Automatic detection, segmentation and motion characterization of the heart from tagged MRI

Cardiac disease is the leading cause of death in the developed countries. To reduce the mortality, early diagnosis is critical. Tagged MRI is a non-invasive technique for the study of cardiac deformation. It generates an MRI-visible tag pattern within the heart that deforms with the tissue during the cardiac cycle in vivo, which gives motion information of the myocardium. It has the potential of early diagnosis and quantitative analysis of various kinds of heart diseases and malfunctions. The difficulty preventing this technique from clinical use is the lack of efficient post-processing methods that automatically extract and analyze cardiac motion from tagged MRI data, which consists of image analysis tasks such as image preprocessing, tagging lines enhancement and tracking, tag removal, heart detection, cardiac boundaries segmentation, and motion or strain estimation. In this dissertation, a system of accurate and reliable automatic/semi-automatic tagged MR image analysis solutions will be given to all these problems. The methodologies of this system involves the interplay between traditional image processing techniques and state-of-the-art statistics, physics and machine learning based methods. In addition, medical prior knowledge and practices have been incorporated into the algorithms. In this research, a wavelet-like Gabor filter-based method has been developed to solve tasks such as tag enhancement, tag removal, myocardial tracking, and strain estimation. Because of its wide applications, Gabor filtering has the potential to become a routine function in tMRI analysis systems. We are also the first that introduced learning-based approaches into the detection and boundary segmentation of the heart in cardiac tMRI, by integrating statistical shape analysis, learning-based local appearance modeling, and sampling-based tracking techniques. For myocardial deformation analysis, we developed both tracking and non-tracking-based strain estimation algorithms, and conducted a quantitative comparison with registered ultrasound elastography. Based on our strain estimates, a novel tensor-based classification framework has been developed to identify and localize regional cardiac abnormalities in human subjects. Experimental results show the automatic detection, segmentation and motion characterization methods that we have developed in this dissertation can automate and largely speed up the image analysis process of tMRI, and achieve robust and accurate results. This research provides a promising avenue to make tMRI clinically accessible.