Improved data representations and data-efficient methods in deep learning for MRI applications
Abstract/Contents
- Abstract
- Magnetic resonance imaging (MRI) is a medical imaging modality which provides high-quality non-invasive soft tissue visualization. The resulting images are used to assess patient health and diagnose various diseases, such as coronary heart disease, brain tumors, and liver disease. Unlike positron emission tomography and computed tomography, MRI does not use harmful ionizing radiation, which makes it a preferable modality in pediatric patients. However, MRI scans are traditionally very slow, requiring patients to lie still for long periods of time to avoid motion artifacts. This is especially difficult and uncomfortable for young children. Therefore, imaging speed remains a main limitation of MRI. Scan times can be significantly reduced by collecting less measurements in the frequency domain; however, this leads to low-quality images. Image reconstruction addresses this by converting undersampled raw data to high-quality images. Deep learning (DL) methods have recently provided rapid and robust image reconstruction compared to traditional iterative methods. However, these DL methods still have several issues. First, most approaches split the complex-valued MRI data into separate real and imaginary channels within some kind of convolutional neural network (CNN). This approach does not accurately represent the underlying complex-valued structure of the data. Second, the vast majority of DL methods for MR image reconstruction are supervised, requiring large amounts of ground truth data. However, ground truth data cannot be acquired for many types of MRI sequences, making it impossible to train existing DL models for reconstruction. In this thesis, both of these issues are addressed in a series of projects. First, work on formulating and analyzing complex-valued CNNs for supervised MR image reconstruction is shown. Complex-valued convolutions, as opposed to real-valued convolutions, are shown to more accurately represent MRI data and thus perform superior reconstructions, especially in terms of phase information. Additionally, it is shown that the superior phase recovery of these complex-valued networks provides more accurate fat-water separation, which is important for applications such as liver fat quantification, as well as more accurate blood flow estimation, an important cardiovascular application. Second, work is presented on unsupervised MR image reconstruction. A framework using generative adversarial networks is formulated to produce high quality reconstructions without ever using any ground truth images during training. Our unsupervised method is compared to compressed sensing (CS), which, being a traditional signal processing method, also requires no ground truth data. The reconstructions from our unsupervised method are superior compared to CS in terms of quantitative image quality metrics, especially at higher accelerations. This method also runs up to 7 times faster compared to CS. An additional reconstruction-related problem in MRI lies in the intrinsic high-dimensional nature of MRI datasets. In MRI, using multiple radio frequency (RF) coil arrays can increase parallel imaging (PI) acceleration and improve signal-to-noise (SNR) ratio. The large number of coils creates prohibitively large MRI datasets in space and infeasible computation time for reconstruction. Additionally, these datasets often contain redundant information across the various acquired images. Coil compression algorithms are effective in mitigating this problem by compressing the datasets to convert the original set of coil images into a smaller set of virtual coil images. This enables smaller datasets and faster computation time. However, traditional iterative coil compression methods are lossy and time-consuming. In this work, we construct an encoder-based neural network for the purposes of dimensionality reduction and apply it to the coil compression task in pursuit of higher reconstruction accuracy and faster coil compression. The learned compression method achieves up to 1.5x lower NRMSE and up to 10 times runtime speed compared to traditional methods on a benchmark test dataset.
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 | 2022; ©2022 |
| Publication date | 2022; 2022 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Cole, Elizabeth Katherine |
|---|---|
| Degree supervisor | Pauly, John (John M.) |
| Thesis advisor | Pauly, John (John M.) |
| Thesis advisor | Nishimura, Dwight George |
| Thesis advisor | Vasanawala, Shreyas |
| Degree committee member | Nishimura, Dwight George |
| Degree committee member | Vasanawala, Shreyas |
| Associated with | Stanford University, Department of Electrical Engineering |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Elizabeth Cole. |
|---|---|
| Note | Submitted to the Department of Electrical Engineering. |
| Thesis | Thesis Ph.D. Stanford University 2022. |
| Location | https://purl.stanford.edu/vj235gf0426 |
Access conditions
- Copyright
- © 2022 by Elizabeth Katherine Cole
- 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...