Acceleration of 3D magnetic resonance angiography

Placeholder Show Content


Magnetic resonance imaging (MRI) is an important medical imaging modality for imaging soft tissue. A fundamental limitation of MRI is that there are tradeoffs between scan time, image resolution, and image signal-to-noise ratio (SNR). Typically, image resolution and SNR are not sacrificed because the images need to be diagnostically useful. Therefore, scan time is typically increased. For magnetic resonance angiography (MRA), this problem is exacerbated. Blood vessels are very small and, because of the ballistics of the heart pumping, the arteries are constantly in motion. Therefore, high resolution is required to resolve the vessels and the images need high SNR such that the vessels are distinguishable from noise. Additionally, while most clinical MRI acquires 2D slices, MRA typically requires 3D imaging because blood vessels typically do not lie in a single plane, which further extends scan time. To accelerate 3D MRA, we approached the problem from several directions. One method was to reduce the signal from all of the other tissue that we did not wish to image. We proposed a combined T2-preparation, outer volume suppression (OVS), fat saturation pulse to improve 3D coronary artery imaging. The pulse sequence consisted of the following pulse sequence: a 90°_-60 180°_{60} composite nonselective tip-down pulse, two 180°_Y hard pulses for refocusing, and a -90° spectral-spatial sinc tip-up pulse. The pulse sequence was played twice, first selective in the x-axis and then selective in the y-axis for a 2D OVS. Compared to another published T2-preparation OVS pulse sequence, the proposed pulse had superior image edge profile acutance values for the right (P< 0.05) and left (P< 0.05) coronary arteries, suggesting superior vessel sharpness. The proposed sequence also had superior SNR (P< 0.05) and passband-to-stopband ratio (P< 0.05). Reader scores and reader preference indicated superior image quality of the proposed sequence for both the right (P< 0.05) and left (P< 0.05) coronary arteries. Therefore, the proposed sequence had superior image quality and suppression, potentially enabling scan acceleration with reduced field of view imaging. Another technique to accelerate MRA was to enable more efficient acquisition trajectories. Non-Cartesian trajectories have the potential to be highly efficient but their efficiency is limited in practice by off-resonance. A residual convolutional neural network was trained to correct off-resonance artifacts in pediatric MRA exams (Off-ResNet). Training data was acquired from exams with a short readout scan (1.18 ms ± 0.38) and a long readout scan (3.35 ms ± 0.74) at 3 T. Short readout scans with longer scan times but negligible off-resonance blurring were used as reference images and augmented with additional off-resonance for supervised training examples. Long readout scans with greater off-resonance artifacts but shorter scan time were corrected by autofocus and Off-ResNet and compared with short readout scans by normalized root-mean-square error (NRMSE), structural similarity index (SSIM), and peak SNR (PSNR). Scans were also compared by scoring on eight anatomical features by two radiologists. Off-ResNet had superior NRMSE, SSIM, and PSNR compared to uncorrected images across ±1 kHz off-resonance (P< 0.01). Off-ResNet also had superior NRMSE over -677 Hz to +1 kHz and superior SSIM and PSNR over ±1 kHz compared with autofocus (P< 0.01). Radiologic scoring demonstrated that long readout scans corrected with Off-ResNet were noninferior to short readout scans (P< 0.05). The long readout scans were 59.3% shorter than the short readout scans, and Off-ResNet was able to correct the long readout scans to a noninferior quality compared to the diagnostically standard short readout scans. One limitation of Off-ResNet was that it used zero-order field maps to generate training data. While zero-order field maps are not realistic, it is also expensive to acquire large datasets with field maps. A generative adversarial network (GAN) was trained to generate realistic but not quantitatively accurate field maps for a given anatomical input image. The input to the generator network was an anatomical image from a single echo of a multi-echo T2*-IDEAL scan and the real examples of the corresponding field maps were also from the same set of T2*-IDEAL scans. We also proposed using a discriminator with Off-ResNet to improve off-resonance correction (Off-ResGAN). To evaluate the contributions of using a dataset with generated field maps and a GAN for off-resonance correction respectively, we trained four networks: 1. Off-ResNet and a dataset created with zero-order field maps, 2. Off-ResNet and a dataset created with generated field maps, 3. Off-ResGAN and a dataset created with zero-order field maps, and 4. Off-ResGAN and a dataset created with generated field maps. The four networks were evaluated on validation datasets created with zero-order field maps and generated field maps. The Off-ResGAN networks had superior NRMSE, SSIM, and PSNR on both datasets. With respect to datasets, each network performed best on the validation dataset that matched its training dataset. However, the networks trained with zero-order field maps tended to have worse performance on the generated field map validation dataset relative to the the performance of networks trained with generated field maps on the zero-order field map validation dataset. These initial results suggested that Off-ResGAN has superior image quality, and the benefit of using a dataset with generated field maps is increased robustness with respect to realistic off-resonant images. Undersampling is an important technique to accelerate scans. Undersampling reconstruction can involve a combination of parallel imaging and compressed sensing. An open question however is an optimal undersampling trajectory as measured by an image quality metric and with respect to a reconstruction technique. Moreover, as the scan proceeds, more information is collected and therefore increasingly better decisions can be made about the optimal undersampling trajectory. We proposed an approach using deep reinforcement learning to train an agent to learn from previous scans and combine the previously learned statistics with the new incoming data to determine the trajectory of the next readout. To frame the reinforcement learning problem, the environment was the reconstruction technique, the reward was the image metric, and the agent was a deep Q-learning convolutional neural network. Initial results with an L2 image quality metric and compressed sensing and deep learning reconstruction techniques indicated that the agent was capable of learning optimal trajectories. L2 was a special metric where we could analytically calculate the optimal trajectory. This suggested that the agent could learn optimal trajectories for other image metrics and reconstruction techniques where computing the optimal trajectory was intractable. In conclusion, the proposed techniques have demonstrated several strategies for accelerating 3D MRA. Through a combination of them, 3D MRA may become more feasible and accessible as a clinical tool


Type of resource text
Form electronic resource; remote; computer; online resource
Extent 1 online resource
Place California
Place [Stanford, California]
Publisher [Stanford University]
Copyright date 2020; ©2020
Publication date 2020; 2020
Issuance monographic
Language English


Author Zeng, David Yaxiong
Degree supervisor Nishimura, Dwight George
Thesis advisor Nishimura, Dwight George
Thesis advisor Pauly, John (John M.)
Thesis advisor Vasanawala, Shreyas
Degree committee member Pauly, John (John M.)
Degree committee member Vasanawala, Shreyas
Associated with Stanford University, Department of Electrical Engineering.


Genre Theses
Genre Text

Bibliographic information

Statement of responsibility David Y. Zeng
Note Submitted to the Department of Electrical Engineering
Thesis Thesis Ph.D. Stanford University 2020
Location electronic resource

Access conditions

© 2020 by David Yaxiong Zeng
This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).

Also listed in

Loading usage metrics...