Enhanced Motion-Corrected Deep-Learning Reconstruction Framework for fast 3D Whole-Heart MRA at 0.55T

Background: Short acquisition and reconstruction times in cardiac MRA (CMRA) are critical for the assessment of cardiovascular disease. Recently, an end-to-end deep learning approach for non-rigid motion estimation and motion-corrected model-based (MoCo-MoDL) reconstruction for image-navigator (iNAV)-based 3D whole-heart CMRA was proposed at 1.5T [1,2] showing promising results for 7-fold-undersampled data. Lower field systems offer advantages in terms of accessibility and cost. However, the direct application of MoCo-MoDL to low-field systems resulted in suboptimal performance due to lower SNR [3]. In this work, we propose an adaptation and enhancement of the previously introduced MoCo-MoDL framework (eMoCo-MoDL) to produce high-quality images from ~2.4min scans with 15s reconstruction time at 0.55T.

Methods: Direct implementation of the 1.5T MoCo-MoDL (Fig. 1a) on 0.55T data results in degraded image quality. To address these limitations, a residual path was integrated into each stage of the U-Net of both networks (Fig. 1b). A perceptual and k-space loss were also added. The dataset includes 27 healthy subjects (HS) and 10 patients with CHD scanned at 0.55T (MAGNETOM Free.Max, Siemens) using the iNAV-based CMRA framework [6] with 1.5mm³ isotropic resolution and 4-fold undersampling for training and testing. In addition, 12 HS were acquired with 7-fold undersampling for evaluation of prospective reconstruction. The 4-fold undersampled volumes were reconstructed using a non-rigid motion corrected framework with patch-based regularization (NR-PROST [7]), to generate a reference image for each subject and each respiratory state. Each 4-fold accelerated k-space dataset was retrospectively undersampled to 7-fold, and binned to obtain zero-filled volumes that served as input to the network. The data was divided into 30 training sets (21 HS) and 7 testing sets. The loss combines: (i) Charbonnier loss between end-expiration reference and warped references, (ii) Charbonnier loss between reference and network output, (iii) MSE between VGG features of reference and output, and (iv) Charbonnier loss between weighted k-spaces of reference and output.

Results: 3D iNAV-based CMRA reconstructions with eMoCo-MoDL and NR-PROST are shown in Fig. 2 and Fig. 3 for retrospective and prospective 7-fold undersampling for 3 representative subjects at 0.55T. Excellent visual quality results were achieved with eMoCo-MoDL for both cases, showing superior image quality than 7-fold NR-PROST (especially for the prospective case) and comparable image quality to the reference 4-fold acquisition. eMoCo-MoDL reconstruction time was 160 times faster than the NR-PROST reference (15s vs 2400s, respectively).

Conclusion: An adapted and enhanced eMoCo-MoDL framework was presented for 7-fold iNAV-based 3D whole-heart CMRA at 0.55T. eMoCo-MoDL enables an acquisition time of 2.4min with a reconstruction time of ~15s. Future work will focus in further validation and adapting eMoCo-MoDL for multi-contrast simultaneous reconstruction.

Figure 1: (A) eMoCo-MoDL framework. Zero-filled bin images are input as patches to the DiRespMe-Net [4], which outputs the estimated 3D non- rigid motion fields between the bin image reference (end-expiration) and the remaining bins. The displacement fields estimated by DiRespMe-Net are incorporated into the data consistency step of the model-based reconstruction [5], which iterates between the data consistency step and a regularization U-Net. At each epoch, a 112x272x96-voxel patch was randomly selected from the reference and zero filled images. (B) The modified U-Net. The architecture employs residual connections at each stage. To match the channel dimensions in the residual paths, a convolution with a kernel size of 3 is applied in both the residual and direct paths.
Figure 2: Coronal slices of iNAV-based 3D whole-heart MRI at 0.55T for 3 representative test subjects with 7-fold retrospective undersampling. 2 healthy subjects and 1 CHD patient (third row) are shown. The figure includes reference dataset (4-fold undersampling) reconstructed with NR-PROST (left column), NR-PROST for 7-fold undersampling (middle column), and proposed eMoCo-MoDL for 7-fold undersampling (right column). The SSIM and the MSE was computed for each case.
Figure 3: Coronal slices of iNAV-based 3D whole-heart MRI at 0.55T for 3 representative subjects with 7-fold prospective undersampling. The figure includes reference dataset (4-fold undersampling) reconstructed with NR-PROST acquired in a separate scan (left column), NR-PROST for 7-fold undersampling (middle column), and proposed eMoCo-MoDL for 7-fold undersampling (right column). Sharper details and better visual image quality is observed with eMoCo-MoDL in comparison to NR-PROST.