Accelerating pediatric magnetic resonance imaging using deep learning-based image reconstruction
Abstract/Contents
- Abstract
- Magnetic resonance imaging (MRI) is a powerful diagnostic tool for visualizing soft tissue anatomy, but physical limits on data acquisition speed result in uncomfortably long MRI exams. This is problematic for many patient populations, but especially for pediatric patients, who often require general anesthesia (GA) to reduce anxiety and body motion. Many attempts at accelerating data acquisition have been made to reduce or eliminate use of GA for pediatric MRI. For example, compressed sensing (CS) methods have been used to iteratively reconstruct rapidly acquired measurements into high-quality images by leveraging sparse priors. More recently, deep learning (DL) methods have been used to train deep neural network models to map the rapidly acquired measurements into even higher-quality images. While DL reconstruction approaches may potentially accelerate data acquisition beyond CS, these approaches have several issues which impede their clinical adoption. First, DL reconstructions require large quantities of high-quality ground truth data for supervised training, which can be costly and time-consuming to acquire. Second, memory requirements during network training limit the applicability of DL reconstruction to low-dimensional MRI data, such as static or dynamic 2-D imaging with limited spatiotemporal resolution. In this thesis, a series of projects demonstrating robust DL reconstruction techniques for acceleration of high-dimensional pediatric MRI will be presented. First, physics-based models are incorporated into deep CNN architectures to enforce consistency between intermediate network outputs and the rapidly acquired measurements in a novel method called DL-ESPIRiT. Physics-based modeling allows DL-ESPIRiT to be trained end-to-end in a supervised fashion with relatively little training data compared to non-physics-driven DL reconstruction. DL-ESPIRiT is applied and validated on 12X prospectively accelerated dynamic 2-D MRI scans acquired at Lucile Packard Children's Hospital. Finally, DL-ESPIRiT is extended to leverage subspace methods within the network to address GPU memory limitations during training. This method, known as deep learning-based subspace reconstruction (DL-Subspace), reconstructs a compressed representation of the MRI data instead of the data directly, thereby reducing the memory footprint during training and accelerating DL inference times. DL-Subspace is demonstrated to reconstruct 2-D dynamic MRI data with 4X higher memory efficiency and inference speed.
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 | 2021; ©2021 |
| Publication date | 2021; 2021 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Sandino, Christopher Michael |
|---|---|
| Degree supervisor | Vasanawala, Shreyas |
| Thesis advisor | Vasanawala, Shreyas |
| Thesis advisor | Nishimura, Dwight George |
| Thesis advisor | Pauly, John (John M.) |
| Degree committee member | Nishimura, Dwight George |
| Degree committee member | Pauly, John (John M.) |
| Associated with | Stanford University, Department of Electrical Engineering |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Christopher Michael Sandino. |
|---|---|
| Note | Submitted to the Department of Electrical Engineering. |
| Thesis | Thesis Ph.D. Stanford University 2021. |
| Location | https://purl.stanford.edu/tc342nz5000 |
Access conditions
- Copyright
- © 2021 by Christopher Michael Sandino
- 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...