Combined super resolution and motion-corrected under sampled DL reconstruction for 18-fold-accelerated 3D whole-heart MRI (RF_SA_446)

High-resolution (HR) whole-heart CMR typically involves long scan times due to the requirement for large amounts of k-space data. Scans can be accelerated by undersampling k-space, acquiring at low image resolution, or both.
 

In the undersampled paradigm, iterative and deep-learning-(DL)-based reconstruction methods have been proposed, including a motion-corrected model-based DL framework (MoCo-MoDL) for the fast reconstruction of whole-heart images.^1,2

Alternatively, super-resolution (SR) techniques can be employed to up-sample low-resolution (LR) images. DL networks have been proposed to perform this task, by learning the relationship between LR and HR images in large sets of training data^3-5. MoCo-MoDL has recently been adapted for fully sampled super-resolved reconstruction⁶. Unlike image-based SR methods, the Super-MoCo-MoDL framework ensures data-consistency, which is achieved by treating SR as a k-space down-sampling problem.

In this work, we adapt Super-MoCo-MoDL for a simultaneous SR and undersampled reconstruction, achieving images with 0.9 mm isotropic resolution from 18-fold-accelerated scans in ~2 minutes.

Methods:

Twenty patients with suspected coronary artery disease were scanned twice on a 1.5-T MRI scanner (MAGNETOM Sola, Siemens Healthcare) using an ECG-triggered free-breathing T₂-prepared bSSFP sequence⁷, with a 4.5-fold undersampled 3D variable-density Cartesian spiral-like trajectory (VD-CASPR)⁸, at isotropic HR (0.9 mm³) and at anisotropic LR (1.8×1.8×0.9 mm³). A 2D image navigator (iNAV) was acquired at each heartbeat to enable respiratory binning and non-rigid respiratory motion correction.

For network training and validation, a 156-patient dataset was formed from 34 T₂-prepared bSSFP 3D whole-heart scans at 1.5 mm isotropic resolution, 95 MTC-BOOST⁹ 3D whole-heart scans at 1.4-1.5 mm isotropic resolution and 27 iT₂-prep-BOOST¹⁰ 3D whole-heart scans at 1.3-1.4 mm isotropic resolution. Each used a 4.5-fold-undersampled VD-CASPR trajectory and was reconstructed with the non-rigid-motion-corrected patch-based low-rank method (NR-PROST)¹¹ to generate ground-truth respiratory-bin images.

Corresponding LR images were formed by cropping, zero-padding and undersampling the k_y-k_z plane to match the sampling density of the LR acquisitions (Fig. 1), followed by a zero-filled reconstruction.

The Super-MoCo-MoDL framework (Fig. 2) consists of a diffeomorphic motion estimation network¹² and a motion-informed model-based reconstruction involving a super-resolving and denoising U-Net.

The U-Net was pre-trained for 400 epochs on 2315 scans from the NYU fastMRI dataset¹³. The complete framework was then trained end-to-end on two 16-GB GPUs for 1000 epochs and tested on the prospective undersampled LR data.

Results:

The average acquisition time for the LR scans was ~2 minutes, compared to ~9 minutes for 4-fold-accelerated HR. Slices of the SR images obtained with Super-MoCo-MoDL from anisotropic LR scans are presented in Fig. 3 for two representative patients. For comparison, the same slices are shown for the HR reference images, the LR zero-filled network-input images and NR-PROST reconstructions of zero-padded LR data.

Conclusion:

The Super-MoCo-MoDL framework achieves sharp high-resolution images with comparable quality to the prospective HR NR-PROST reconstructions, despite the ~4.5-fold-faster acquisition times and ~150-fold-faster reconstruction times.