UC Berkeley

Magnetic resonance imaging (MRI) is an effective imaging modality offering tremendous benefits to both science and medicine. It provides exceptional contrast for visualizing soft tissue, can capture images from any orientation, and does not involve any ionizing radiation. Its remarkable versatility enables a wide range of applications, including assessing blood flow, imaging brain activity with functional MRI (fMRI), and quantifying susceptibility mapping, ushering in a new era of clinical diagnosis and brain research. However, due to its physics limitations, acquiring MRI data is inherently time-consuming, which significantly extends scan times and limits throughput in hospitals. As a result, there is great interest in reconstructing diagnostic-quality images from limited measurements (k-space data) to shorten scan times.

For instance, parallel imaging (PI) capitalizes on spatially sensitive receive coil arrays to simultaneously acquire multiple MRI measurements. Compressed Sensing (CS) techniques have been employed to iteratively reconstruct under-sampled data into high-quality images by utilizing sparse priors. More recently, end-to-end deep learning (DL) based reconstruction techniques have been introduced, leveraging deep neural networks to learn the reconstruction pipeline directly from extensive training datasets, rather than relying on hand-crafted prior knowledge.

Although DL-based methods have demonstrated significant success surpassing PI and CS capabilities, several challenges persist that limit the fidelity and efficiency, for example: 1) Loss functions used in DL-based reconstruction are mostly hand-crafted, either pixel-wise or based on local statistics (e.g., SSIM loss), inadequately capture perceptual information, leading to compromised image quality and blurring; 2) Memory constraints during network training restrict the applicability of DL reconstruction for high-dimensional MRI (e.g., 2D+time, 3D, 3D+time); 3) The confidence or reliability of reconstructed structures remains insufficiently investigated, posing a challenge for DL-based approaches in clinical applications. 4) Unlike natural images, MRI data is inherently complex-valued and faces challenges due to the limited availability of fully-sampled ground truth. This constraint inevitably restricts the applications of deep learning-based MRI techniques to tasks without access to adequate ground truth.

In this dissertation, we introduce a series of projects aimed at overcoming existing obstacles and achieving enhanced fidelity and efficiency in Magnetic Resonance (MR) image reconstruction. Chapter 3 begins by reconstructing high-fidelity contrast-weighted images from highly under-sampled Magnetic Resonance Fingerprinting (MRF) scan. It introduces a supervised learning method that directly synthesizes contrast-weighted images (T1-weighted, T2-weighted, and FLAIR) from an MRF scan. This technique generates multi-contrast images with significantly reduced scan times, as detailed in the.

Chapters 4-6 feature physics-informed DL-based reconstruction from undersampled k-space data. First, A novel patch-based Unsupervised Feature Loss (UFLoss) is proposed as a novel perceptual loss function and incorporated into the training of DL-based reconstruction frameworks in order to preserve perceptual similarity and high-order statistics. Next, I employ our previously proposed memory-efficient learning framework to minimize the memory required for backpropagation, facilitating the training of DL-based unrolled reconstructions for large-scale 3D MRI and 2D+time cardiac cine MRI.

Then, I present our uncertainty estimation framework to identify when and where a reconstruction model is producing potentially misleading results. Our framework produces confidence intervals at each pixel of a reconstruction image with a rigorous finite-sample statistical guarantee.

Our in-vivo knee and brain results probe the quality of our uncertainty estimation model, which allows us to identify specific regions where the model performs poorly. A distinctive aspect of MRI lies in its inherently complex-valued data. The final section of this dissertation concentrates on the representation learning of complex-valued data. In contrast to deep learning applied to natural images, MRI faces the challenge of a scarcity of fully-sampled ground truth and well-annotated data.

To tackle this challenge, I introduce Complex-valued Scattering Representation (CSR) as a universal complex-valued representation, which so far demonstrates superior performance in both real-valued (e.g., RGB image) and complex-valued (e.g., MRI) image classification tasks compared to its counterparts, particularly when training samples are limited. Although this dissertation has not applied CSR to DL-based reconstruction, it represents a promising direction for future research.

Collectively, these approaches embody the central theme and progress toward MR image reconstruction with high fidelity, high efficiency, and high reliability.