Improvement of MR images with machine learning methods integrating multiple datasets

Abstract

MRI has gained an importance in clinical imaging and medical research because of its ability (i) to produce high resolution images with various tissue contrasts and (ii) to visualize physiological information. However, acquisitions of multi-contrast images need a long scan time, and MRI sequences sometimes require repetitive scans due to low signal-to-noise ratio or artifacts. Therefore, acceleration of MRI scan has been of importance in the MRI community. The research goal of this dissertation is to improve the efficiency of MRI scan and thus to reduce MRI scan time using machine learning. Machine learning is a method to learn problem solving by automatically extracting hierarchical feature representations from large-scale data. Information from multiple images can be integrated in the feature spaces, potentially providing better image quality. We propose three machine learning approaches that improve MRI images by integrating multiple datasets. First, machine learning algorithms are proposed to suppress banding artifacts, typically detected in the fast MRI sequence called “balanced steady state free precession (bSSFP)”, using fewer phase cycled (PC) bSSFP datasets than previously used. The results showed superior performance over the conventional methods. Second, machine learning algorithms are proposed to produce arterial spin labeling (ASL) perfusion images with fewer data acquisitions than previously used. ASL suffers from inherently low SNR, but the proposed method could quantify cerebral blood flow accurately with a shorter scan time. Lastly, machine learnings are proposed to reconstruct high-resolution MR images by incorporating another image acquired with a different sequence. The proposed methods produced better MR images from data with different phase cycles, data with different time points, and the down-sampled data. Machine learning algorithms could be successfully implemented to improve MR image quantitatively and qualitatively in various MRI techniques, therefore, they can be a good strategy for accelerating MR imaging that combines multiple datasets.