Improving Parallel Magnetic Resonance Imaging Reconstruction Using Nonlinear Time Series Analysis

Abstract

Linear model is generally used in parallel magnetic resonance image (pMRI) reconstruction. Data acquired from multiple coils are learned and fitted for predicting and reconstructing missing k-space signal. Without sampling full k-space data, MRI speed is therefore accelerated and clinical scan cost can be reduced. However, due to noise and outliers existing multiple coil data, reconstructed image is deteriorated by noise and aliasing artifacts. To reduce noise and artifacts, some complicated models may remove noise and outliers in raw coil data and then predict missing data accurately. In this paper, a nonlinear time series analysis model was proposed from system identification and analysis perspective. The conventional pMRI reconstruction was formulated as a system identification task, but the proposed nonlinear time series analysis identifies model structured with removing noise and outliers. Experimental results demonstrated that the proposed model outperformed the conventional method. Noise and outliers were suppressed with high quality of the cardiac and brain applications.