Synopsis

Keywords: Image Reconstruction, Machine Learning/Artificial Intelligence

Ultra-high field 7T susceptibility-weighted imaging (SWI) has shown great potential in visualizing and evaluating a broad range of pathology, but suffers from long acquisition times. In this study, a complex-valued convolutional neural network (ComplexNet) model was proposed to reconstruct highly accelerated 7T SWI data. The average reconstruction time of ComplexNet was 0.56 seconds per slice (45.16 seconds per participant). Meanwhile, ComplexNet can provide high-quality 7T SWI for visualizing subtle pathology, including cerebral microbleeds, asymmetric deep medullary veins, and swallow tail sign.

Introduction

At ultra-high field 7T MRI, SWI can provide superior depiction of venous microvasculature, cerebral microbleeds, or deep gray matter structures (e.g., substantia nigra, subthalamic nucleus, and thalamus).¹ However, 7T SWI suffers from long scan times (usually 7-10 minutes with parallel imaging) due to sub-millimeter resolution acquisitions and long echo time.² The long acquisition time can easily lead to motion artifacts even in compliant subjects, and impose great burdens on patients with brain diseases. In this study, a complex-valued convolutional neural network (ComplexNet) was developed to fast and accurately reconstruct highly accelerated 7T SWI data, and its performance in terms of reconstruction quality and pathology visualization was investigated.

Methods

To faithfully reconstruct both MR magnitude and phase images, a deep learning model based on ComplexNet was proposed in this work.³ ComplexNet consists of a deep cascade of convolutional neural network (CNN) modules and data consistency layers. We adopt the complex residual network as the CNN modules to leverage the inherently complex-valued nature of SWI data and learn richer representations. Furthermore, each CNN module is followed by a data consistency layer to ensure data consistency in k-space domain. The mean-squared error between the ComplexNet reconstructed results and the fully sampled images was chosen as the loss function.
The SWI data were acquired at a whole-body 7T scanner (MAGNETOM Terra, Siemens Healthcare, Erlangen, Germany) equipped with a 32-channel head coil (Nova Medical, Wilmington, MA, USA) using a flow-compensated 3D SWI sequence. Imaging parameters were as follows: repetition time = 21.0 ms, echo time = 14.0 ms, field of view = 220 × 179 mm², slice thickness = 1.5 mm, voxel size = 0.25 × 0.25 × 1.5 mm³, flip angle = 15°, bandwidth = 210 Hz/Px, number of slices = 80, 3 × undersampling in the first phase-encoding direction, and a total acquisition time of approximately 7 minutes 29 seconds.
The SWI data were acquired from 130 participants, including patients with tumor, cerebrovascular disease, Parkinson’s disease, multiple sclerosis, and other conditions. The dataset was split into 2D axial slices of 91 participants for training, 39 participants for testing. The undersampled data were generated by retrospectively undersampling the reference k-space data with acceleration rates (R) of 6 and 8. The reconstruction results of ComplexNet were evaluated using two quantitative metrics: peak signal-to-noise ratio (PSNR) and structural similarity (SSIM).

Results

Mean PSNR values were 39.48 ± 2.17 and 38.79 ± 1.92 for ComplexNet at R = 6 and R = 8, respectively. Mean SSIM values were 0.9563 ± 0.0233 and 0.9517 ± 0.0235 for ComplexNet at R = 6 and R = 8, respectively. Average reconstruction time was 0.56 seconds per slice (45.16 seconds per participant) for ComplexNet in TensorFlow on the RTX 3090 GPU.
Figure 1 shows representative SWI images obtained using fully sampled and ComplexNet approaches at R = 6 and R =8 in a 63-year-old woman with Parkinson’s disease. It can be seen that the swallow tail sign is not clearly shown in the fully sampled images (Figure 1a) due to the degeneration of nigrosome 1 in PD. This potential imaging biomarker can also be consistently identified in the SWI images obtained using ComplexNet at both R = 6 (Figure 1b) and R = 8 (Figure 1c). In Figure 2, ComplexNet shows a similar diagnostic image quality as the fully sampled approach for visualization of subtle pathology, including asymmetric deep medullary veins, cerebral venous malformation, and multiple cerebral microbleeds.

Discussion and conclusion

SWI particularly benefits from increased SNR as well as the enhanced sensitivity of susceptibility effects at higher field strengths.⁴ Therefore, high-resolution SWI at 7T is a powerful tool to improve lesion characterization and treatment planning. In this study, we developed a ComplexNet model for fast and accurate reconstruction of highly accelerated 7T SWI data. Our results show that ComplexNet can offer comparable performance to the fully sampled SWI for visualizing a wide range of pathology. In a prospective way, the scan time of 7T SWI could be reduced by approximately 83.3% for R = 6, and 87.5% for R = 8 compared to the fully sampled acquisition, which is of greatest interest regarding clinical practice.

Acknowledgements

This work is supported by the National Natural Science Foundation of China (81825012, 81730048 and 82151309).