Synopsis

Keywords: Heart, Machine Learning/Artificial Intelligence, Deep Learning, Unrolled, Self Attention, CINE

A deep learning based ESPIRiT (DL-ESPIRiT) was recently proposed to reconstruct dynamic MRI data with higher reconstruction accuracy. However, the method still has difficulty resolving fine anatomic structures. We propose incorporating self-attention to the network using a computationally lightweight squeeze-excitation block (DL-ESPIRiT SE), which uses global information to select more important features while suppressing less important ones. We demonstrate improved reconstruction with DL-ESPIRiT SE, which is most pronounced during faster cardiac motion such as in ventricular ejection.

Introduction

Cardiac CINE MRI is widely used for characterizing heart morphology and function, but requires multiple breath holds to minimize motion artifacts from respiration. A deep learning based ESPIRiT (DL-ESPIRiT) reconstruction method was recently proposed to reconstruct dynamic MRI data with up to 12x undersampling and was shown to have higher reconstruction accuracy than state of the art compressed sensing algorithms.¹ However, DL-ESPIRiT still has difficulty resolving fast moving, fine anatomic structures such as the papillary muscles. We propose incorporating self-attention into the network to improve the reconstruction fidelity of such structures. Attention works under the premise that certain features of an input or image are more important and relevant than others. Specifically, we incorporate a computationally lightweight architectural unit called a squeeze-excitation block² into the DL-ESPIRiT network (DL-ESPIRiT SE), which uses global information to self-select more important features and suppress less important ones.

Methods

Squeeze-Excitation Block

Given an intermediate feature map , the squeeze-excitation (SE) block calculates 1D channel weights, $$$W\in\mathbb{R}^{1 \times 1\times 1\times C}$$$ and multiplies them to the feature map as follows:
$$U'=W(U)\times U$$
Where $$$U'$$$ is the final output. The element wise multiplication operator is performed after appropriately broadcasting the channel weights across the spatiotemporal dimensions. $$$W$$$ is calculated by first squeezing spatial information into a channel descriptor through average pooling:

$$z_c=\frac{1}{H \times W \times D} \sum^H_{i=1} \sum^W_{j=1} \sum^D_{k=1} u_c(i,j,k)$$

where $$$z_c$$$ is the squeezed channel descriptor and $$$H \times W \times D$$$ is the spatiotemporal dimensions of the feature maps $$$U=[u_1,u_2,...,u_c]$$$. Subsequently, the excitation network maps the channel descriptor to a learned set of channel weights, $$$W$$$, to capture channel wise dependencies. This excitation step is performed through a bottleneck composed of a dimension reducing fully connected layer with reduction ratio $$$r$$$, a ReLU, and then a fully connected dimension increasing layer followed by a sigmoid activation. The bottleneck is used to limit model complexity and increase model generalization.

$$s=\sigma(F_2\delta(F_1z))$$

where $$$F_1\in\mathbb{R}^{C \times C/r}$$$, is the dimension reducing fully connected layer, $$$F_2\in\mathbb{R}^{C/r \times C}$$$ is the dimension increasing fully connected layer, $$$\delta$$$ is the ReLU function, and $$$\sigma$$$ is the sigmoid activation.

SE blocks were incorporated into every ResNET module as shown in Figure 1 and within the DL-ESPIRiT network as shown in Figure 2. Specifications of the DL-ESPIRiT network itself are explained in more detail within reference 1.

Dataset
Fully sampled 2D cardiac cine datasets were acquired from 22 healthy volunteers on 1.5T and 3.0T GE MRI scanners using a balanced steady state free precession sequence with a 32-channel cardiac coil. Data were acquired in different standard cardiac views including short axis, 2-chamber, 3-chamber, and 4-chamber views. All data were compressed from 32 to 8 virtual coils. The 22 volunteers were divided into three cohorts for training, validation, and testing (10, 2, 10 split). Variable density undersampling masks were applied to simulate 10-15X acceleration across k-t space.

Results

Figures 3 and 4 show representative reconstructions of DL-ESPIRiT, and DL-ESPIRiT SE on retrospectively undersampled cardiac CINE with varying acceleration factors in the short axis and a 3-chamber view respectively. Figure 3 demonstrates that fine papillary muscle structures are reconstructed with higher fidelity with DL-SPIRiT SE even at higher acceleration. The advantage of DL-ESPIRiT SE is most pronounced when motion is greatest during the cardiac cycle. In Figure 4, the anterior wall of the aortic root is blurred during ventricular ejection when reconstructed with DL-ESPIRiT, which is reconstructed with higher fidelity with DL-ESPIRiT SE.

Conclusion

In this work, we demonstrate that self-attention using a light-weight squeeze-excitation block can improve the reconstruction quality of cardiac CINE images, which is most pronounced during faster cardiac motion such as in ventricular ejection.

Acknowledgements

No acknowledgement found.