Applying machine learning and deep learning for improved acquisition, reconstruction and quantification in MRI
Abstract/Contents
- Abstract
- Magnetic Resonance Imaging, MRI, is a powerful imaging modality that is frequently used in both clinical and academic settings. With its advantages of flexibility in signal encoding, we can use MRI to non-invasively visualize various soft-tissue contrasts, showing not only anatomical but also metabolic and functional information. In addition, MRI is a radiation-free modality which makes it favorable in numbers of clinical applications because of the reduced radiation-risk compared with other radiology modalities such as X-ray, Computed Tomography (CT), Positron Emission Tomography (PET) etc. Despite the advantages of MRI techniques, there are still several challenges preventing MRI from becoming more efficient and accessible. First, the scan time for MRI is usually longer than other modalities such as X-ray and CT, since it requires enough measurements to resolve high-quality images for diagnostic tasks. In order to accelerate MRI, various fast-imaging techniques, such as Parallel Imaging (PI) and Compressed Sensing (CS) have been proposed to speed up MRI acquisition using under-sampling. However, it is still unclear what is the best approach to conduct the under-sampling as different under-sampling patterns may result in different reconstruction quality. Second, the reconstruction methods for under-sampled MRI need further improvement. The reconstruction algorithms are formed as nonlinear optimization problems using iterative optimization that can be time-consuming. Fixed and handcrafted penalty terms are usually used to regularize the optimization, which are hard to tune. There are often trade-offs between the speed of the algorithm and the quality of resulting images. In many cases, the imperfect artifact suppression or over-smoothing slows down the clinical adoption of these fast-imaging techniques. Third, MR images are typically not quantitative. Most clinical MRI protocols used nowadays are contrast-weighted sequences, which incorporate the tissue contrasts in qualitative ways. Therefore, the resulting MR images may vary a lot between different protocols and scanners, which makes it very difficult for radiologists to conduct quantitative analysis or longitudinal comparison. In this work, we propose to resolve these remaining challenges to further improve MRI technologies. We utilized state-of-the-art Machine Learning and Deep Learning algorithms to significantly improve these three essential components in MRI: faster acquisition, better reconstruction, and more accurate qualification. Specifically, we firstly propose a machine learning based method to optimize the undersampling pattern for accelerated acquisition. The results, validated on in-vivo multi-contrast brain and prostate MRI datasets, demonstrate that the proposed method can generalize well for different anatomy. It enables efficient (5sec-10sec) and adaptive under-sampling pattern optimization at per-subject/per-scan level, and achieves 30%-50% lower PI+CS reconstruction error at the same acceleration factor. To improve MRI acquisition with a safer protocol and lower contrast dose, a deep learning model is developed to enhance the MRI. The proposed Deep Learning method yielded significant (N=50, p< 0.001) improvements over the low-dose (10%) images (> 5dB PSNR gains and > 11.0% SSIM). Ratings on image quality and contrast enhancement are significantly (N=20, p< 0.001) increased. Comparing to true full-dose images, the synthesized full-dose images have a slight but not significant reduction in image quality (N=20, p=0.083) and contrast enhancement (N=20, p=0.068). Slightly better (N=20, p=0.039) motion-artifact suppression was noted in the synthesized images. Non-inferiority test rejects the inferiority of the synthesized to true full-dose images for image quality (95% CI: -14%-9%), artifacts suppression (95%CI: -5%-20%) and contrast enhancement (95% CI: -13%-6%). Then for reconstruction, a deep learning based method, Deep Generative Adversarial Neural Networks for Compressive Sensing MRI (GANCS), is proposed to solve Compressed Sensing using state-of-the-art deep learning models including the usage of Generative Adversarial Network. Extensive experiments based on a large cohort of abdominal MR data and knee datasets, with the evaluations performed by expert radiologists, confirm that the GANCS retrieves images with noticeable diagnostic quality improvements. Finally, a deep learning algorithm, Quantitative Susceptibility Mapping using Deep Neural Network (QSMnet), is described to further push the accuracy and robustness of quantitative MRI techniques, with Quantitative Susceptibility Mapping (QSM) as an example. Quantitative and qualitative image quality comparisons on in-vivo datasets demonstrate that the QSMnet results have superior image quality to those of TKD or MEDI results, and have comparable image quality to those of COSMOS. Additionally, QSMnet maps reveal substantially better consistency across the multiple head orientations than those from TKD or MEDI, demonstrating better stability for applications that require quantitative biomarkers.
Description
| 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 | 2018; ©2018 |
| Publication date | 2018; 2018 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Gong, Enhao |
|---|---|
| Degree supervisor | Pauly, John (John M.) |
| Thesis advisor | Pauly, John (John M.) |
| Thesis advisor | Nishimura, Dwight George |
| Thesis advisor | Zaharchuk, Greg |
| Degree committee member | Nishimura, Dwight George |
| Degree committee member | Zaharchuk, Greg |
| Associated with | Stanford University, Department of Electrical Engineering. |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Enhao Gong. |
|---|---|
| Note | Submitted to the Department of Electrical Engineering. |
| Thesis | Thesis Ph.D. Stanford University 2018. |
| Location | electronic resource |
Access conditions
- Copyright
- © 2018 by Enhao Gong
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...