Deep learning applications for quantitative magnetization transfer imaging

Abstract

Magnetization transfer (MT) is a phenomenon in which magnetization is exchanged due to the interaction between two proton pools: the free (water) pool and the bound (macro-molecular) pool. Magnetization transfer is often an undesirable effect, reducing the longitudinal magnetization and signal intensity or serving as a confounding factor. However, MT can be utilized as a contrast mechanism, differentiating tissues based on the macro-molecular content. This sensitivity is attractive for detecting and monitoring diseases that change the content of the brain tissues, such as multiple sclerosis or tumor. Quantitative magnetization transfer (qMT) imaging seeks to model the MT effect to produce quantitative parameters characterizing the exchange process. The dynamic of the two-pool system can be described by the addition of two parameters to the standard Bloch-McConnell equation: the magnetization exchange rate from the free to the bound pool, denoted as $k_f$, and the ratio of the bound pool and the free pool, denoted as F. Several acquisition methods have been proposed to probe the two-pool system. However, most methods require a long scan time to acquire the data at different strengths of the MT effect. The post-processing time also poses a problem, as the model needs to be fitted to the acquired data for every pixel. This fitting process is a time-consuming task, especially if performed with whole-brain data. This dissertation presents several attempts at tackling these issues through deep learning. The first study explored the application of deep learning to accelerate the acquisition and fitting of inter-slice qMT data. Two deep learning models, qMTNet-acq and qMTNet-fit, were proposed to accelerate each task, respectively. The models were combined to produce qMTNet-1 and qMTNet-2, which could perform both tasks. The proposed networks generated qMT parameters consistent with the conventional fitting method at a fraction of the processing time. The network also produced good fitting results with 3 times less acquired data. The second and third studies focused on on-resonant multiple phase-cycles bSSFP (mPC-bSSFP-qMT), a recently introduced qMT method that can also provide $T_1$, $T_2$, and $B_0$ maps. In the second study, a physics-informed deep neural network was proposed to improve the fitting of mPC-bSSFP-qMT. The network was trained using synthetic data generated from the signal model. This approach allowed the creation of potentially unlimited training data and matching ground truth without acquiring actual in-vivo data. The proposed method demonstrated better accuracy than the conventional ellipse fitting method on both simulated and in-vivo data. We also showed that the network was more robust to noise and with fewer acquisitions, which is attractive for integrating the method into clinical practice. In the third study, we explored further improvements to the mPC-bSSFP-qMT sequence, specifically through the optimization of the acquisition schedule with a learning-based method, as well as accelerating the data acquisition through multi-contrast sampling pattern optimization and reconstruction. Results showed that the proposed method could discover acquisition schedules that can reduce the fitting error of the parameters. The multi-contrast reconstruction method performed well on both retrospective and prospective data. Combining the two approaches enabled highly accelerated mPCbSSFP-qMT with high spatial resolution. Overall, the proposed methods described in this dissertation have shown several technical improvements for accelerating and improving qMT imaging.