Development of machine learning-based analysis and processing tools applied to magnetization transfer MRI : Z-spectral denoising and fitting
(2025)- Abstract
- Chemical exchange saturation transfer MRI is a promising magnetization transfer (MT) imaging technique that utilizes, by radiofrequency irradiation, selective saturation of exchangeable protons in solutes. This enables the indirect detection of the solutes through the water signal after saturation transfer by chemical exchange, cross-relaxation, or the combination of the two. Typically, the water-signal change is studied within a frequency range where different molecular protons have visible resonances in the proton nuclear magnetic resonance spectrum due to the molecular mobility in the solution. In saturation transfer experiments, a normalized water signal as a function of saturation frequency, also referred to as a Z-spectrum, is... (More)
- Chemical exchange saturation transfer MRI is a promising magnetization transfer (MT) imaging technique that utilizes, by radiofrequency irradiation, selective saturation of exchangeable protons in solutes. This enables the indirect detection of the solutes through the water signal after saturation transfer by chemical exchange, cross-relaxation, or the combination of the two. Typically, the water-signal change is studied within a frequency range where different molecular protons have visible resonances in the proton nuclear magnetic resonance spectrum due to the molecular mobility in the solution. In saturation transfer experiments, a normalized water signal as a function of saturation frequency, also referred to as a Z-spectrum, is obtained. Like any other medical imaging modality, the MT imaging technique may be hampered by several drawbacks and limitations. A small effect size, together with the unavoidable noise in digital imaging, might render Z-spectra useless, and various data processing strategies are thus warranted. In the work described in this thesis the deep learning (DL)-based constrained loss autoencoder residual denoiser was developed, which combined the strength of latent mapping by autoencoders with subtractive residual denoising to obtain improved denoising performance and enable an increased level of signal recovery compared to other state-of-the-art approaches. Moreover, even for processed or high-quality Z-spectra, valuable biochemical information from various convolved contributions is commonly extracted by fitting the spectral data to a model. Conventional algorithms such as least squares (LS) fitting hamper efficient analysis due to inherent limitations such as a high dependence on data quality and sampling density, as well as long fitting times. To increase the feasibility of implementing MT-techniques, promote their clinical applications, and allow for studies of larger cohorts, it is necessary to streamline and standardize the analysis. A DL-based fitting approach for direct water saturation (DS) Z-spectra was developed and provided increased robustness and accelerated fitting compared to LS. The developed method found an application in a subsequent study by showing tangible differences in linewidth changes of DS spectra (pre- and post-glucose infusion) for healthy brain tissue and tumor. Finally, a multi-pool machine learning-based fitting approach using gradient boosted decision trees was also developed in the work of this thesis. The reduction in algorithmic complexity resulted in training times of approximately one minute, thus providing more freedom to change acquisition protocols. The fitting time was also reduced to approximately one second per brain compared to several hours with LS. The goodness-of-fit of four components was also empirically compared across the Lorentzian and Voigt spectral models, showing a statistically significant improvement for the latter. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/record/a668b182-c603-4694-ad95-ea0abd8e498e
- author
- Mohammed Ali, Sajad LU
- supervisor
- opponent
-
- Professor Chappell, Michael, University of Nottingham
- organization
- publishing date
- 2025-09-17
- type
- Thesis
- publication status
- published
- subject
- keywords
- Magnetic resonance imaging (MRI), Magnetization transfer, Chemical Exchange Saturation Transfer (CEST), Z-spectra, Artificial intelligence (AI), Machine Learning (ML), spectral fitting, Denoising, Deep learning
- pages
- 111 pages
- publisher
- Lund Learning, Lund University
- defense location
- Sal F5, Skånes Universitetssjukhus, Lund
- defense date
- 2025-10-14 09:00:00
- ISBN
- 978-91-8104-635-9
- 978-91-8104-634-2
- language
- English
- LU publication?
- yes
- id
- a668b182-c603-4694-ad95-ea0abd8e498e
- date added to LUP
- 2025-09-16 22:58:21
- date last changed
- 2025-09-19 09:52:10
@phdthesis{a668b182-c603-4694-ad95-ea0abd8e498e,
abstract = {{Chemical exchange saturation transfer MRI is a promising magnetization transfer (MT) imaging technique that utilizes, by radiofrequency irradiation, selective saturation of exchangeable protons in solutes. This enables the indirect detection of the solutes through the water signal after saturation transfer by chemical exchange, cross-relaxation, or the combination of the two. Typically, the water-signal change is studied within a frequency range where different molecular protons have visible resonances in the proton nuclear magnetic resonance spectrum due to the molecular mobility in the solution. In saturation transfer experiments, a normalized water signal as a function of saturation frequency, also referred to as a Z-spectrum, is obtained. Like any other medical imaging modality, the MT imaging technique may be hampered by several drawbacks and limitations. A small effect size, together with the unavoidable noise in digital imaging, might render Z-spectra useless, and various data processing strategies are thus warranted. In the work described in this thesis the deep learning (DL)-based constrained loss autoencoder residual denoiser was developed, which combined the strength of latent mapping by autoencoders with subtractive residual denoising to obtain improved denoising performance and enable an increased level of signal recovery compared to other state-of-the-art approaches. Moreover, even for processed or high-quality Z-spectra, valuable biochemical information from various convolved contributions is commonly extracted by fitting the spectral data to a model. Conventional algorithms such as least squares (LS) fitting hamper efficient analysis due to inherent limitations such as a high dependence on data quality and sampling density, as well as long fitting times. To increase the feasibility of implementing MT-techniques, promote their clinical applications, and allow for studies of larger cohorts, it is necessary to streamline and standardize the analysis. A DL-based fitting approach for direct water saturation (DS) Z-spectra was developed and provided increased robustness and accelerated fitting compared to LS. The developed method found an application in a subsequent study by showing tangible differences in linewidth changes of DS spectra (pre- and post-glucose infusion) for healthy brain tissue and tumor. Finally, a multi-pool machine learning-based fitting approach using gradient boosted decision trees was also developed in the work of this thesis. The reduction in algorithmic complexity resulted in training times of approximately one minute, thus providing more freedom to change acquisition protocols. The fitting time was also reduced to approximately one second per brain compared to several hours with LS. The goodness-of-fit of four components was also empirically compared across the Lorentzian and Voigt spectral models, showing a statistically significant improvement for the latter.}},
author = {{Mohammed Ali, Sajad}},
isbn = {{978-91-8104-635-9}},
keywords = {{Magnetic resonance imaging (MRI); Magnetization transfer; Chemical Exchange Saturation Transfer (CEST); Z-spectra; Artificial intelligence (AI); Machine Learning (ML); spectral fitting; Denoising; Deep learning}},
language = {{eng}},
month = {{09}},
publisher = {{Lund Learning, Lund University}},
school = {{Lund University}},
title = {{Development of machine learning-based analysis and processing tools applied to magnetization transfer MRI : Z-spectral denoising and fitting}},
url = {{https://lup.lub.lu.se/search/files/227931572/Thesis_V5_with_cover.pdf}},
year = {{2025}},
}