Comparison of Image Segmentation and Registration Based Methods for Analysis of Misaligned Dynamic H215O Cardiac PET Images

In this study, we compared quantitatively image segmentation and registration based methods to find misalignment between two dynamic (H2O)-O-15 cardiac PET images. Due to a low contrast between tissues in oxygen-15-labeled images, we first applied independent component analysis (ICA) to separate the different cardiac structures. The misalignment was then defined from the separated ICA component images using two different methods. We used Deformable Models based Dual Surface Minimization (DM-DSM) and normalized mutual information based image registration algorithms in the comparison. The evaluation was done using realistic phantom data, generated using the MCAT phantom and the PET SORTEO Monte Carlo simulator. The simulated data consisted patient movement between the image sets and in addition, to produce more realistic data the movement within one time frame due the respiratory and cardiac movement. The quantitative results showed that the image registration based method was more accurate to find the misalignment between the image sets than the segmentation based method. One reason for this was that the segmentation algorithm was more dependent on the quality of the ICA separation result.