Unsupervised cardiac PET image segmentation

Automatic, unsupervised image segmentation plays an important role in medical imaging, but remains a challenging task in positron emission tomography (PET) due to unpredictable object shapes and inconsistent image quality resulting from noise and sampling artifacts. The main objective of this work is to develop a segmentation method for the mouse myocardium PET images based on deformable models. Two moving curves, one from inside of the left ventricle and one from the outside of the heart will be deformed to track heart boundaries. More precisely, topology constraints are incorporated to the energy functional governing the evolution of the contours to avoid any collision while allowing them to compete against each other until stabilization. First, we locate the heart, which is the region of interest (ROI) for our study, using level sets with high internal energy initialized from the extremities of the image. It is followed by an optimal thresholding and the application of the mean shift clustering algorithm to locate the center of the left ventricle region. This is where a second contour (interior contour) is initialized. The coupled contours allow to detect the correct myocardial boundaries and compute a number of useful quantities such as the ejection-fraction of the left ventricle and the myocardium wall thickness. The model was applied successfully to the automatic segmentation of the PET images of a mouse myocardium as measured by the Sherbrooke LabPET scanner.