Abstract
Chemical exchange saturation transfer (CEST) magnetic resonance imaging (MRI) is an advanced MRI technique capable of detecting low-concentration molecules with exchangeable protons in water. CEST uniquely identifies molecular changes associated with diseases such as tumors and Alzheimer's disease (AD). However, two factors limit the clinical application of CEST. First, CEST MRI requires multiple scans at various offsets to gather sufficient data for analysis, resulting in longer scanning times compared to traditional MRI techniques like T1-weighted (T1W) and T2-weighted (T2W). Second, raw CEST MRI data cannot be directly used for disease diagnosis; relevant information must be extracted through post-processing methods. To address these challenges, it is essential to develop advanced techniques that accelerate scanning time and optimize complicated post-processing accuracy. In this thesis, we designed accelerated post-processing and Artificial Intelligence (AI) based methods to enhance molecular imaging in brain tumors and AD.Firstly, we developed a post-processing method for rapid CEST MRI scans in brain tumors. CEST signals, including amide proton transfer (APT), relayed nuclear Overhauser enhancement (rNOE), and creatine, have shown associations with brain tumors. Rapid acquisition with relaxation enhancement (RARE) is a commonly used CEST acquisition model that is insensitive to B0 field inhomogeneity, providing a high signal-to-noise ratio (SNR) but requiring more time. Echo planar imaging (EPI) is a rapid acquisition sequence that offers lower SNR and is susceptible to distortion from B0 field inhomogeneity. Common distortion correction methods, such as field mapping necessitate additional scans for calculation. Field mapping requires an extra scan of the B0 map to compute the phase shift along the y-axis. Notably, the analysis of CEST MRI also requires a B0 map for B0 shift correction, which can be directly calculated from the Z-spectrum without an additional scan. In this study, we assessed the effectiveness of utilizing a B0 map generated from single-shot CEST-EPI to achieve distortion self-correction (DISC). In the creatine phantom, the structural similarity index (SSIM) of the raw CEST image increased from 0.818 to 0.921, while the SSIM of the creatine map improved from 0.875 to 0.921 after DISC. Following DISC, the spatial correlation of CEST values increased from 0.9682 to 0.9843. In the mouse study, the SSIM of the raw CEST image rose from 0.634±0.068 to 0.681±0.069 (P=0.0002), while the APT map and rNOE map improved from 0.909±0.014 to 0.940±0.012 (P<0.0001) and from 0.894±0.010 to 0.929±0.014 (P<0.0001) after DISC, respectively. After DISC, the spatial correlation of CEST values in the regions of interest increased from 0.8867 to 0.9158. Compared to CEST-RARE, DISC-CEST-EPI exhibited higher SSIM and spatial consistency than CEST-EPI.
Secondly, we focused on optimizing the analysis of CEST MRI in AD. AD is the most prevalent form of dementia, characterized by the accumulation of amyloid-β (Aβ) plaques in the brain. In addition to amyloid alterations, AD is associated with various other molecular changes that occur during its progression. Glucose CEST (glucoCEST), dynamic glucose-enhanced (DGE) MRI, and APT CEST have been utilized in AD diagnosis, along with their respective biomarkers. However, the CEST signal changes in AD can be subtle, and the connections between these CEST signal changes and different brain regions may not be apparent. Consequently, we employed deep learning as an approach for AD classification due to its ability to detect subtle changes and identify regional connections. Both spectral and regional CEST changes provide valuable information for disease diagnosis, and utilizing deep learning approaches can enhance CEST analysis. We developed a deep-learning method that effectively identifies abnormal changes by capturing both spectral and regional alterations across multiple CEST contrasts. By employing a 3-D convolutional neural network (CNN) with a combined spectrum-region input that encompasses a majority of the molecular information, accurate and rapid identification of AD mice from normal controls can be achieved. Spectrum-region-based deep learning models can be utilized for identifying AD from wild-type (WT) controls. Specifically, the LeNet-5 model utilizing MTRrex from all brain regions achieved the highest performance, with an F1-score of 86.7%.
In conclusion, the first part of this thesis demonstrated that CEST-EPI is a rapid imaging technique but is prone to image distortion due to field inhomogeneities. We evaluated the effectiveness of utilizing the field map generated from the Z-spectrum to achieve DISC in single-shot CEST-EPI, eliminating the need for additional field map acquisition. Results from phantom and mouse brain experiments indicated that DISC-CEST-EPI outperformed CEST-EPI in terms of image geometry and CEST quantification. The implementation of this DISC strategy could enhance lesion identification and improve diagnostic accuracy when utilizing single-shot EPI for CEST acquisition. The second part of this thesis demonstrated that spectrum-region-based deep learning models can be employed for the accurate and rapid identification of AD from WT controls. Specifically, the LeNet-5 model utilizing MTRrex from all brain regions achieved the highest performance in this study. Overall, the developed post-processing and AI methods have improved molecular imaging in brain tumors and AD, and both methods hold potential for accelerating the extraction of essential molecular changes in the brain for a wider clinical application of CEST MRI.
| Date of Award | 19 Sept 2025 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Wai Yan Kannie CHAN (Supervisor) |