Multi-contrast magnetic resonance imaging with deep generative learning

Abstract

Magnetic Resonance Imaging (MRI) is a powerful biomedical imaging modality capable of producing high-resolution images with excellent soft tissue contrast across various anatomies of interest. Despite its superior resolution and versatile contrast options, MRI is inherently a slow imaging technique due to the sequential nature of data acquisition. This dissertation investigates two distinct paradigms aimed at accelerating MRI image generation. The first approach, known as Synthetic MRI, introduces a scanning paradigm that utilizes a fast multi-contrast sequence to estimate underlying quantitative tissue parameter maps. These maps can then be used to retrospectively synthesize any desired clinical contrast by adjusting scan parameters in silico. Two benefits of this approach are the reduced exam time and the ability to generate arbitrary contrasts offline. However, synthetically generated contrasts are known to deviate from the contrast of experimental scans. The reason for contrast mismatch is the necessary exclusion of some unmodeled physical effects such as partial voluming, diffusion, flow, susceptibility, magnetization transfer, and more. The inclusion of these effects in signal encoding would improve the synthetic images, but would make the quantitative imaging protocol impractical due to long scan times. Therefore, in this work, we propose a novel deep learning approach that generates a multiplicative correction term to capture unmodeled effects and correct the synthetic contrast images to better match experimental contrasts for arbitrary scan parameters. The physics inspired deep learning model implicitly accounts for some unmodeled physical effects occurring during the scan. As a proof of principle, we validate our approach on synthesizing arbitrary inversion recovery fast spin-echo scans using a commercially available 2D multi-contrast sequence. We observe that the proposed correction visually and numerically reduces the mismatch with experimentally collected contrasts compared to conventional synthetic MRI. Finally, we show the results of a preliminary reader study and find that the proposed method statistically significantly improves in contrast and SNR as compared to synthetic MR images. Hospitals remain hesitant to integrate Synthetic MRI into their clinical workflows due to several challenges, including a lack of standardization, the absence of key contrasts such as SWI and DWI, additional costs, and limited clinical acceptance. To address these limitations, we next explore the paradigm of undersampling in the k-space domain by acquiring fewer lines to increase acquisition speed. To reconstruct the images, traditional methods, such as parallel imaging and compressed sensing, can only accelerate acquisition up to a certain point before aliasing and noise artifacts emerge. This limitation stems from the constrained ability of these hand-crafted priors to effectively capture and represent the structures inherent in MR images. To address these limitations, deep learning-based methods have been proposed that learn priors directly from training data. These methods typically utilize conventional convolutional neural networks, generative adversarial networks, and other training paradigms. However, they are constrained to reconstruct undersampled data only when the sampling pattern matches the one used during training. To overcome this limitation, diffusion probabilistic models (DPM) have recently been introduced. In these models, the posterior sampling algorithm for reconstruction is designed so that the measurement process is independent of neural network training. The model only learns the gradient of the log-probability distribution. However, training the neural network to accurately learn the prior distribution requires large datasets with diverse contrasts and anatomies, which is challenging to acquire for MR images. In this dissertation, we propose a training technique to enable DPM models to learn priors from limited data. The model is initially trained on a large public dataset and then fine-tuned on local data with a reduced learning rate and for 2% of the total training time. As a proof of concept, we first present our results on the FLAIR contrast of the FastMRI dataset. We then validate our approach on multiple contrasts from the clinical MRI stroke protocol. The DPM-enabled reconstruction not only achieved high-fidelity reconstructions at higher acceleration factors but also effectively removed motion artifacts. Additionally, we demonstrate the application of the proposed training methods on prospectively accelerated data from a volunteer subject. Finally, we conducted a reader study in which a trained neuroradiologist compared the standard-of-care images with the accelerated reconstructions. The study found no significant difference in diagnostic quality. Moreover, the radiologist favored the accelerated images reconstructed with the DPM model due to fewer motion artifacts and higher resolution.