INTRODUCTION

CMR cine is the primary technique for assessing cardiac volume and function. However, it often necessitates multiple breath-holds to cover the whole heart, resulting in extended acquisition times. Additionally, scans from patients with irregular heart rhythms or difficulties in breath-holding may suffer from image artifacts. Deep learning (DL) has been used to accelerate cine images beyond what can be accomplished with current acceleration techniques¹. While DL achieves higher acceleration rates, the conventional sensitivity-encoding (SENSE)-based approach only utilizes a single coil-combined image, thereby limiting the full potential of the acquired multi-channel k-space data. We sought to develop an iterative self-consistent parallel imaging reconstruction (SPIRiT)-based approach that provides insights into each channel image within regularization by applying 3D convolutions across spatial and channel dimensions.

METHODS

Spatio-channel Regularized Deep Learning (SCR-DL):
Regularized MRI reconstruction is formulated as:

$$\arg \min _{\mathbf{\kappa}} \left\|\mathbf{D}\mathbf{\kappa}-\mathbf{y}\right\|_2^2 +\left\|\mathbf{G}\mathbf{\kappa}-\mathbf{\kappa}\right\|_2^2 + \lambda R(F^{-1}(\kappa)), \quad\quad (1)$$
where D is the subsampling operator that relates reconstructed k-space ($$$\kappa$$$) to the acquired data, G is the SPIRiT self-consistency operator that convolves entire k-space with calibration kernels across all channels², $$$\lambda$$$ is a weight term, $$$R(\cdot)$$$ is a regularizer and $$$F^{-1}(\cdot)$$$ is the inverse Fourier transform. SENSE-based DL (SENSE-DL) algorithms lead to a coil combined image relying on data-consistency with the acquired points. This is followed by a regularization process which is implicitly solved using a convolutional neural network operating on the coil-combined image. In contrast to SENSE-DL opponents, the proposed approach leverages the 3D convolutions within its regularizer, facilitating a comprehensive exploration of all channels, rather than relying on 2D convolutions applied to coil-combined images. Coil-wise self-consistency is followed by a regularizer that operates on individual channel images which are inverse Fourier transform of the interpolated k-spaces. A schematic of the proposed approach is depicted in Fig. 1.

Imaging Experiments:
Breath-hold ECG-gated segmented cine were collected at 3T (MAGNETOM Vida Siemens Healthineers, Erlangen, Germany) in the left ventricular (LV) short axis with the following imaging parameters: 1.8-fold GRAPPA acceleration, resolution=1.7×1.7mm², matrix size=156-208×208, and slice-thickness=8mm. These data were further retrospectively undersampled to R=5 by sampling every other 8^th line while keeping 24 center lines in k-space. Prospectively accelerated data were acquired on 5 subjects with the same imaging parameters. For comparison, baseline images were acquired at 1.8-fold GRAPPA in the same session. The current imaging pipeline acquires one slice per breath-hold, resulting in 10-12 breath-holds. This highly accelerated acquisition translates to only 2-3 breath-holds, depending on the subject's heart rate.

Implementation Details:
A supervised training was performed between fully-sampled segmented cine and retrospectively accelerated cine using 32 subjects, all slices, and every other 5^th time-frame. Variable splitting with quadratic penalty was used to unroll the algorithm in Eq. 1 that alternates between data consistency and regularizer 10 times. A 3D spatio-channel regularizer was used based on a 3D ResNet structure across 8-channels that were coil-compressed from 32 channels³. Adam optimizer, LR=1⋅10^-4, mixed normalized $$$\ell_1-\ell_2$$$ loss in both image and frequency domain were used. SPIRiT kernels^2,4 were calibrated on the full range of ACS region using 9×9 rectangular kernels and Tikhonov regularization with a penalty of 10^-3. SENSE-DL reconstruction was implemented with the same hyper-parameters, and sensitivity maps were generated, followed by the ESPIRiT implementation⁵.

Testing was performed on 10 subjects retrospectively in whom k-space data were undersampled by 5-fold, using all slices and time-frames unseen by the network. Normalized mean square error (NMSE) and structural similarity index measure (SSIM) were calculated and assessed using a paired t-test (P<.05 considered significant). LV ejection fraction (LVEF) and LV mass were manually calculated and assessed with Bland-Altman compared to 1.8-fold accelerated images in prospectively accelerated acquisitions.

RESULTS

Fig. 2 shows representative cine from 5-fold retrospectively accelerated acquisition where SCR-DL improves upon SENSE-DL, particularly in myocardium regions (yellow arrows). Mean NMSE and SSIM measurements are given in Table 1, showing SCR-DL performance. Prospective subjects at 5-fold acceleration are depicted in Fig. 3, where SCR-DL shows improved image quality compared to GRAPPA and SENSE-DL. Bland-Altman analyses (Fig. 4) show good agreement between the SCR-DL at 5-fold acceleration and 1.8-fold accelerated images for LVEF and mass.

CONCLUSION

The spatio-channel regularization applied to individual coils improved the image quality in accelerated cardiac cine and provided quantitative and qualitative improvements over existing methods.

Acknowledgements

No acknowledgement found.