SCMR/ISMRM Co-Provided Workshop
SCMR 22nd Annual Scientific Sessions
3D whole-heart coronary MR angiography (CMRA) has shown significant potential to noninvasively visualize coronary arteries. However, acquisition remains lengthy as a large amount of data need to be acquired to obtain high-resolution images. Undersampled compressed sensing (CS) reconstruction approaches have been applied to accelerate CMRA1-2. For high acceleration factors, however, CS-based techniques suffer from residual aliasing artifacts and long reconstruction times.
We propose to use a variational neural network (VNN)3 as a probabilistic prior embedded in an iterative direction method of multipliers (ADMM) optimization framework. We utilize the ADMM+VNN approach for the reconstruction of 3D high-resolution images from prospectively undersampled translational motion-corrected CMRA data; and we compare its results with CS for an undersampling factor of 9x.
Acquisition: Four healthy subjects underwent free-breathing 3D bSSFP acquisition with spiral-like Cartesian sampling (VD-CASPR4) in a 1.5T scanner (Siemens Magnetom Aera) with 1.2mm3 isotropic resolution, TR/TE/FA= 3.35ms/1.47ms/90 °, 9-fold undersampling and acquisition time of ~3 mins. A 2D image-navigator enabled beat-to-beat 2D translational respiratory motion correction and thus 100% respiratory scan efficiency5.
VNN Training: The proposed VNN (Fig.1) consists of 3 network layers. Each layer presents 24 real-valued 11x11 filters and their corresponding activation functions. The network was trained on 2D magnitude images obtained from retrospectively undersampled (9-fold) 3D CMRA data of 20 healthy subjects. During the training procedure, the output of the VNN is compared to the corresponding fully-sampled reference images. A combination of the normalized root mean-squared error (RMSE) and of the structural similarity index (SSIM) was used as loss function.
Image Reconstruction: ADMM iterates between a data-consistency step, which performs MR-reconstruction using a conjugate gradient descent method with the output of the second step as prior knowledge; and a “denoising/de-aliasing” step, where the VNN provides a high-resolution 3D volume. The two sub-problems are solved iteratively in an Augmented Lagrangian scheme (Fig.1).
Experiments: Image quality of the proposed approach was compared to zero-filled and Wavelet-based CS6 reconstructions in terms of coronary artery length and sharpness.
Results: The proposed reconstruction presents higher image quality than zero-filled and CS (Fig. 2) with coronary artery length and sharpness comparable to the fully-sampled reference (Fig.3). The average total reconstruction time was ~1.5 min.
We demonstrated that the VNN prior can be successfully trained on retrospectively undersampled data. The whole-heart 3D coronary images obtained with ADMM+VNN have shown improved image quality with respect to zero-filled and CS reconstructions. The proposed method proved to be fast and efficient offering easy integration into clinical workflow.