Motion-Corrected Deep-Learning Reconstruction Framework for 3D Whole-Heart MRA at 0.55T (RF_SA_448)

Background: Short cardiac MR angiography (CMRA) acquisition and reconstruction times are critical for the assessment of cardiovascular disease in clinical settings. Recently, an end-to-end deep learning approach for non-rigid motion estimation and motion-corrected model-based (MoCo-MoDL) reconstruction for whole-heart CMRA was proposed at 1.5T^1,2. However, this method has not yet been tested on lower field systems, which may offer potential advantages in terms of accessibility and cost at the expense of lower SNR. In this work, we show a proof of concept of the MoCo-MoDL framework to produce good quality images in shorter times than state-of-the-art methods in healthy subjects and patients with congenital heart disease (CHD) at 0.55T.

Methods: MoCo-MoDL (Fig. 1) consists of: 1) a non-rigid motion estimation network (RespMe-Net), which estimates 3D displacement fields from respiratory bin images, and 2) an unrolled reconstruction framework that alternates between data consistency and denoising using a U-Net.
The dataset includes 12 healthy subjects and 9 patients scanned at 0.55T (MAGNETOM Free.Max, Siemens) using the CMRA framework^3,4 with 4-fold undersampling, except one healthy subject scanned with 7-fold undersampling. As preprocessing, the 4-fold undersampled volumes were reconstructed using a non-rigid motion corrected framework with patch-based regularization (NR-PROST)⁵, to generate a reference image for each subject and each respiratory state. Each 4-fold accelerated k-space dataset was retrospectively undersampled to 7-fold², and after binning and soft gating, zero-filled reconstructions for each respiratory state served as input to the network.The data was divided into 19 training sets and 2 testing sets, including the prospectively 7-fold undersampled data for testing. The loss function includes the Charbonnier loss between the end-expiration-bin reference and the warped reference of the remaining respiratory states (L_mot), the Charbonnier loss between the end-expiration reference and the network output (L_recon), the SSIM between the end-expiration-reference and the network output (L_SSIM) and a regularization loss (L_reg) that was calculated as the l^2-norm of the convolution kernel parameters in both networks.
L = λ₁L_mot + λ₂L_recon + λ₃L_SSIM + λ₄L_reg
MoCo-MoDL was trained on a 64 GB NVIDIA RTX A6000 using a batch size=1, learning rate=2e-4, MoDL iterations = 4, λ₁=100,λ₂=10, λ₃=5, and λ₄=1. At each epoch, a 48 x 272 x 96-voxel path was randomly selected from the reference and zero filled images.

Results: 3D CMRA at 0.55T reconstructions with MoCo-MoDL at 0.55T and the corresponding references are shown in Fig. 2 and Fig. 3 for retrospectively and prospectively 7-fold undersampled data. Acceptable results were found for both cases, though improvements are still needed to visualize small structures and details, specifically for the prospective case. Increasing the number of training samples, along with utilizing data augmentation techniques, may improve the results. MoCo-MoDL reconstruction time was 65 times faster than that of the PROST reference (46 seconds and 3000 seconds, respectively).

Conclusion: In this work, we show a proof of concept of the MoCo-MoDL framework to enable 7-fold accelerated 3D whole-heart CMRA at 0.55T, enabling an acquisition time of 1.36 min and reconstruction time of ~46s, 65 times faster than the state-of-the-art. Future work will focus on further optimizing the training process and evaluation in a larger cohort of subjects.