Purpose: Cardiac positron emission tomography (PET) images usually show two kinds of artifacts: the limited resolution of PET leads to partial volume effects and the motion of the heart induces blurring. These phenomena degrade the PET images and induce errors in the quantification. One method of reducing this problem is to use gated PET data. However, the reduction of information per phase leads to an increase in noise on the reconstructed images. Alternatively, the PET data have to be corrected for motion and partial volume effects.
Methods: Optical flow methods have been shown to accurately estimate the motion between PET image frames. These methods assume that the brightness of the objects remains constant between the frames. This condition is not fulfilled in cardiac PET data because the brightness of the cardiac muscle tissue (myocardium) is not accurately resolved due to the partial volume effect. Therefore, the use of a newly developed optical flow method based upon the conservation of mass condition is proposed to correct the cardiac PET data. Mass conservation is applicable to PET images as the total activity in the field of view may be assumed to remain almost constant, if the data are precorrected for radioactive decay. Two variants of the method using the quadratic and the nonquadratic penalization are presented. The methods were evaluated with respect to correlation coefficient, myocardial thickness and the blood pool activity in the left ventricle on phantom data and on 14 patient image volumes.
Results: The proposed methods showed that the cardiac motion can be efficiently corrected despite partial volume effects. The correlation coefficient between the image volumes increased from 0.87 to 0.98 on average. The change in myocardial thickness was reduced from 28% to 3%. The variation in blood pool activity was reduced from 80% to 8%. The algorithm needed only about 4 s for execution.
Conclusions: A mass preserving optical flow method of cardiac motion correction in 3D PET data has been presented and tested on phantom as well as patient data. The results show that the motion was corrected for all datasets effectively.