Image reconstruction methods for accelerated MR parallel imaging

Abstract

Magnetic resonance imaging (MRI) is a powerful and highly versatile imaging technique that has been widely applied in clinical diagnosis and preclinical research. Over the past several decades, parallel imaging has been continuously developed for accelerating MR data acquisition beyond the Nyquist sampling rate. Conventional parallel imaging techniques apply the spatial encoding capability of multiple receiver elements to reconstruct partially acquired data, thus requiring accurate coil sensitivity information for reconstruction. The main research scope of this dissertation is to develop advanced image reconstruction methods for highly accelerated parallel imaging without the need of separate coil sensitivity calibration.

First, a joint calibrationless low-rank reconstruction method was proposed for highly undersampled multi-channel multi-contrast datasets. Multi-contrast imaging has been applied routinely in many MRI applications, in which independent datasets of distinct contrast at the same slice location are acquired for providing differential diagnostic information. However, such multiple and independent scans are time-consuming and reconstruction of each single-contrast dataset may not be sufficient to achieve very high acceleration due to severe undersampling artifacts and intrinsic noise amplification. To jointly reconstruct highly undersampled multi-channel multi-contrast datasets, a novel low-rank Hankel tensor completion (MC-HTC) framework is formulated that exploits sharable information with respect to their highly correlated image structure, common spatial support, and shared coil sensitivity. The proposed MC-HTC effectively reduces both artifacts and noise amplification, thus enabling higher acceleration beyond what single-contrast reconstruction alone can offer.

Second, a novel deep learning framework, which estimates high-quality multi-channel spatial support directly from undersampled data, was presented for fast and calibrationless low-rank parallel imaging while preserving its reconstruction stability. Calibrationless low-rank parallel imaging has recently emerged that formulates reconstruction as a structured low-rank matrix/tensor completion problem. However, all existing low-rank reconstruction methods incur a cumbersome iterative matrix/tensor completion process and the targeted rank threshold must be carefully chosen in a manual and trial-and-error manner, severely hindering the adoption of calibrationless parallel imaging in routine clinical applications. In the proposed deep learning framework, accurate multi-channel spatial support is obtained from undersampled data by exploiting correlations of datasets acquired from the same MRI receiving system, rather than estimated from the null-space of incomplete matrix/tensor in a sequential iterative manner. Consequently, the proposed framework allows a direct least-square reconstruction without involving iterative low-rank completion and bypasses the need for rank determination.

Third, a simple and effective k-space sampling strategy, where the phase-encoding direction is alternated orthogonally among different contrasts or across adjacent slices, is presented. Conventional low-rank and sparse modeling for reconstruction require incoherent random undersampling patterns, which can be performed only along one designated phase-encoding direction in Cartesian 2D MRI. With our proposed sampling strategy, both the aforementioned MC-HTC and compressed sensing multi-slice reconstruction can be uniquely applied for Cartesian uniform undersampling, which conventionally has been considered an ill-conditioned reconstruction problem.

In conclusion, these studies have developed advanced reconstruction methods, making parallel imaging more efficient, simple, and robust in practice.