SCMR 22nd Annual Scientific Sessions
Background: Over the past decade, there have been numerous advances in 3D cardiac imaging. However, 3D imaging has long scan and reconstruction time, hampering its clinical adoption. Accelerated imaging such as compressed sensing (CS) has been extensively used to reduce the scan time . Despite excellent image quality and high achievable acceleration factors, the reconstruction time remains very long [2,3]. In this study, we sought to develop an Ultrafast Reconstruction of UnderSampled k-space data (URUS) based on deep complex neural network for 3D Late Gadolinium Enhancement (LGE).
Fig 1a shows the proposed URUS reconstruction algorithm that uses the complex undersampled k-space data as input to create an artifact-free 3D dataset via deep learning. First, the acquired k-space data are zero-filled, FFTed, combined into a single image using B1-weighted combination, and finally a total variation regularization is performed . This dataset is then used as input to stack of complex convolution layers with three down-sampling and up-sampling stages of U-net architecture (Fig 1b) to learn a mapping function between the aliased complex input and Low-dimensional-structure self-learning and thresholding compressed-sensing (LOST) reconstructed magnitude images . In this model, the convolution and batch normalization layers were implemented to perform in the complex domain to maintain the intrinsic information of the complex data.
Results: Figure 2 shows reformatted 3D LGE images (acceleration factor of 5) with isotropic spatial resolution of 1.4 mm3 reconstructed using the proposed reconstruction algorithm in a patient with hypertrophic cardiomyopathy. Figure 3 shows example images for URUS- and LOST-reconstructed images for patients with scars. The reconstruction time was ~1 min for URUS vs. 1 hour for LOST. The mean square error was significantly decreased for URUS output (0.073±0.028) compared to its input (i.e. after coil combination and TV regularization step) (0.238±0.062, P<0.001). The structure similarity index and peak signal-to-noise ratios significantly increased for URUS output (0.831±0.037 and 11.63±1.48) compared to the input (0.794±0.060 and 6.42±1.3, respectively; P<0.001 for both).
URUS enables fast reconstruction of high-resolution accelerated 3D LGE data with reconstruction times in the order of minutes.