Synopsis

Keywords: Image Reconstruction, Cardiovascular

A deep learning reconstruction framework, trained in an end-to-end fashion and incorporating both a non-rigid respiratory motion estimation network and a motion-informed model-based reconstruction network, has been previously demonstrated to enable good quality images from seven-fold undersampled acquisitions for coronary magnetic resonance angiography applications. Herein, we apply the framework to whole-heart MRI scans of patients with congenital heart disease, enabling fast reconstruction of 7×-accelerated acquisitions and achieving image quality comparable to that of state-of-the-art patch-based low-rank iterative techniques.

Introduction

Cardiovascular magnetic resonance angiography (CMRA) is established for anatomical assessment in patients with congenital heart disease (CHD)¹. Conventional acquisition strategies rely on diaphragmatic respiratory gating (dNAV T₂prep-bSSFP), leading to long and unpredictable scan times and residual motion artefacts. Furthermore, achieving scan acceleration via undersampling k-space often involves the use of iterative reconstruction methods with long reconstruction times.
Recently, a deep learning reconstruction framework incorporating a motion estimation network and a motion-informed model-based reconstruction network, trained in an end-to-end manner, has been proposed for free-breathing whole-heart coronary magnetic resonance angiography with 100% respiratory scan efficiency². In that study, a combination of fully sampled and two-to-three-fold undersampled data was utilised to train the network, which was then successfully applied to reconstruct images from 7×-undersampled acquisitions (a ~2.5-minute scan). Herein, we propose to apply this reconstruction framework to 7×-undersampled 3D whole-heart scans of CHD patients to enable 3D whole-heart acquisition in 1-2 minutes.

Methods

40 adult patients with CHD were scanned on a 1.5 T system (MAGNETOM Aera, Siemens Healthcare) using an ECG-triggered free-breathing T₂-prepared bSSFP sequence³ and the following imaging parameters: FOV = 400 mm × 300 mm × 72-108 mm, resolution = 1.5 mm × 1.5 mm × 1.5 mm, flip angle = 90°, T₂-prep duration = 40 ms, TE = 1.75 ms, TR = 238ms, coronal orientation. For each patient, two scans were acquired: one with three-or-four-fold undersampling and another with seven-fold undersampling. A 2D image navigator (iNAV) was acquired during each heartbeat for respiratory binning, and to ensure each respiratory bin exhibited incoherent undersampling artefacts a variable-density Cartesian trajectory with spiral-like sampling (VD-CASPR)⁴ was utilised.
The data were randomly sorted into a training set (25 patients) and a testing set (15 patients). The training data (3-4× undersampled) were used to generate a ground truth and zero-filled image for each of four respiratory bins, as shown in Figure 1 (a). Since fully sampled data sets were not obtained, the ground truth images were reconstructed using the iterative patch-based low-rank method 3D-PROST³.
The deep learning reconstruction network (MoCo-MoDL)² consisted of a diffeomorphic motion estimation network⁵, which estimated the motion fields between respiratory bins, and a motion-informed model-based reconstruction network, which enforced data consistency between the reconstruction and the input zero-filled respiratory-bin images. A schematic of the network structure is shown in Figure 2. The network was trained on the 25-patient training set over 200 epochs and then tested on 7× prospectively undersampled scans of 15 patients in the manner depicted in Figure 1 (b). The output of the network was a motion-corrected whole-heart image in the reference respiratory phase.

Results

Average acquisition times were 1.5±0.3 minutes (7×-undersampled) and 3.2±0.7 minutes (3-4×-undersampled). Coronal cardiac images obtained using the MoCo-MoDL network are presented for four representative patients in Figure 3. In each case, the input to the network (a 7×-undersampled zero-filled soft-gated respiratory-bin image) is shown for comparison. Additionally, a “ground truth” PROST reconstruction calculated from the 3-4× undersampled scan of the same patient is presented. We note that while these ground truth reconstructions are generated following the same steps as depicted in Figure 1 (a) for the training datasets, they are not included in the network training and are only calculated for presentation.

Discussion

Visual inspection of the image quality in Figure 3 suggests that the MoCo-MoDL network achieves comparable image quality to the 3D-PROST “ground truth” images despite the higher undersampling factor. Additionally, the network reconstruction avoids the long reconstruction time associated with iterative algorithms such as 3D-PROST. However, we note that the images produced by the network tend to not be as sharp as those from the 3D-PROST reconstructions with a lower undersampling factor.
Future work will focus on tuning the hyperparameters of the network and incorporating non-rigid motion correction into the ground truth 3D-PROST reconstruction.

Conclusion

Fast reconstruction of 7×-undersampled whole-heart scans of CHD patients was achieved by utilising an end-to-end trained deep learning network to simultaneously estimate non-rigid respiratory motion fields and implement these fields in a motion-corrected model-based reconstruction.

Acknowledgements

This work was supported by the following grants: (1) EPSRC P/V044087/1; (2) BHF programme grant RG/20/1/34802, (3) Wellcome/EPSRC Centre for Medical Engineering (WT 203148/Z/16/Z), (4) Millennium Institute for Intelligent Healthcare Engineering ICN2021_004, (5) FONDECYT 1210637 and 1210638, (6) IMPACT, Center of Interventional Medicine for Precision and Advanced Cellular Therapy, ANID FB210024.