HyperCMR: Enhanced Multi-Contrast CMR Reconstruction with Eagle Loss

TL;DR

HyperCMR is presented, a novel framework designed to accelerate the reconstruction of multi-contrast cardiac magnetic resonance (CMR) images by incorporating advanced loss functions, notably the innovative Eagle Loss, which is specifically designed to recover missing high-frequency information in undersampled k-space.

Abstract

Accelerating image acquisition for cardiac magnetic resonance imaging (CMRI) is a critical task. CMRxRecon2024 challenge aims to set the state of the art for multi-contrast CMR reconstruction. This paper presents HyperCMR, a novel framework designed to accelerate the reconstruction of multi-contrast cardiac magnetic resonance (CMR) images. HyperCMR enhances the existing PromptMR model by incorporating advanced loss functions, notably the innovative Eagle Loss, which is specifically designed to recover missing high-frequency information in undersampled k-space. Extensive experiments conducted on the CMRxRecon2024 challenge dataset demonstrate that HyperCMR consistently outperforms the baseline across multiple evaluation metrics, achieving superior SSIM and PSNR scores.

Figures (5)

• Figure 1: Overview of the HyperCMR framework. Our pipeline includes generation of sensitivity maps and repaired k-space generation with promptUnet.
• Figure 2: Visualization of the impact of different patch sizes and cutoff frequencies on the high-pass filter and resulting magnitude maps. The selected patch size of 5 and cutoff of 0.35 provide balanced performance in capturing mid-level frequencies and spatial details in the HyperCMR framework.
• Figure 3: The figure illustrates the detailed workflow of the Eagle Loss implementation for calculating gradients in the x-direction. The process begins by computing the gradients for both the predicted and target images, followed by calculating variance across non-overlapping patches. These variance maps are then transformed using a 2D FFT, and their magnitudes are filtered with a Butterworth high-pass filter. Finally, the L1 loss is calculated by comparing the filtered magnitudes in the frequency domain. The same process is applied to the y-direction gradients. The total Eagle Loss is computed as the sum of the loss for both x and y-direction gradients.
• Figure 4: Visual comparison of reconstructed images for three different CMR modalities (aorta_tra, cine_lvot, cine_sax) at 10x acceleration. The SSIM values and zoomed-in regions highlight the superior performance of our model (HyperCMR) compared to the PromptMR baseline. This figure provides a clear example of the improvements achieved by our method in preserving fine details and enhancing overall image quality, particularly in the high-frequency regions of the undersampled k-space.
• Figure 5: Task2 result. Some examples