Generalizing Deep Learning Reconstruction for Free-Breathing and Quantitative MRI

点击翻译按钮获取中文摘要

Project Summary/Abstract The goal of this project is to increase the precision and resolution of quantitative magnetic resonance imaging (MRI). Quantitative information such as tissue relaxation parameters (e.g., T1 and T2) measure tissue function and indicate disease-related changes in the heart, liver, brain, and other organs. For instance, T1 changes can provide evidence of diffuse fibrosis in the myocardium that can signal heart disease. Quantitative maps also are reproducible, directly comparable longitudinally and across subjects, and less affected by the properties of the scanner used, when compared versus common weighted (non-quantitative) clinical imaging. But, quantitative imaging involves more complicated and time-consuming pulse sequences. To accomplish this goal, this project will develop new machine learning algorithms for high-quality parameter mapping from free-breathing data. The first aim of this project will increase parameter map resolution achievable from highly accelerated, noisy data. The proposed method will integrate existing deep cascade network-based image reconstructions with convolutional network-based blocks for super-resolution and parameter map estimation. Preliminary studies suggest these new blocks improve sharpness and mitigate artifacts in the reconstructed parameter maps. The next aim will improve the training precision of such artificial neural networks to account for the significant per-voxel nonlinear fit variability in quantitative MRI. The proposed method will reweight the loss function used for calibrating these networks by the goodness-of-fit (coefficient of determination) of the reference maps obtained from fully sampled training data. Preliminary results demonstrate that quality-aware reweighting significantly improves reconstructed image quality when working with noisy training data. Experiments will evaluate the precision of both of these innovations against existing deep-learning-based reconstructions on T1 maps obtained from pre- and post-contrast cardiac images of volunteer patients. The final aim will address motion during the acquisition by estimating and tracking nonrigid motion in the data consistency stages of the deep cascade artificial neural network architecture. Two methods are proposed: deformable motion estimation already demonstrated on compressive model-based image reconstructions, and a new “re-blurring” convolutional neural network that automatically introduces artifacts into a “clean” image to match the motion-corrupted data. Both of these methods enforce consistency between motion-affected data and a motion-free image during the reconstruction. Both methods will be validated on both cardiac and abdominal images for motion artifacts and reconstruction quality against breath-held parameter mapping acquisitions.