Machine Learning-Based Adaptation of Data Sampling and Reconstruction for Efficient Dynamic MRI

点击翻译按钮获取中文摘要

ABSTRACT Magnetic resonance imaging (MRI) is essential for the detection and diagnosis of diseases. Clinical MRI scanners use fixed sequential data sampling patterns with long acquisition times, and employ nonadaptive reconstruction algorithms to generate images. The acquisitions are not usually tailored for the specific clinical task and patient characteristics, leading to sub-optimal images; they are often low-resolution, blurry, or contain errors that can reduce their diagnostic efficacy. Dynamic imaging applications, in which many images must be captured quickly to depict the motion of organs such as the heart, tend to suffer the most from these ill-effects. We propose to replace the conventional dynamic MRI acquisitions with a machine learning-based acquisition system, where the data sampling is efficiently optimized together with the reconstruction approach and task prediction, for optimized image quality and clinical task performance. First, we will explore and compare different ways of learning fast sampling of MRI frames to optimize image reconstruction quality metrics using large public data sets and current sophisticated (iterative) reconstruction algorithms. We will as- certain the sampling learning strategies that achieve the best image reconstruction quality at high data undersampling factors. Second, we will further extend machine learning throughout the MRI pipeline and develop approaches for joint adaptation of the data acquisition and image reconstruction and finally the task (e.g., quantification task) predictor as well. A key approach will use highly undersampled initial acquisitions (of current frame) and/or past (frame) data as input to the learned acquisition model to rapidly predict a patient- and frame-adaptive optimized sampling pattern. Then the samples from the scanner will be used to rapidly produce machine-learned reconstructions followed by task predictions. Particularly, for dynamic MRI, the temporal information from preceding images (frames) will be effectively incorporated and exploited in the proposed machine-learned models to drive efficient on-the-fly adaptive acquisitions and reconstruc- tions. We propose the mathematical formulations and algorithmic framework to accomplish these goals. The developed learning-based methods will be comprehensively evaluated and cross-compared in terms of image quality metrics (e.g., root mean squared error) and dynamic cardiac MRI task performance (ejection fraction estimation) at several undersam- pling or acceleration rates, and benchmarked using existing data sets as well as using newly collected cardiac MRI data. The development of smart imaging technologies that infuse learning across the imaging pipeline could enable rapid and effective task-driven adaptive imaging for dynamic cardiac MRI and related applications. Such a machine-learning MRI system could potentially improve clinical diagnosis and treatment, by helping enable the imaging system and acquisition to adapt in real-time to optimally detect and image various features at high resolution. Our goal in this project is to conduct the initial comprehensive studies to determine and analyze the potential, robustness, and algorithm behavior of the proposed machine learning dynamic MRI framework and techniques.