Data-Driven Joint Optimization of Acquisition and Reconstruction of Quantitative MRI

Abstract

Quantitative magnetic resonance imaging (qMRI) measures physical, physiological or biological properties of tissues and thus provides reproducible imaging biomarkers for disease diagnosis and therapy response monitoring. However, long scanning and reconstruction time, low reproducibility, spatial resolution, and volume of coverage limit the clinical translation of qMRI.

The overall goal of this dissertation is to improve qMRI by exploiting its sparsity by data-driven deep learning methods. Such methods can provide more accurate and precise tissue parameters from highly undersampled or accelerated scans using a fraction of reconstruction time of conventional methods. Sampling patterns, image reconstruction and parameter estimation could be jointly optimized to directly minimize parameter quantification error under the same scan time.

Temporal sparsity in dynamic contrast enhanced magnetic resonance imaging (DCE MRI) was exploited by a long short-term memory (LSTM) neural network-based approach to provide more robust tissue parameter estimation. The network was trained on simulated DCE signals and tested on both simulated and real data. Compared to a conventional linear least squares (LLSQ) fitting method, the LSTM-based approach had higher accuracy for the data with temporally subsampling, total acquisition time truncation, or high noise level. Also, the LSTM-based method reduced the inference time by ~14 times compared to the LLSQ fitting. Validation of the method on real data demonstrated its clinical feasibility to provide high-quality tissue parameter maps.

Beyond temporal sparsity, the spatiotemporal sparsity of DCE MRI was further exploited by convolutional recurrent neural network. 2D Cartesian phase encoding k-space subsampling patterns were jointly optimized with image reconstruction to identify the most informative k-space data to acquire beyond the learned population prior knowledge. Both reconstruction image quality and parameter estimation accuracy were used to guide network training. The proposed method was trained and tested by multi-coil complex digital reference objects of DCE images. The proposed method achieved lower parameter estimation error at four temporal resolutions (2s, 3s, 4s, and 5s) compared with two benchmark methods and reduced parameter estimation bias and uncertainty in tumor regions at temporal resolution of 2s. The proposed method also showed robustness to contrast arrival timing variations across patients.

Compared with Cartesian sampling, non-Cartesian sampling provides more flexibility in trajectory design. This work also develops a deep learning framework that is able to synergistically optimize rotation angles of 3D spiral trajectories, image reconstruction, and parameter estimation of magnetic resonance fingerprinting. To counter the large problem size, an efficient model-based deep learning (MBDL) image reconstruction framework was developed. The MBDL image reconstruction provided more accurate parameter estimation than a state-of-the-art reconstruction method on both simulated and in vivo data. On simulated data, joint optimization of image-parameter reconstruction or sampling trajectory-image reconstruction were incorporated into the baseline MBDL framework and further improved tissue parameter estimation.