Registration and segmentation in perfusion mri: kidneys and hearts

Perfusion MRI is a promising noninvasive diagnostic tool for monitoring the status of the heart and of the kidney. This thesis focuses on developing the essential enabling technology for a quantitative analysis of the MR perfusion of a heart or a kidney, namely automatic registration and segmentation of time-series MR perfusion images. We consider three types of applications: renal perfusion MRI in rats, renal perfusion MRI in humans, and cardiac perfusion MRI in humans. In this thesis, the registration and segmentation algorithms are derived based on the energy minimization approach and the level set method. These algorithms are designed to address some unique challenges that commonly arise in perfusion MRI studies such as problems with rapidly changing intensity and contrast across the image sequence, at the same time, they are highly tuned to the specificity of each application context. The key contribution is the use of the time characteristics of the perfusion signals in developing registration and segmentation algorithms. For renal perfusion MRI in rats, we present a subpixel registration algorithm that compensates for the breathing motion during image acquisition. To segment the rat kidney and its anatomical structures, we develop an automatic image segmentation algorithm that utilizes the correlation information among pixels in the same image and the temporal correlation across the images in the sequence. For both cardiac and renal perfusion MRI in humans, we propose an interleaved registration and segmentation algorithm that exploits image features that are invariant to a rapidly changing image contrast and utilizes image segmentation results for the construction of the affine registration templates. The algorithm has been tested with various real cardiac and renal perfusion MR images, and has demonstrated very good performance in identifying the motion and the boundaries of the heart or the kidney in each frame of the sequence. Finally, we formulate a joint segmentation and registration algorithm to simultaneously segment and register the entire image sequence, it exploits all the information available: spatial or intra-image, and temporal or across the images of the sequence, as well as prior anatomical constraints.