated the capability of training stability and high-quality sample generation across various vision tasks (9-11). In this study, we sought to develop TagGen, a cascaded diffusion-based conditional generative model that uses low-resolution (LR) MR tagging images as guidance to generate corresponding high-resolution (HR) MR tagging images.

Methods:

Dataset: a) Retrospective Data: 120 long-axis slices from 50 patients underwent CMR tagging scans were used (1.5T Aera, Siemens), with a random split of 8:2 per subject level for train and test sets. We synthesized the LR-HR pairs by keeping the center 30% of phase-encoding lines of HR image k-space and zero-padded the outer k-space, and then center cropping both the synthesized LR images and source HR image to 128x128 (Figure 1A). The low-resolution acquisition ratio is set at 30% to cover the majority of the harmonic peak regions which contains the tagline information. b) Prospective 10-fold Acceleration Testing Data: 4 patients (3T Vida, Siemens) and 2 volunteers (1.5T Aera, Siemens) were scanned with a 30% central phase-encoding k-space (R=3.3) and GRAPPA-3, with each slice acquired in 3 heartbeats.

Model: We used the conditional Denoising Diffusion Probabilistic Models model as shown in Figure 1B, which reconstruct HR tagging image using the LR tagging image as the guidance. To improve the generation quality, we employed cascaded strategies (N=3) to iteratively apply the trained model multiple times and enforcing k-space data consistency on the output of the previous stage (Figure 1C). For model hyperparameter settings, we used five levels in U-Net, with the number of channels in each level being [64,128,256,512,512], and the model was trained for 100,000 iterations with AdamW optimizer with a learning rate of 3e-5 and a batch size of 128. 

Results:

For synthetic 3.3-fold accelerated data, TagGen demonstrated superior performance than REGAIN with sharper tag grids, better tag grid fidelity, and higher overall quality, as referenced by HR images (Figure 2A). TagGen outperformed REGAIN in terms of nRMSE, PSNR, and SSIM (Figure 2B). For prospectively 10-fold accelerated data, TagGen showed better tag grid quality for the volunteer and patient examples (Figure 3A, B). TagGen provided better tag grid quality, SNR, and overall image quality than REGAIN as blinded scored by two radiologists (Figure 3C).

Conclusion:

Leveraging the harmonic peak physical property, we developed TagGen, a conditional diffusion-based super-resolution model designed to recover tagging grids and enhance image quality. Incorporating a low-resolution tagging acquisition that integrates with parallel imaging, TagGen enables highly accelerated MR tagging while providing SNR loss compensation. We demonstrated the feasibility of 10-fold highly accelerated tag grid-based MR tagging, achieving the acquisition of a single slice in 3 heartbeats with enhanced tag grids.