A Novel Image Segmentation Method based on Energy Minimization with an Application to Cardiac PET Image Segmentation

Image segmentation is a challenging research area in computer vision, image processing, and computer graphics. Interactive image segmentation is one important research branch of image segmentation, which has been applied in many areas such as medical imaging, movie industry, object tracking and sports. In this master’s thesis, a novel image segmentation method has been developed. This segmentation method is based on a combination of geometric interactive image segmentation, classification and Gaussian Mixture Models. It has been successfully applied to rodents myocardium Positron Emission Tomography (PET) image segmentation and has been shown as an effective and accurate method for PET image segmentation. One main difference between standard image segmentation and interactive image segmentation is that, interactive image segmentation requires users to provide some information about foreground and background by using hard constraints. Typically, the ways to set hard constraints can be divided into two categories: scribbles and bounding box. However, conventional scribbles and bounding box cannot satisfy the requirements in some applications. One of the most significant features of this novel interactive image segmentation method is the way to set hard constraints. Unlike conventional ways of setting hard constraints, this segmentation method requires less input information from users, which makes it more user-friendly and efficient. In fact the method only requires users to provide a bounding box which encloses partially the target (foreground), and can yield an accurate segmentation result. Based on the above facts, it has a broad application prospect.