Bayesian inference for physiological MRI by using deep learning techniques
<p>Medical imaging has become an indispensable tool in routine medical practice. It reveals the interior structure of human body under the skin and bones and provides biophysical information to diagnose and treat disease. Physiological MRI including perfusion MRI, diffusion MRI, functional MRI...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
2023
|
| Subjects: |
| _version_ | 1826313081321422848 |
|---|---|
| author | Zhang, Y |
| author2 | Michael, C |
| author_facet | Michael, C Zhang, Y |
| author_sort | Zhang, Y |
| collection | OXFORD |
| description | <p>Medical imaging has become an indispensable tool in routine medical practice. It reveals the interior structure of human body under the skin and bones and provides biophysical information to diagnose and treat disease. Physiological MRI including perfusion MRI, diffusion MRI, functional MRI, CEST, etc, is a branch of MRI techniques that aims to provide not only anatomical structure information as conventional MRI but also quantitative measurements of physiological and functional states of tissue.</p>
<p>Physiological MRI relies on the basic principles of nuclear magnetic resonance, which applies a strong magnetic field and radiofrequency pulses at resonant frequency to certain atomic nuclei for example hydrogen nuclei. Through monitoring the interaction between labeled nuclei and the surrounding environment, the physiological information of tissues can be captured in detail. Some mathematical models describe the dynamics of labeled nuclei behind physiological process within biological tissues, through which the MRI signals can be converted to meaningful physiological parameters, for example perfusion parameter provided by perfusion MRI or microstructural parameters provided by diffusion MRI.</p>
<p>In physiological MRI, parameter estimates are often quantified through statistical model inversion techniques. Non-linear least square and maximum likelihood estimation are popular choices to achieve point estimates. The Bayesian framework incorporates prior knowledge and enables the modeling of uncertainty while providing parameter estimates. However, the computational cost associated with these inference processes is often expensive, with the most representative example being Markov Chain Monte Carlo (MCMC) in Bayesian inference. For example, it has been reported that the computational time to process three 3D diffusion-weighted images by using non- linear least square method is 8h. The high computational burden mainly comes from two aspects: the complexity of model and the iterative optimization technique. Iterative optimization involves successive updates of parameter estimate until a minimum loss is met, which requires repetitive evaluations of model.</p>
<p>Based on the above discussion, this thesis primarily addresses the challenge of computational cost in the context of parameter estimation problem for physiological MRI, specifically within a Bayesian inference framework. In this thesis, the second chapter introduces the background of parameter estimation problem within physiological MRI. The third chapter presents a rapid parameter quantification framework with neural network as the tracer kinetic model. The fourth chapter introduces a Variational- Autoencoder-like framework designed for direct parameter quantification. The fifth chapter assesses the reliability and detectability of neural network based models and the final chapter outlines the possible future directions.</p> |
| first_indexed | 2024-09-25T04:07:15Z |
| format | Thesis |
| id | oxford-uuid:8503be2c-37a8-4d15-9264-929957c3a568 |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-09-25T04:07:15Z |
| publishDate | 2023 |
| record_format | dspace |
| spelling | oxford-uuid:8503be2c-37a8-4d15-9264-929957c3a5682024-06-04T09:39:24ZBayesian inference for physiological MRI by using deep learning techniques Thesishttp://purl.org/coar/resource_type/c_db06uuid:8503be2c-37a8-4d15-9264-929957c3a568Magnetic resonance imagingNeural networks (Computer science)Parameter estimationEnglishHyrax Deposit2023Zhang, YMichael, CGrau, V<p>Medical imaging has become an indispensable tool in routine medical practice. It reveals the interior structure of human body under the skin and bones and provides biophysical information to diagnose and treat disease. Physiological MRI including perfusion MRI, diffusion MRI, functional MRI, CEST, etc, is a branch of MRI techniques that aims to provide not only anatomical structure information as conventional MRI but also quantitative measurements of physiological and functional states of tissue.</p> <p>Physiological MRI relies on the basic principles of nuclear magnetic resonance, which applies a strong magnetic field and radiofrequency pulses at resonant frequency to certain atomic nuclei for example hydrogen nuclei. Through monitoring the interaction between labeled nuclei and the surrounding environment, the physiological information of tissues can be captured in detail. Some mathematical models describe the dynamics of labeled nuclei behind physiological process within biological tissues, through which the MRI signals can be converted to meaningful physiological parameters, for example perfusion parameter provided by perfusion MRI or microstructural parameters provided by diffusion MRI.</p> <p>In physiological MRI, parameter estimates are often quantified through statistical model inversion techniques. Non-linear least square and maximum likelihood estimation are popular choices to achieve point estimates. The Bayesian framework incorporates prior knowledge and enables the modeling of uncertainty while providing parameter estimates. However, the computational cost associated with these inference processes is often expensive, with the most representative example being Markov Chain Monte Carlo (MCMC) in Bayesian inference. For example, it has been reported that the computational time to process three 3D diffusion-weighted images by using non- linear least square method is 8h. The high computational burden mainly comes from two aspects: the complexity of model and the iterative optimization technique. Iterative optimization involves successive updates of parameter estimate until a minimum loss is met, which requires repetitive evaluations of model.</p> <p>Based on the above discussion, this thesis primarily addresses the challenge of computational cost in the context of parameter estimation problem for physiological MRI, specifically within a Bayesian inference framework. In this thesis, the second chapter introduces the background of parameter estimation problem within physiological MRI. The third chapter presents a rapid parameter quantification framework with neural network as the tracer kinetic model. The fourth chapter introduces a Variational- Autoencoder-like framework designed for direct parameter quantification. The fifth chapter assesses the reliability and detectability of neural network based models and the final chapter outlines the possible future directions.</p> |
| spellingShingle | Magnetic resonance imaging Neural networks (Computer science) Parameter estimation Zhang, Y Bayesian inference for physiological MRI by using deep learning techniques |
| title | Bayesian inference for physiological MRI by using deep learning techniques
|
| title_full | Bayesian inference for physiological MRI by using deep learning techniques
|
| title_fullStr | Bayesian inference for physiological MRI by using deep learning techniques
|
| title_full_unstemmed | Bayesian inference for physiological MRI by using deep learning techniques
|
| title_short | Bayesian inference for physiological MRI by using deep learning techniques
|
| title_sort | bayesian inference for physiological mri by using deep learning techniques |
| topic | Magnetic resonance imaging Neural networks (Computer science) Parameter estimation |
| work_keys_str_mv | AT zhangy bayesianinferenceforphysiologicalmribyusingdeeplearningtechniques |